Stage Harbor release files 50501-51000
Browse filesThis view is limited to 50 files because it contains too many changes. See raw diff
- tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/big_bird/test_modeling_big_bird.py +969 -0
- tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/big_bird/test_modeling_flax_big_bird.py +224 -0
- tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/big_bird/test_tokenization_big_bird.py +233 -0
- tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/bigbird_pegasus/__init__.py +0 -0
- tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/bigbird_pegasus/test_modeling_bigbird_pegasus.py +0 -0
- tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/biogpt/__init__.py +0 -0
- tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/biogpt/test_modeling_biogpt.py +461 -0
- tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/biogpt/test_tokenization_biogpt.py +99 -0
- tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/bit/__init__.py +0 -0
- tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/bit/test_modeling_bit.py +310 -0
- tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/blenderbot/__init__.py +0 -0
- tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/blenderbot/test_modeling_blenderbot.py +559 -0
- tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/blenderbot/test_modeling_flax_blenderbot.py +418 -0
- tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/blenderbot/test_modeling_tf_blenderbot.py +235 -0
- tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/blenderbot/test_tokenization_blenderbot.py +75 -0
- tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/blenderbot_small/__init__.py +0 -0
- tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/blenderbot_small/test_modeling_blenderbot_small.py +568 -0
- tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/blenderbot_small/test_modeling_flax_blenderbot_small.py +411 -0
- tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/blenderbot_small/test_modeling_tf_blenderbot_small.py +257 -0
- tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/blenderbot_small/test_tokenization_blenderbot_small.py +88 -0
- tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/blip/__init__.py +0 -0
- tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/blip/test_image_processing_blip.py +148 -0
- tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/blip/test_modeling_blip.py +1435 -0
- tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/blip/test_modeling_blip_text.py +183 -0
- tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/blip/test_modeling_tf_blip.py +901 -0
- tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/blip/test_modeling_tf_blip_text.py +181 -0
- tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/blip/test_processor_blip.py +167 -0
- tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/blip_2/__init__.py +0 -0
- tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/blip_2/test_modeling_blip_2.py +0 -0
- tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/blip_2/test_processor_blip_2.py +141 -0
- tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/bloom/__init__.py +0 -0
- tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/bloom/test_modeling_bloom.py +825 -0
- tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/bloom/test_modeling_flax_bloom.py +251 -0
- tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/bloom/test_tokenization_bloom.py +162 -0
- tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/bridgetower/__init__.py +0 -0
- tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/bridgetower/test_image_processing_bridgetower.py +158 -0
- tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/bridgetower/test_modeling_bridgetower.py +692 -0
- tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/bridgetower/test_processor_bridgetower.py +203 -0
- tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/bros/__init__.py +0 -0
- tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/bros/test_modeling_bros.py +455 -0
- tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/byt5/__init__.py +0 -0
- tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/byt5/test_tokenization_byt5.py +362 -0
- tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/camembert/__init__.py +0 -0
- tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/camembert/test_modeling_camembert.py +84 -0
- tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/camembert/test_modeling_tf_camembert.py +56 -0
- tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/camembert/test_tokenization_camembert.py +220 -0
- tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/canine/__init__.py +0 -0
- tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/canine/test_modeling_canine.py +573 -0
- tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/canine/test_tokenization_canine.py +334 -0
- tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/chameleon/__init__.py +0 -0
tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/big_bird/test_modeling_big_bird.py
ADDED
|
@@ -0,0 +1,969 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# coding=utf-8
|
| 2 |
+
# Copyright 2021 The HuggingFace Inc. team. All rights reserved.
|
| 3 |
+
#
|
| 4 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
| 5 |
+
# you may not use this file except in compliance with the License.
|
| 6 |
+
# You may obtain a copy of the License at
|
| 7 |
+
#
|
| 8 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
| 9 |
+
#
|
| 10 |
+
# Unless required by applicable law or agreed to in writing, software
|
| 11 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
| 12 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
| 13 |
+
# See the License for the specific language governing permissions and
|
| 14 |
+
# limitations under the License.
|
| 15 |
+
"""Testing suite for the PyTorch BigBird model."""
|
| 16 |
+
|
| 17 |
+
import unittest
|
| 18 |
+
|
| 19 |
+
from transformers import BigBirdConfig, is_torch_available
|
| 20 |
+
from transformers.models.auto import get_values
|
| 21 |
+
from transformers.models.big_bird.tokenization_big_bird import BigBirdTokenizer
|
| 22 |
+
from transformers.testing_utils import require_torch, slow, torch_device
|
| 23 |
+
|
| 24 |
+
from ...test_configuration_common import ConfigTester
|
| 25 |
+
from ...test_modeling_common import ModelTesterMixin, floats_tensor, ids_tensor, random_attention_mask
|
| 26 |
+
from ...test_pipeline_mixin import PipelineTesterMixin
|
| 27 |
+
|
| 28 |
+
|
| 29 |
+
if is_torch_available():
|
| 30 |
+
import torch
|
| 31 |
+
|
| 32 |
+
from transformers import (
|
| 33 |
+
MODEL_FOR_PRETRAINING_MAPPING,
|
| 34 |
+
BigBirdForCausalLM,
|
| 35 |
+
BigBirdForMaskedLM,
|
| 36 |
+
BigBirdForMultipleChoice,
|
| 37 |
+
BigBirdForPreTraining,
|
| 38 |
+
BigBirdForQuestionAnswering,
|
| 39 |
+
BigBirdForSequenceClassification,
|
| 40 |
+
BigBirdForTokenClassification,
|
| 41 |
+
BigBirdModel,
|
| 42 |
+
)
|
| 43 |
+
|
| 44 |
+
|
| 45 |
+
class BigBirdModelTester:
|
| 46 |
+
def __init__(
|
| 47 |
+
self,
|
| 48 |
+
parent,
|
| 49 |
+
batch_size=7,
|
| 50 |
+
seq_length=128,
|
| 51 |
+
is_training=True,
|
| 52 |
+
use_input_mask=True,
|
| 53 |
+
use_token_type_ids=True,
|
| 54 |
+
use_labels=True,
|
| 55 |
+
vocab_size=99,
|
| 56 |
+
hidden_size=32,
|
| 57 |
+
num_hidden_layers=2,
|
| 58 |
+
num_attention_heads=4,
|
| 59 |
+
intermediate_size=37,
|
| 60 |
+
hidden_act="gelu_new",
|
| 61 |
+
hidden_dropout_prob=0.1,
|
| 62 |
+
attention_probs_dropout_prob=0.1,
|
| 63 |
+
max_position_embeddings=256,
|
| 64 |
+
type_vocab_size=16,
|
| 65 |
+
type_sequence_label_size=2,
|
| 66 |
+
initializer_range=0.02,
|
| 67 |
+
num_labels=3,
|
| 68 |
+
num_choices=4,
|
| 69 |
+
attention_type="block_sparse",
|
| 70 |
+
use_bias=True,
|
| 71 |
+
rescale_embeddings=False,
|
| 72 |
+
block_size=8,
|
| 73 |
+
num_rand_blocks=3,
|
| 74 |
+
position_embedding_type="absolute",
|
| 75 |
+
scope=None,
|
| 76 |
+
):
|
| 77 |
+
self.parent = parent
|
| 78 |
+
self.batch_size = batch_size
|
| 79 |
+
self.seq_length = seq_length
|
| 80 |
+
self.is_training = is_training
|
| 81 |
+
self.use_input_mask = use_input_mask
|
| 82 |
+
self.use_token_type_ids = use_token_type_ids
|
| 83 |
+
self.use_labels = use_labels
|
| 84 |
+
self.vocab_size = vocab_size
|
| 85 |
+
self.hidden_size = hidden_size
|
| 86 |
+
self.num_hidden_layers = num_hidden_layers
|
| 87 |
+
self.num_attention_heads = num_attention_heads
|
| 88 |
+
self.intermediate_size = intermediate_size
|
| 89 |
+
self.hidden_act = hidden_act
|
| 90 |
+
self.hidden_dropout_prob = hidden_dropout_prob
|
| 91 |
+
self.attention_probs_dropout_prob = attention_probs_dropout_prob
|
| 92 |
+
self.max_position_embeddings = max_position_embeddings
|
| 93 |
+
self.type_vocab_size = type_vocab_size
|
| 94 |
+
self.type_sequence_label_size = type_sequence_label_size
|
| 95 |
+
self.initializer_range = initializer_range
|
| 96 |
+
self.num_labels = num_labels
|
| 97 |
+
self.num_choices = num_choices
|
| 98 |
+
self.scope = scope
|
| 99 |
+
|
| 100 |
+
self.attention_type = attention_type
|
| 101 |
+
self.use_bias = use_bias
|
| 102 |
+
self.rescale_embeddings = rescale_embeddings
|
| 103 |
+
self.block_size = block_size
|
| 104 |
+
self.num_rand_blocks = num_rand_blocks
|
| 105 |
+
self.position_embedding_type = position_embedding_type
|
| 106 |
+
|
| 107 |
+
def prepare_config_and_inputs(self):
|
| 108 |
+
input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
|
| 109 |
+
|
| 110 |
+
input_mask = None
|
| 111 |
+
if self.use_input_mask:
|
| 112 |
+
input_mask = random_attention_mask([self.batch_size, self.seq_length])
|
| 113 |
+
|
| 114 |
+
token_type_ids = None
|
| 115 |
+
if self.use_token_type_ids:
|
| 116 |
+
token_type_ids = ids_tensor([self.batch_size, self.seq_length], self.type_vocab_size)
|
| 117 |
+
|
| 118 |
+
sequence_labels = None
|
| 119 |
+
token_labels = None
|
| 120 |
+
choice_labels = None
|
| 121 |
+
if self.use_labels:
|
| 122 |
+
sequence_labels = ids_tensor([self.batch_size], self.type_sequence_label_size)
|
| 123 |
+
token_labels = ids_tensor([self.batch_size, self.seq_length], self.num_labels)
|
| 124 |
+
choice_labels = ids_tensor([self.batch_size], self.num_choices)
|
| 125 |
+
|
| 126 |
+
config = self.get_config()
|
| 127 |
+
|
| 128 |
+
return config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
|
| 129 |
+
|
| 130 |
+
def get_config(self):
|
| 131 |
+
return BigBirdConfig(
|
| 132 |
+
vocab_size=self.vocab_size,
|
| 133 |
+
hidden_size=self.hidden_size,
|
| 134 |
+
num_hidden_layers=self.num_hidden_layers,
|
| 135 |
+
num_attention_heads=self.num_attention_heads,
|
| 136 |
+
intermediate_size=self.intermediate_size,
|
| 137 |
+
hidden_act=self.hidden_act,
|
| 138 |
+
hidden_dropout_prob=self.hidden_dropout_prob,
|
| 139 |
+
attention_probs_dropout_prob=self.attention_probs_dropout_prob,
|
| 140 |
+
max_position_embeddings=self.max_position_embeddings,
|
| 141 |
+
type_vocab_size=self.type_vocab_size,
|
| 142 |
+
is_encoder_decoder=False,
|
| 143 |
+
initializer_range=self.initializer_range,
|
| 144 |
+
attention_type=self.attention_type,
|
| 145 |
+
use_bias=self.use_bias,
|
| 146 |
+
rescale_embeddings=self.rescale_embeddings,
|
| 147 |
+
block_size=self.block_size,
|
| 148 |
+
num_random_blocks=self.num_rand_blocks,
|
| 149 |
+
position_embedding_type=self.position_embedding_type,
|
| 150 |
+
)
|
| 151 |
+
|
| 152 |
+
def prepare_config_and_inputs_for_decoder(self):
|
| 153 |
+
(
|
| 154 |
+
config,
|
| 155 |
+
input_ids,
|
| 156 |
+
token_type_ids,
|
| 157 |
+
input_mask,
|
| 158 |
+
sequence_labels,
|
| 159 |
+
token_labels,
|
| 160 |
+
choice_labels,
|
| 161 |
+
) = self.prepare_config_and_inputs()
|
| 162 |
+
|
| 163 |
+
config.is_decoder = True
|
| 164 |
+
encoder_hidden_states = floats_tensor([self.batch_size, self.seq_length, self.hidden_size])
|
| 165 |
+
encoder_attention_mask = ids_tensor([self.batch_size, self.seq_length], vocab_size=2)
|
| 166 |
+
|
| 167 |
+
return (
|
| 168 |
+
config,
|
| 169 |
+
input_ids,
|
| 170 |
+
token_type_ids,
|
| 171 |
+
input_mask,
|
| 172 |
+
sequence_labels,
|
| 173 |
+
token_labels,
|
| 174 |
+
choice_labels,
|
| 175 |
+
encoder_hidden_states,
|
| 176 |
+
encoder_attention_mask,
|
| 177 |
+
)
|
| 178 |
+
|
| 179 |
+
def create_and_check_model(
|
| 180 |
+
self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
|
| 181 |
+
):
|
| 182 |
+
model = BigBirdModel(config=config)
|
| 183 |
+
model.to(torch_device)
|
| 184 |
+
model.eval()
|
| 185 |
+
result = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids)
|
| 186 |
+
result = model(input_ids, token_type_ids=token_type_ids)
|
| 187 |
+
result = model(input_ids)
|
| 188 |
+
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
|
| 189 |
+
|
| 190 |
+
def create_and_check_for_pretraining(
|
| 191 |
+
self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
|
| 192 |
+
):
|
| 193 |
+
model = BigBirdForPreTraining(config=config)
|
| 194 |
+
model.to(torch_device)
|
| 195 |
+
model.eval()
|
| 196 |
+
result = model(
|
| 197 |
+
input_ids,
|
| 198 |
+
attention_mask=input_mask,
|
| 199 |
+
token_type_ids=token_type_ids,
|
| 200 |
+
labels=token_labels,
|
| 201 |
+
next_sentence_label=sequence_labels,
|
| 202 |
+
)
|
| 203 |
+
self.parent.assertEqual(result.prediction_logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
|
| 204 |
+
self.parent.assertEqual(result.seq_relationship_logits.shape, (self.batch_size, config.num_labels))
|
| 205 |
+
|
| 206 |
+
def create_and_check_model_as_decoder(
|
| 207 |
+
self,
|
| 208 |
+
config,
|
| 209 |
+
input_ids,
|
| 210 |
+
token_type_ids,
|
| 211 |
+
input_mask,
|
| 212 |
+
sequence_labels,
|
| 213 |
+
token_labels,
|
| 214 |
+
choice_labels,
|
| 215 |
+
encoder_hidden_states,
|
| 216 |
+
encoder_attention_mask,
|
| 217 |
+
):
|
| 218 |
+
config.add_cross_attention = True
|
| 219 |
+
model = BigBirdModel(config)
|
| 220 |
+
model.to(torch_device)
|
| 221 |
+
model.eval()
|
| 222 |
+
result = model(
|
| 223 |
+
input_ids,
|
| 224 |
+
attention_mask=input_mask,
|
| 225 |
+
token_type_ids=token_type_ids,
|
| 226 |
+
encoder_hidden_states=encoder_hidden_states,
|
| 227 |
+
encoder_attention_mask=encoder_attention_mask,
|
| 228 |
+
)
|
| 229 |
+
result = model(
|
| 230 |
+
input_ids,
|
| 231 |
+
attention_mask=input_mask,
|
| 232 |
+
token_type_ids=token_type_ids,
|
| 233 |
+
encoder_hidden_states=encoder_hidden_states,
|
| 234 |
+
)
|
| 235 |
+
result = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids)
|
| 236 |
+
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
|
| 237 |
+
|
| 238 |
+
def create_and_check_for_causal_lm(
|
| 239 |
+
self,
|
| 240 |
+
config,
|
| 241 |
+
input_ids,
|
| 242 |
+
token_type_ids,
|
| 243 |
+
input_mask,
|
| 244 |
+
sequence_labels,
|
| 245 |
+
token_labels,
|
| 246 |
+
choice_labels,
|
| 247 |
+
encoder_hidden_states,
|
| 248 |
+
encoder_attention_mask,
|
| 249 |
+
):
|
| 250 |
+
model = BigBirdForCausalLM(config=config)
|
| 251 |
+
model.to(torch_device)
|
| 252 |
+
model.eval()
|
| 253 |
+
result = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids, labels=token_labels)
|
| 254 |
+
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
|
| 255 |
+
|
| 256 |
+
def create_and_check_for_masked_lm(
|
| 257 |
+
self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
|
| 258 |
+
):
|
| 259 |
+
model = BigBirdForMaskedLM(config=config)
|
| 260 |
+
model.to(torch_device)
|
| 261 |
+
model.eval()
|
| 262 |
+
result = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids, labels=token_labels)
|
| 263 |
+
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
|
| 264 |
+
|
| 265 |
+
def create_and_check_decoder_model_past_large_inputs(
|
| 266 |
+
self,
|
| 267 |
+
config,
|
| 268 |
+
input_ids,
|
| 269 |
+
token_type_ids,
|
| 270 |
+
input_mask,
|
| 271 |
+
sequence_labels,
|
| 272 |
+
token_labels,
|
| 273 |
+
choice_labels,
|
| 274 |
+
encoder_hidden_states,
|
| 275 |
+
encoder_attention_mask,
|
| 276 |
+
):
|
| 277 |
+
config.is_decoder = True
|
| 278 |
+
config.add_cross_attention = True
|
| 279 |
+
model = BigBirdForCausalLM(config=config)
|
| 280 |
+
model.to(torch_device)
|
| 281 |
+
model.eval()
|
| 282 |
+
|
| 283 |
+
# first forward pass
|
| 284 |
+
outputs = model(
|
| 285 |
+
input_ids,
|
| 286 |
+
attention_mask=input_mask,
|
| 287 |
+
encoder_hidden_states=encoder_hidden_states,
|
| 288 |
+
encoder_attention_mask=encoder_attention_mask,
|
| 289 |
+
use_cache=True,
|
| 290 |
+
)
|
| 291 |
+
past_key_values = outputs.past_key_values
|
| 292 |
+
|
| 293 |
+
# create hypothetical multiple next token and extent to next_input_ids
|
| 294 |
+
next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size)
|
| 295 |
+
next_mask = ids_tensor((self.batch_size, 3), vocab_size=2)
|
| 296 |
+
|
| 297 |
+
# append to next input_ids and
|
| 298 |
+
next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
|
| 299 |
+
next_attention_mask = torch.cat([input_mask, next_mask], dim=-1)
|
| 300 |
+
|
| 301 |
+
output_from_no_past = model(
|
| 302 |
+
next_input_ids,
|
| 303 |
+
attention_mask=next_attention_mask,
|
| 304 |
+
encoder_hidden_states=encoder_hidden_states,
|
| 305 |
+
encoder_attention_mask=encoder_attention_mask,
|
| 306 |
+
output_hidden_states=True,
|
| 307 |
+
)["hidden_states"][0]
|
| 308 |
+
output_from_past = model(
|
| 309 |
+
next_tokens,
|
| 310 |
+
attention_mask=next_attention_mask,
|
| 311 |
+
encoder_hidden_states=encoder_hidden_states,
|
| 312 |
+
encoder_attention_mask=encoder_attention_mask,
|
| 313 |
+
past_key_values=past_key_values,
|
| 314 |
+
output_hidden_states=True,
|
| 315 |
+
)["hidden_states"][0]
|
| 316 |
+
|
| 317 |
+
# select random slice
|
| 318 |
+
random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()
|
| 319 |
+
output_from_no_past_slice = output_from_no_past[:, -3:, random_slice_idx].detach()
|
| 320 |
+
output_from_past_slice = output_from_past[:, :, random_slice_idx].detach()
|
| 321 |
+
|
| 322 |
+
self.parent.assertTrue(output_from_past_slice.shape[1] == next_tokens.shape[1])
|
| 323 |
+
|
| 324 |
+
# test that outputs are equal for slice
|
| 325 |
+
self.parent.assertTrue(torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3))
|
| 326 |
+
|
| 327 |
+
def create_and_check_for_question_answering(
|
| 328 |
+
self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
|
| 329 |
+
):
|
| 330 |
+
model = BigBirdForQuestionAnswering(config=config)
|
| 331 |
+
model.to(torch_device)
|
| 332 |
+
model.eval()
|
| 333 |
+
result = model(
|
| 334 |
+
input_ids,
|
| 335 |
+
attention_mask=input_mask,
|
| 336 |
+
token_type_ids=token_type_ids,
|
| 337 |
+
start_positions=sequence_labels,
|
| 338 |
+
end_positions=sequence_labels,
|
| 339 |
+
)
|
| 340 |
+
self.parent.assertEqual(result.start_logits.shape, (self.batch_size, self.seq_length))
|
| 341 |
+
self.parent.assertEqual(result.end_logits.shape, (self.batch_size, self.seq_length))
|
| 342 |
+
|
| 343 |
+
def create_and_check_for_sequence_classification(
|
| 344 |
+
self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
|
| 345 |
+
):
|
| 346 |
+
config.num_labels = self.num_labels
|
| 347 |
+
model = BigBirdForSequenceClassification(config)
|
| 348 |
+
model.to(torch_device)
|
| 349 |
+
model.eval()
|
| 350 |
+
result = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids, labels=sequence_labels)
|
| 351 |
+
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.num_labels))
|
| 352 |
+
|
| 353 |
+
def create_and_check_for_token_classification(
|
| 354 |
+
self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
|
| 355 |
+
):
|
| 356 |
+
config.num_labels = self.num_labels
|
| 357 |
+
model = BigBirdForTokenClassification(config=config)
|
| 358 |
+
model.to(torch_device)
|
| 359 |
+
model.eval()
|
| 360 |
+
result = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids, labels=token_labels)
|
| 361 |
+
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.num_labels))
|
| 362 |
+
|
| 363 |
+
def create_and_check_for_multiple_choice(
|
| 364 |
+
self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
|
| 365 |
+
):
|
| 366 |
+
config.num_choices = self.num_choices
|
| 367 |
+
model = BigBirdForMultipleChoice(config=config)
|
| 368 |
+
model.to(torch_device)
|
| 369 |
+
model.eval()
|
| 370 |
+
multiple_choice_inputs_ids = input_ids.unsqueeze(1).expand(-1, self.num_choices, -1).contiguous()
|
| 371 |
+
multiple_choice_token_type_ids = token_type_ids.unsqueeze(1).expand(-1, self.num_choices, -1).contiguous()
|
| 372 |
+
multiple_choice_input_mask = input_mask.unsqueeze(1).expand(-1, self.num_choices, -1).contiguous()
|
| 373 |
+
result = model(
|
| 374 |
+
multiple_choice_inputs_ids,
|
| 375 |
+
attention_mask=multiple_choice_input_mask,
|
| 376 |
+
token_type_ids=multiple_choice_token_type_ids,
|
| 377 |
+
labels=choice_labels,
|
| 378 |
+
)
|
| 379 |
+
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.num_choices))
|
| 380 |
+
|
| 381 |
+
def prepare_config_and_inputs_for_common(self):
|
| 382 |
+
config_and_inputs = self.prepare_config_and_inputs()
|
| 383 |
+
(
|
| 384 |
+
config,
|
| 385 |
+
input_ids,
|
| 386 |
+
token_type_ids,
|
| 387 |
+
input_mask,
|
| 388 |
+
sequence_labels,
|
| 389 |
+
token_labels,
|
| 390 |
+
choice_labels,
|
| 391 |
+
) = config_and_inputs
|
| 392 |
+
inputs_dict = {"input_ids": input_ids, "token_type_ids": token_type_ids, "attention_mask": input_mask}
|
| 393 |
+
return config, inputs_dict
|
| 394 |
+
|
| 395 |
+
def create_and_check_for_auto_padding(
|
| 396 |
+
self,
|
| 397 |
+
config,
|
| 398 |
+
input_ids,
|
| 399 |
+
token_type_ids,
|
| 400 |
+
input_mask,
|
| 401 |
+
sequence_labels,
|
| 402 |
+
token_labels,
|
| 403 |
+
choice_labels,
|
| 404 |
+
):
|
| 405 |
+
model = BigBirdModel(config)
|
| 406 |
+
model.to(torch_device)
|
| 407 |
+
model.eval()
|
| 408 |
+
result = model(input_ids)
|
| 409 |
+
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
|
| 410 |
+
|
| 411 |
+
def create_and_check_for_change_to_full_attn(
|
| 412 |
+
self,
|
| 413 |
+
config,
|
| 414 |
+
input_ids,
|
| 415 |
+
token_type_ids,
|
| 416 |
+
input_mask,
|
| 417 |
+
sequence_labels,
|
| 418 |
+
token_labels,
|
| 419 |
+
choice_labels,
|
| 420 |
+
):
|
| 421 |
+
model = BigBirdModel(config)
|
| 422 |
+
model.to(torch_device)
|
| 423 |
+
model.eval()
|
| 424 |
+
result = model(input_ids)
|
| 425 |
+
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
|
| 426 |
+
# the config should not be changed
|
| 427 |
+
self.parent.assertTrue(model.config.attention_type == "block_sparse")
|
| 428 |
+
|
| 429 |
+
|
| 430 |
+
@require_torch
|
| 431 |
+
class BigBirdModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase):
|
| 432 |
+
# head masking & pruning is currently not supported for big bird
|
| 433 |
+
test_head_masking = False
|
| 434 |
+
test_pruning = False
|
| 435 |
+
|
| 436 |
+
# torchscript should be possible, but takes prohibitively long to test.
|
| 437 |
+
# Also torchscript is not an important feature to have in the beginning.
|
| 438 |
+
test_torchscript = False
|
| 439 |
+
|
| 440 |
+
all_model_classes = (
|
| 441 |
+
(
|
| 442 |
+
BigBirdModel,
|
| 443 |
+
BigBirdForPreTraining,
|
| 444 |
+
BigBirdForMaskedLM,
|
| 445 |
+
BigBirdForCausalLM,
|
| 446 |
+
BigBirdForMultipleChoice,
|
| 447 |
+
BigBirdForQuestionAnswering,
|
| 448 |
+
BigBirdForSequenceClassification,
|
| 449 |
+
BigBirdForTokenClassification,
|
| 450 |
+
)
|
| 451 |
+
if is_torch_available()
|
| 452 |
+
else ()
|
| 453 |
+
)
|
| 454 |
+
all_generative_model_classes = (BigBirdForCausalLM,) if is_torch_available() else ()
|
| 455 |
+
pipeline_model_mapping = (
|
| 456 |
+
{
|
| 457 |
+
"feature-extraction": BigBirdModel,
|
| 458 |
+
"fill-mask": BigBirdForMaskedLM,
|
| 459 |
+
"question-answering": BigBirdForQuestionAnswering,
|
| 460 |
+
"text-classification": BigBirdForSequenceClassification,
|
| 461 |
+
"text-generation": BigBirdForCausalLM,
|
| 462 |
+
"token-classification": BigBirdForTokenClassification,
|
| 463 |
+
"zero-shot": BigBirdForSequenceClassification,
|
| 464 |
+
}
|
| 465 |
+
if is_torch_available()
|
| 466 |
+
else {}
|
| 467 |
+
)
|
| 468 |
+
|
| 469 |
+
# special case for ForPreTraining model
|
| 470 |
+
def _prepare_for_class(self, inputs_dict, model_class, return_labels=False):
|
| 471 |
+
inputs_dict = super()._prepare_for_class(inputs_dict, model_class, return_labels=return_labels)
|
| 472 |
+
|
| 473 |
+
if return_labels:
|
| 474 |
+
if model_class in get_values(MODEL_FOR_PRETRAINING_MAPPING):
|
| 475 |
+
inputs_dict["labels"] = torch.zeros(
|
| 476 |
+
(self.model_tester.batch_size, self.model_tester.seq_length), dtype=torch.long, device=torch_device
|
| 477 |
+
)
|
| 478 |
+
inputs_dict["next_sentence_label"] = torch.zeros(
|
| 479 |
+
self.model_tester.batch_size, dtype=torch.long, device=torch_device
|
| 480 |
+
)
|
| 481 |
+
return inputs_dict
|
| 482 |
+
|
| 483 |
+
def setUp(self):
|
| 484 |
+
self.model_tester = BigBirdModelTester(self)
|
| 485 |
+
self.config_tester = ConfigTester(self, config_class=BigBirdConfig, hidden_size=37)
|
| 486 |
+
|
| 487 |
+
def test_config(self):
|
| 488 |
+
self.config_tester.run_common_tests()
|
| 489 |
+
|
| 490 |
+
def test_model(self):
|
| 491 |
+
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
| 492 |
+
self.model_tester.create_and_check_model(*config_and_inputs)
|
| 493 |
+
|
| 494 |
+
def test_for_pretraining(self):
|
| 495 |
+
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
| 496 |
+
self.model_tester.create_and_check_for_pretraining(*config_and_inputs)
|
| 497 |
+
|
| 498 |
+
def test_for_masked_lm(self):
|
| 499 |
+
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
| 500 |
+
self.model_tester.create_and_check_for_masked_lm(*config_and_inputs)
|
| 501 |
+
|
| 502 |
+
def test_for_multiple_choice(self):
|
| 503 |
+
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
| 504 |
+
self.model_tester.create_and_check_for_multiple_choice(*config_and_inputs)
|
| 505 |
+
|
| 506 |
+
def test_decoder_model_past_with_large_inputs(self):
|
| 507 |
+
config_and_inputs = self.model_tester.prepare_config_and_inputs_for_decoder()
|
| 508 |
+
self.model_tester.create_and_check_decoder_model_past_large_inputs(*config_and_inputs)
|
| 509 |
+
|
| 510 |
+
def test_for_question_answering(self):
|
| 511 |
+
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
| 512 |
+
self.model_tester.create_and_check_for_question_answering(*config_and_inputs)
|
| 513 |
+
|
| 514 |
+
def test_for_sequence_classification(self):
|
| 515 |
+
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
| 516 |
+
self.model_tester.create_and_check_for_sequence_classification(*config_and_inputs)
|
| 517 |
+
|
| 518 |
+
def test_for_token_classification(self):
|
| 519 |
+
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
| 520 |
+
self.model_tester.create_and_check_for_token_classification(*config_and_inputs)
|
| 521 |
+
|
| 522 |
+
def test_model_as_decoder(self):
|
| 523 |
+
config_and_inputs = self.model_tester.prepare_config_and_inputs_for_decoder()
|
| 524 |
+
self.model_tester.create_and_check_model_as_decoder(*config_and_inputs)
|
| 525 |
+
|
| 526 |
+
def test_model_as_decoder_with_default_input_mask(self):
|
| 527 |
+
# This regression test was failing with PyTorch < 1.3
|
| 528 |
+
(
|
| 529 |
+
config,
|
| 530 |
+
input_ids,
|
| 531 |
+
token_type_ids,
|
| 532 |
+
input_mask,
|
| 533 |
+
sequence_labels,
|
| 534 |
+
token_labels,
|
| 535 |
+
choice_labels,
|
| 536 |
+
encoder_hidden_states,
|
| 537 |
+
encoder_attention_mask,
|
| 538 |
+
) = self.model_tester.prepare_config_and_inputs_for_decoder()
|
| 539 |
+
|
| 540 |
+
input_mask = None
|
| 541 |
+
|
| 542 |
+
self.model_tester.create_and_check_model_as_decoder(
|
| 543 |
+
config,
|
| 544 |
+
input_ids,
|
| 545 |
+
token_type_ids,
|
| 546 |
+
input_mask,
|
| 547 |
+
sequence_labels,
|
| 548 |
+
token_labels,
|
| 549 |
+
choice_labels,
|
| 550 |
+
encoder_hidden_states,
|
| 551 |
+
encoder_attention_mask,
|
| 552 |
+
)
|
| 553 |
+
|
| 554 |
+
def test_retain_grad_hidden_states_attentions(self):
|
| 555 |
+
# bigbird cannot keep gradients in attentions when `attention_type=block_sparse`
|
| 556 |
+
|
| 557 |
+
if self.model_tester.attention_type == "original_full":
|
| 558 |
+
super().test_retain_grad_hidden_states_attentions()
|
| 559 |
+
|
| 560 |
+
@slow
|
| 561 |
+
def test_model_from_pretrained(self):
|
| 562 |
+
model_name = "google/bigbird-roberta-base"
|
| 563 |
+
model = BigBirdForPreTraining.from_pretrained(model_name)
|
| 564 |
+
self.assertIsNotNone(model)
|
| 565 |
+
|
| 566 |
+
def test_model_various_attn_type(self):
|
| 567 |
+
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
| 568 |
+
for type in ["original_full", "block_sparse"]:
|
| 569 |
+
config_and_inputs[0].attention_type = type
|
| 570 |
+
self.model_tester.create_and_check_model(*config_and_inputs)
|
| 571 |
+
|
| 572 |
+
def test_fast_integration(self):
|
| 573 |
+
# fmt: off
|
| 574 |
+
input_ids = torch.tensor(
|
| 575 |
+
[[6, 117, 33, 36, 70, 22, 63, 31, 71, 72, 88, 58, 109, 49, 48, 116, 92, 6, 19, 95, 118, 100, 80, 111, 93, 2, 31, 84, 26, 5, 6, 82, 46, 96, 109, 4, 39, 19, 109, 13, 92, 31, 36, 90, 111, 18, 75, 6, 56, 74, 16, 42, 56, 92, 69, 108, 127, 81, 82, 41, 106, 19, 44, 24, 82, 121, 120, 65, 36, 26, 72, 13, 36, 98, 43, 64, 8, 53, 100, 92, 51, 122, 66, 17, 61, 50, 104, 127, 26, 35, 94, 23, 110, 71, 80, 67, 109, 111, 44, 19, 51, 41, 86, 71, 76, 44, 18, 68, 44, 77, 107, 81, 98, 126, 100, 2, 49, 98, 84, 39, 23, 98, 52, 46, 10, 82, 121, 73],[6, 117, 33, 36, 70, 22, 63, 31, 71, 72, 88, 58, 109, 49, 48, 116, 92, 6, 19, 95, 118, 100, 80, 111, 93, 2, 31, 84, 26, 5, 6, 82, 46, 96, 109, 4, 39, 19, 109, 13, 92, 31, 36, 90, 111, 18, 75, 6, 56, 74, 16, 42, 56, 92, 69, 108, 127, 81, 82, 41, 106, 19, 44, 24, 82, 121, 120, 65, 36, 26, 72, 13, 36, 98, 43, 64, 8, 53, 100, 92, 51, 12, 66, 17, 61, 50, 104, 127, 26, 35, 94, 23, 110, 71, 80, 67, 109, 111, 44, 19, 51, 41, 86, 71, 76, 28, 18, 68, 44, 77, 107, 81, 98, 126, 100, 2, 49, 18, 84, 39, 23, 98, 52, 46, 10, 82, 121, 73]], # noqa: E231
|
| 576 |
+
dtype=torch.long,
|
| 577 |
+
device=torch_device,
|
| 578 |
+
)
|
| 579 |
+
# fmt: on
|
| 580 |
+
input_ids = input_ids % self.model_tester.vocab_size
|
| 581 |
+
input_ids[1] = input_ids[1] - 1
|
| 582 |
+
|
| 583 |
+
attention_mask = torch.ones((input_ids.shape), device=torch_device)
|
| 584 |
+
attention_mask[:, :-10] = 0
|
| 585 |
+
|
| 586 |
+
config, _, _, _, _, _, _ = self.model_tester.prepare_config_and_inputs()
|
| 587 |
+
torch.manual_seed(0)
|
| 588 |
+
model = BigBirdModel(config).eval().to(torch_device)
|
| 589 |
+
|
| 590 |
+
with torch.no_grad():
|
| 591 |
+
hidden_states = model(input_ids, attention_mask=attention_mask).last_hidden_state
|
| 592 |
+
self.assertTrue(
|
| 593 |
+
torch.allclose(
|
| 594 |
+
hidden_states[0, 0, :5],
|
| 595 |
+
torch.tensor([1.4825, 0.0774, 0.8226, -0.2962, -0.9593], device=torch_device),
|
| 596 |
+
atol=1e-3,
|
| 597 |
+
)
|
| 598 |
+
)
|
| 599 |
+
|
| 600 |
+
def test_auto_padding(self):
|
| 601 |
+
self.model_tester.seq_length = 241
|
| 602 |
+
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
| 603 |
+
self.model_tester.create_and_check_for_auto_padding(*config_and_inputs)
|
| 604 |
+
|
| 605 |
+
def test_for_change_to_full_attn(self):
|
| 606 |
+
self.model_tester.seq_length = 9
|
| 607 |
+
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
| 608 |
+
self.model_tester.create_and_check_for_change_to_full_attn(*config_and_inputs)
|
| 609 |
+
|
| 610 |
+
@unittest.skip(
|
| 611 |
+
reason="This architecure seem to not compute gradients properly when using GC, check: https://github.com/huggingface/transformers/pull/27124"
|
| 612 |
+
)
|
| 613 |
+
def test_training_gradient_checkpointing(self):
|
| 614 |
+
pass
|
| 615 |
+
|
| 616 |
+
@unittest.skip(
|
| 617 |
+
reason="This architecure seem to not compute gradients properly when using GC, check: https://github.com/huggingface/transformers/pull/27124"
|
| 618 |
+
)
|
| 619 |
+
def test_training_gradient_checkpointing_use_reentrant(self):
|
| 620 |
+
pass
|
| 621 |
+
|
| 622 |
+
@unittest.skip(
|
| 623 |
+
reason="This architecure seem to not compute gradients properly when using GC, check: https://github.com/huggingface/transformers/pull/27124"
|
| 624 |
+
)
|
| 625 |
+
def test_training_gradient_checkpointing_use_reentrant_false(self):
|
| 626 |
+
pass
|
| 627 |
+
|
| 628 |
+
# overwrite from common in order to skip the check on `attentions`
|
| 629 |
+
def check_pt_flax_outputs(self, fx_outputs, pt_outputs, model_class, tol=1e-5, name="outputs", attributes=None):
|
| 630 |
+
# `bigbird_block_sparse_attention` in `FlaxBigBird` returns `attention_probs = None`, while in PyTorch version,
|
| 631 |
+
# an effort was done to return `attention_probs` (yet to be verified).
|
| 632 |
+
if name.startswith("outputs.attentions"):
|
| 633 |
+
return
|
| 634 |
+
else:
|
| 635 |
+
super().check_pt_flax_outputs(fx_outputs, pt_outputs, model_class, tol, name, attributes)
|
| 636 |
+
|
| 637 |
+
|
| 638 |
+
@require_torch
|
| 639 |
+
@slow
|
| 640 |
+
class BigBirdModelIntegrationTest(unittest.TestCase):
|
| 641 |
+
# we can have this true once block_sparse attn_probs works accurately
|
| 642 |
+
test_attention_probs = False
|
| 643 |
+
|
| 644 |
+
def _get_dummy_input_ids(self):
|
| 645 |
+
# fmt: off
|
| 646 |
+
ids = torch.tensor(
|
| 647 |
+
[[6, 117, 33, 36, 70, 22, 63, 31, 71, 72, 88, 58, 109, 49, 48, 116, 92, 6, 19, 95, 118, 100, 80, 111, 93, 2, 31, 84, 26, 5, 6, 82, 46, 96, 109, 4, 39, 19, 109, 13, 92, 31, 36, 90, 111, 18, 75, 6, 56, 74, 16, 42, 56, 92, 69, 108, 127, 81, 82, 41, 106, 19, 44, 24, 82, 121, 120, 65, 36, 26, 72, 13, 36, 98, 43, 64, 8, 53, 100, 92, 51, 122, 66, 17, 61, 50, 104, 127, 26, 35, 94, 23, 110, 71, 80, 67, 109, 111, 44, 19, 51, 41, 86, 71, 76, 44, 18, 68, 44, 77, 107, 81, 98, 126, 100, 2, 49, 98, 84, 39, 23, 98, 52, 46, 10, 82, 121, 73]], # noqa: E231
|
| 648 |
+
dtype=torch.long,
|
| 649 |
+
device=torch_device,
|
| 650 |
+
)
|
| 651 |
+
# fmt: on
|
| 652 |
+
return ids
|
| 653 |
+
|
| 654 |
+
def test_inference_block_sparse_pretraining(self):
|
| 655 |
+
model = BigBirdForPreTraining.from_pretrained("google/bigbird-roberta-base", attention_type="block_sparse")
|
| 656 |
+
model.to(torch_device)
|
| 657 |
+
|
| 658 |
+
input_ids = torch.tensor([[20920, 232, 328, 1437] * 1024], dtype=torch.long, device=torch_device)
|
| 659 |
+
with torch.no_grad():
|
| 660 |
+
outputs = model(input_ids)
|
| 661 |
+
prediction_logits = outputs.prediction_logits
|
| 662 |
+
seq_relationship_logits = outputs.seq_relationship_logits
|
| 663 |
+
|
| 664 |
+
self.assertEqual(prediction_logits.shape, torch.Size((1, 4096, 50358)))
|
| 665 |
+
self.assertEqual(seq_relationship_logits.shape, torch.Size((1, 2)))
|
| 666 |
+
|
| 667 |
+
expected_prediction_logits_slice = torch.tensor(
|
| 668 |
+
[
|
| 669 |
+
[-0.5583, 0.0475, -0.2508, 7.4423],
|
| 670 |
+
[0.7409, 1.4460, -0.7593, 7.7010],
|
| 671 |
+
[1.9150, 3.1395, 5.8840, 9.3498],
|
| 672 |
+
[-0.1854, -1.4640, -2.2052, 3.7968],
|
| 673 |
+
],
|
| 674 |
+
device=torch_device,
|
| 675 |
+
)
|
| 676 |
+
|
| 677 |
+
self.assertTrue(
|
| 678 |
+
torch.allclose(prediction_logits[0, 128:132, 128:132], expected_prediction_logits_slice, atol=1e-4)
|
| 679 |
+
)
|
| 680 |
+
|
| 681 |
+
expected_seq_relationship_logits = torch.tensor([[46.9465, 47.9517]], device=torch_device)
|
| 682 |
+
self.assertTrue(torch.allclose(seq_relationship_logits, expected_seq_relationship_logits, atol=1e-4))
|
| 683 |
+
|
| 684 |
+
def test_inference_full_pretraining(self):
|
| 685 |
+
model = BigBirdForPreTraining.from_pretrained("google/bigbird-roberta-base", attention_type="original_full")
|
| 686 |
+
model.to(torch_device)
|
| 687 |
+
|
| 688 |
+
input_ids = torch.tensor([[20920, 232, 328, 1437] * 512], dtype=torch.long, device=torch_device)
|
| 689 |
+
with torch.no_grad():
|
| 690 |
+
outputs = model(input_ids)
|
| 691 |
+
prediction_logits = outputs.prediction_logits
|
| 692 |
+
seq_relationship_logits = outputs.seq_relationship_logits
|
| 693 |
+
|
| 694 |
+
self.assertEqual(prediction_logits.shape, torch.Size((1, 512 * 4, 50358)))
|
| 695 |
+
self.assertEqual(seq_relationship_logits.shape, torch.Size((1, 2)))
|
| 696 |
+
|
| 697 |
+
expected_prediction_logits_slice = torch.tensor(
|
| 698 |
+
[
|
| 699 |
+
[0.1499, -1.1217, 0.1990, 8.4499],
|
| 700 |
+
[-2.7757, -3.0687, -4.8577, 7.5156],
|
| 701 |
+
[1.5446, 0.1982, 4.3016, 10.4281],
|
| 702 |
+
[-1.3705, -4.0130, -3.9629, 5.1526],
|
| 703 |
+
],
|
| 704 |
+
device=torch_device,
|
| 705 |
+
)
|
| 706 |
+
self.assertTrue(
|
| 707 |
+
torch.allclose(prediction_logits[0, 128:132, 128:132], expected_prediction_logits_slice, atol=1e-4)
|
| 708 |
+
)
|
| 709 |
+
|
| 710 |
+
expected_seq_relationship_logits = torch.tensor([[41.4503, 41.2406]], device=torch_device)
|
| 711 |
+
self.assertTrue(torch.allclose(seq_relationship_logits, expected_seq_relationship_logits, atol=1e-4))
|
| 712 |
+
|
| 713 |
+
def test_block_sparse_attention_probs(self):
|
| 714 |
+
"""
|
| 715 |
+
Asserting if outputted attention matrix is similar to hard coded attention matrix
|
| 716 |
+
"""
|
| 717 |
+
|
| 718 |
+
if not self.test_attention_probs:
|
| 719 |
+
self.skipTest("test_attention_probs is set to False")
|
| 720 |
+
|
| 721 |
+
model = BigBirdModel.from_pretrained(
|
| 722 |
+
"google/bigbird-roberta-base", attention_type="block_sparse", num_random_blocks=3, block_size=16
|
| 723 |
+
)
|
| 724 |
+
model.to(torch_device)
|
| 725 |
+
model.eval()
|
| 726 |
+
config = model.config
|
| 727 |
+
|
| 728 |
+
input_ids = self._get_dummy_input_ids()
|
| 729 |
+
|
| 730 |
+
hidden_states = model.embeddings(input_ids)
|
| 731 |
+
|
| 732 |
+
batch_size, seqlen, _ = hidden_states.size()
|
| 733 |
+
attn_mask = torch.ones(batch_size, seqlen, device=torch_device, dtype=torch.float)
|
| 734 |
+
to_seq_length = from_seq_length = seqlen
|
| 735 |
+
from_block_size = to_block_size = config.block_size
|
| 736 |
+
|
| 737 |
+
blocked_mask, band_mask, from_mask, to_mask = model.create_masks_for_block_sparse_attn(
|
| 738 |
+
attn_mask, config.block_size
|
| 739 |
+
)
|
| 740 |
+
from_blocked_mask = to_blocked_mask = blocked_mask
|
| 741 |
+
|
| 742 |
+
for i in range(config.num_hidden_layers):
|
| 743 |
+
pointer = model.encoder.layer[i].attention.self
|
| 744 |
+
|
| 745 |
+
query_layer = pointer.transpose_for_scores(pointer.query(hidden_states))
|
| 746 |
+
key_layer = pointer.transpose_for_scores(pointer.key(hidden_states))
|
| 747 |
+
value_layer = pointer.transpose_for_scores(pointer.value(hidden_states))
|
| 748 |
+
|
| 749 |
+
context_layer, attention_probs = pointer.bigbird_block_sparse_attention(
|
| 750 |
+
query_layer,
|
| 751 |
+
key_layer,
|
| 752 |
+
value_layer,
|
| 753 |
+
band_mask,
|
| 754 |
+
from_mask,
|
| 755 |
+
to_mask,
|
| 756 |
+
from_blocked_mask,
|
| 757 |
+
to_blocked_mask,
|
| 758 |
+
pointer.num_attention_heads,
|
| 759 |
+
pointer.num_random_blocks,
|
| 760 |
+
pointer.attention_head_size,
|
| 761 |
+
from_block_size,
|
| 762 |
+
to_block_size,
|
| 763 |
+
batch_size,
|
| 764 |
+
from_seq_length,
|
| 765 |
+
to_seq_length,
|
| 766 |
+
seed=pointer.seed,
|
| 767 |
+
plan_from_length=None,
|
| 768 |
+
plan_num_rand_blocks=None,
|
| 769 |
+
output_attentions=True,
|
| 770 |
+
)
|
| 771 |
+
|
| 772 |
+
context_layer = context_layer.contiguous().view(batch_size, from_seq_length, -1)
|
| 773 |
+
cl = torch.einsum("bhqk,bhkd->bhqd", attention_probs, value_layer)
|
| 774 |
+
cl = cl.view(context_layer.size())
|
| 775 |
+
|
| 776 |
+
self.assertTrue(torch.allclose(context_layer, cl, atol=0.001))
|
| 777 |
+
|
| 778 |
+
def test_block_sparse_context_layer(self):
|
| 779 |
+
model = BigBirdModel.from_pretrained(
|
| 780 |
+
"google/bigbird-roberta-base", attention_type="block_sparse", num_random_blocks=3, block_size=16
|
| 781 |
+
)
|
| 782 |
+
model.to(torch_device)
|
| 783 |
+
model.eval()
|
| 784 |
+
config = model.config
|
| 785 |
+
|
| 786 |
+
input_ids = self._get_dummy_input_ids()
|
| 787 |
+
dummy_hidden_states = model.embeddings(input_ids)
|
| 788 |
+
|
| 789 |
+
attn_mask = torch.ones_like(input_ids, device=torch_device)
|
| 790 |
+
blocked_mask, band_mask, from_mask, to_mask = model.create_masks_for_block_sparse_attn(
|
| 791 |
+
attn_mask, config.block_size
|
| 792 |
+
)
|
| 793 |
+
|
| 794 |
+
targeted_cl = torch.tensor(
|
| 795 |
+
[
|
| 796 |
+
[0.1870, 1.5248, 0.2333, -0.0483, -0.0952, 1.8359, -0.0142, 0.1239, 0.0083, -0.0045],
|
| 797 |
+
[-0.0601, 0.1243, 0.1329, -0.1524, 0.2347, 0.0894, -0.2248, -0.2461, -0.0645, -0.0109],
|
| 798 |
+
[-0.0418, 0.1463, 0.1290, -0.1638, 0.2489, 0.0799, -0.2341, -0.2406, -0.0524, 0.0106],
|
| 799 |
+
[0.1859, 1.5182, 0.2324, -0.0473, -0.0952, 1.8295, -0.0148, 0.1242, 0.0080, -0.0045],
|
| 800 |
+
[0.1879, 1.5300, 0.2334, -0.0480, -0.0967, 1.8428, -0.0137, 0.1256, 0.0087, -0.0050],
|
| 801 |
+
[0.1852, 1.5149, 0.2330, -0.0492, -0.0936, 1.8236, -0.0154, 0.1210, 0.0080, -0.0048],
|
| 802 |
+
[0.1857, 1.5186, 0.2331, -0.0484, -0.0940, 1.8285, -0.0148, 0.1224, 0.0077, -0.0045],
|
| 803 |
+
[0.1884, 1.5336, 0.2334, -0.0469, -0.0974, 1.8477, -0.0132, 0.1266, 0.0085, -0.0046],
|
| 804 |
+
[0.1881, 1.5308, 0.2334, -0.0479, -0.0969, 1.8438, -0.0136, 0.1258, 0.0088, -0.0050],
|
| 805 |
+
[0.1849, 1.5143, 0.2329, -0.0491, -0.0930, 1.8230, -0.0156, 0.1209, 0.0074, -0.0047],
|
| 806 |
+
[0.1878, 1.5299, 0.2333, -0.0472, -0.0967, 1.8434, -0.0137, 0.1257, 0.0084, -0.0048],
|
| 807 |
+
[0.1873, 1.5260, 0.2333, -0.0478, -0.0961, 1.8383, -0.0142, 0.1245, 0.0083, -0.0048],
|
| 808 |
+
[0.1849, 1.5145, 0.2327, -0.0491, -0.0935, 1.8237, -0.0156, 0.1215, 0.0083, -0.0046],
|
| 809 |
+
[0.1866, 1.5232, 0.2332, -0.0488, -0.0950, 1.8342, -0.0143, 0.1237, 0.0084, -0.0047],
|
| 810 |
+
],
|
| 811 |
+
device=torch_device,
|
| 812 |
+
)
|
| 813 |
+
|
| 814 |
+
context_layer = model.encoder.layer[0].attention.self(
|
| 815 |
+
dummy_hidden_states,
|
| 816 |
+
band_mask=band_mask,
|
| 817 |
+
from_mask=from_mask,
|
| 818 |
+
to_mask=to_mask,
|
| 819 |
+
from_blocked_mask=blocked_mask,
|
| 820 |
+
to_blocked_mask=blocked_mask,
|
| 821 |
+
)
|
| 822 |
+
context_layer = context_layer[0]
|
| 823 |
+
|
| 824 |
+
self.assertEqual(context_layer.shape, torch.Size((1, 128, 768)))
|
| 825 |
+
self.assertTrue(torch.allclose(context_layer[0, 64:78, 300:310], targeted_cl, atol=0.0001))
|
| 826 |
+
|
| 827 |
+
def test_tokenizer_inference(self):
|
| 828 |
+
tokenizer = BigBirdTokenizer.from_pretrained("google/bigbird-roberta-base")
|
| 829 |
+
model = BigBirdModel.from_pretrained(
|
| 830 |
+
"google/bigbird-roberta-base", attention_type="block_sparse", num_random_blocks=3, block_size=16
|
| 831 |
+
)
|
| 832 |
+
model.to(torch_device)
|
| 833 |
+
|
| 834 |
+
text = [
|
| 835 |
+
"Transformer-based models are unable to process long sequences due to their self-attention operation,"
|
| 836 |
+
" which scales quadratically with the sequence length. To address this limitation, we introduce the"
|
| 837 |
+
" Longformer with an attention mechanism that scales linearly with sequence length, making it easy to"
|
| 838 |
+
" process documents of thousands of tokens or longer. Longformer’s attention mechanism is a drop-in"
|
| 839 |
+
" replacement for the standard self-attention and combines a local windowed attention with a task"
|
| 840 |
+
" motivated global attention. Following prior work on long-sequence transformers, we evaluate Longformer"
|
| 841 |
+
" on character-level language modeling and achieve state-of-the-art results on text8 and enwik8. In"
|
| 842 |
+
" contrast to most prior work, we also pretrain Longformer and finetune it on a variety of downstream"
|
| 843 |
+
" tasks. Our pretrained Longformer consistently outperforms RoBERTa on long document tasks and sets new"
|
| 844 |
+
" state-of-the-art results on WikiHop and TriviaQA."
|
| 845 |
+
]
|
| 846 |
+
inputs = tokenizer(text)
|
| 847 |
+
|
| 848 |
+
for k in inputs:
|
| 849 |
+
inputs[k] = torch.tensor(inputs[k], device=torch_device, dtype=torch.long)
|
| 850 |
+
|
| 851 |
+
prediction = model(**inputs)
|
| 852 |
+
prediction = prediction[0]
|
| 853 |
+
|
| 854 |
+
self.assertEqual(prediction.shape, torch.Size((1, 199, 768)))
|
| 855 |
+
|
| 856 |
+
expected_prediction = torch.tensor(
|
| 857 |
+
[
|
| 858 |
+
[0.1887, -0.0474, 0.2604, 0.1453],
|
| 859 |
+
[0.0651, 0.1999, 0.1797, 0.1161],
|
| 860 |
+
[0.2833, -0.3036, 0.6910, 0.1123],
|
| 861 |
+
[0.2836, -0.4644, -0.0111, 0.1530],
|
| 862 |
+
[0.3919, -0.2823, 0.4192, 0.1687],
|
| 863 |
+
[0.2168, -0.1956, 0.4050, 0.0925],
|
| 864 |
+
[0.2597, -0.0884, 0.1258, 0.1119],
|
| 865 |
+
[0.1127, -0.1203, 0.1924, 0.2859],
|
| 866 |
+
[0.1362, -0.1315, 0.2693, 0.1027],
|
| 867 |
+
[-0.3169, -0.2266, 0.4419, 0.6740],
|
| 868 |
+
[0.2366, -0.1452, 0.2589, 0.0579],
|
| 869 |
+
[0.0358, -0.2021, 0.3112, -0.1392],
|
| 870 |
+
],
|
| 871 |
+
device=torch_device,
|
| 872 |
+
)
|
| 873 |
+
|
| 874 |
+
self.assertTrue(torch.allclose(prediction[0, 52:64, 320:324], expected_prediction, atol=1e-4))
|
| 875 |
+
|
| 876 |
+
def test_inference_question_answering(self):
|
| 877 |
+
tokenizer = BigBirdTokenizer.from_pretrained("google/bigbird-base-trivia-itc")
|
| 878 |
+
model = BigBirdForQuestionAnswering.from_pretrained(
|
| 879 |
+
"google/bigbird-base-trivia-itc", attention_type="block_sparse", block_size=16, num_random_blocks=3
|
| 880 |
+
)
|
| 881 |
+
model.to(torch_device)
|
| 882 |
+
|
| 883 |
+
context = (
|
| 884 |
+
"The BigBird model was proposed in Big Bird: Transformers for Longer Sequences by Zaheer, Manzil and"
|
| 885 |
+
" Guruganesh, Guru and Dubey, Kumar Avinava and Ainslie, Joshua and Alberti, Chris and Ontanon, Santiago"
|
| 886 |
+
" and Pham, Philip and Ravula, Anirudh and Wang, Qifan and Yang, Li and others. BigBird, is a"
|
| 887 |
+
" sparse-attention based transformer which extends Transformer based models, such as BERT to much longer"
|
| 888 |
+
" sequences. In addition to sparse attention, BigBird also applies global attention as well as random"
|
| 889 |
+
" attention to the input sequence. Theoretically, it has been shown that applying sparse, global, and"
|
| 890 |
+
" random attention approximates full attention, while being computationally much more efficient for longer"
|
| 891 |
+
" sequences. As a consequence of the capability to handle longer context, BigBird has shown improved"
|
| 892 |
+
" performance on various long document NLP tasks, such as question answering and summarization, compared"
|
| 893 |
+
" to BERT or RoBERTa."
|
| 894 |
+
)
|
| 895 |
+
|
| 896 |
+
question = [
|
| 897 |
+
"Which is better for longer sequences- BigBird or BERT?",
|
| 898 |
+
"What is the benefit of using BigBird over BERT?",
|
| 899 |
+
]
|
| 900 |
+
inputs = tokenizer(
|
| 901 |
+
question,
|
| 902 |
+
[context, context],
|
| 903 |
+
padding=True,
|
| 904 |
+
return_tensors="pt",
|
| 905 |
+
add_special_tokens=True,
|
| 906 |
+
max_length=256,
|
| 907 |
+
truncation=True,
|
| 908 |
+
)
|
| 909 |
+
|
| 910 |
+
inputs = {k: v.to(torch_device) for k, v in inputs.items()}
|
| 911 |
+
|
| 912 |
+
start_logits, end_logits = model(**inputs).to_tuple()
|
| 913 |
+
|
| 914 |
+
# fmt: off
|
| 915 |
+
target_start_logits = torch.tensor(
|
| 916 |
+
[[-8.5622, -9.6209, -14.3351, -8.7032, -11.8596, -7.7446, -9.6730, -13.6063, -8.9651, -11.7417, -8.2641, -8.7056, -13.4116, -5.6600, -8.8316, -10.4148, -12.2180, -7.7979, -12.5274, -6.0685, -10.3373, -11.3128, -6.6456, -14.4030, -6.8292, -14.5383, -11.5638, -6.3326, 11.5293, -1.8434, -10.0013, -7.6150], [-10.7384, -13.1179, -10.1837, -13.7700, -10.0186, -11.7335, -13.3411, -10.0188, -13.4235, -9.9381, -10.4252, -13.1281, -8.2022, -10.4326, -11.5542, -14.1549, -10.7546, -13.4691, -8.2744, -11.4324, -13.3773, -9.8284, -14.5825, -8.7471, -14.7050, -8.0364, -11.3627, -6.4638, -11.7031, -14.3446, -9.9425, -8.0088]], # noqa: E231
|
| 917 |
+
device=torch_device,
|
| 918 |
+
)
|
| 919 |
+
|
| 920 |
+
target_end_logits = torch.tensor(
|
| 921 |
+
[[-12.1736, -8.8487, -14.8877, -11.6713, -15.1165, -12.2396, -7.6828, -15.4153, -12.2528, -14.3671, -12.3596, -7.4272, -14.9615, -13.6356, -11.7939, -9.9767, -14.8112, -8.9567, -15.8798, -11.5291, -9.4249, -14.7544, -7.9387, -16.2789, -8.9702, -15.3111, -11.5585, -7.9992, -4.1127, 10.3209, -8.3926, -10.2005], [-11.1375, -15.4027, -12.6861, -16.9884, -13.7093, -10.3560, -15.7228, -12.9290, -15.8519, -13.7953, -10.2460, -15.7198, -14.2078, -12.8477, -11.4861, -16.1017, -11.8900, -16.4488, -13.2959, -10.3980, -15.4874, -10.3539, -16.8263, -10.9973, -17.0344, -9.2751, -10.1196, -13.8907, -12.1025, -13.0628, -12.8530, -13.8173]], # noqa: E321
|
| 922 |
+
device=torch_device,
|
| 923 |
+
)
|
| 924 |
+
# fmt: on
|
| 925 |
+
|
| 926 |
+
self.assertTrue(torch.allclose(start_logits[:, 64:96], target_start_logits, atol=1e-4))
|
| 927 |
+
self.assertTrue(torch.allclose(end_logits[:, 64:96], target_end_logits, atol=1e-4))
|
| 928 |
+
|
| 929 |
+
input_ids = inputs["input_ids"].tolist()
|
| 930 |
+
answer = [
|
| 931 |
+
input_ids[i][torch.argmax(start_logits, dim=-1)[i] : torch.argmax(end_logits, dim=-1)[i] + 1]
|
| 932 |
+
for i in range(len(input_ids))
|
| 933 |
+
]
|
| 934 |
+
answer = tokenizer.batch_decode(answer)
|
| 935 |
+
|
| 936 |
+
self.assertTrue(answer == ["BigBird", "global attention"])
|
| 937 |
+
|
| 938 |
+
def test_fill_mask(self):
|
| 939 |
+
tokenizer = BigBirdTokenizer.from_pretrained("google/bigbird-roberta-base")
|
| 940 |
+
model = BigBirdForMaskedLM.from_pretrained("google/bigbird-roberta-base")
|
| 941 |
+
model.to(torch_device)
|
| 942 |
+
|
| 943 |
+
input_ids = tokenizer("The goal of life is [MASK] .", return_tensors="pt").input_ids.to(torch_device)
|
| 944 |
+
logits = model(input_ids).logits
|
| 945 |
+
|
| 946 |
+
# [MASK] is token at 6th position
|
| 947 |
+
pred_token = tokenizer.decode(torch.argmax(logits[0, 6:7], axis=-1))
|
| 948 |
+
self.assertEqual(pred_token, "happiness")
|
| 949 |
+
|
| 950 |
+
def test_auto_padding(self):
|
| 951 |
+
model = BigBirdModel.from_pretrained(
|
| 952 |
+
"google/bigbird-roberta-base", attention_type="block_sparse", num_random_blocks=3, block_size=16
|
| 953 |
+
)
|
| 954 |
+
model.to(torch_device)
|
| 955 |
+
model.eval()
|
| 956 |
+
|
| 957 |
+
input_ids = torch.tensor([200 * [10] + 40 * [2] + [1]], device=torch_device, dtype=torch.long)
|
| 958 |
+
with torch.no_grad():
|
| 959 |
+
output = model(input_ids).to_tuple()[0]
|
| 960 |
+
|
| 961 |
+
# fmt: off
|
| 962 |
+
target = torch.tensor(
|
| 963 |
+
[[-0.129420, -0.164740, 0.042422, -0.336030, 0.094379, 0.033794, 0.384590, 0.229660, -0.196500, 0.108020], [-0.000154, -0.168800, 0.165820, -0.313670, 0.101240, 0.035145, 0.381880, 0.213730, -0.201080, 0.077443], [0.053754, -0.166350, 0.225520, -0.272900, 0.119670, 0.019987, 0.348670, 0.199190, -0.181600, 0.084640], [0.063636, -0.187110, 0.237010, -0.297380, 0.126300, 0.020025, 0.268490, 0.191820, -0.192300, 0.035077], [0.073893, -0.184790, 0.188870, -0.297860, 0.134280, 0.028972, 0.174650, 0.186890, -0.180530, 0.006851], [0.005253, -0.169360, 0.123100, -0.302550, 0.126930, 0.024188, 0.133410, 0.200600, -0.168210, -0.001006], [-0.093336, -0.175370, -0.004768, -0.333170, 0.114330, 0.034168, 0.120960, 0.203570, -0.162810, -0.005757], [-0.160210, -0.169310, -0.049064, -0.331950, 0.115730, 0.027062, 0.143600, 0.205310, -0.144580, 0.026746], [-0.193200, -0.156820, -0.079422, -0.351600, 0.106450, 0.032174, 0.245690, 0.210250, -0.173480, 0.043914], [-0.167980, -0.153050, -0.059764, -0.357890,0.103910, 0.031481, 0.334190, 0.208960,-0.178180, 0.072165], [-0.136990, -0.156950, -0.012099, -0.353140,0.096996, 0.025864, 0.376340, 0.216050, -0.171820, 0.089963], [-0.041143, -0.167060, 0.079754, -0.353220, 0.093247, 0.019867, 0.385810, 0.214340, -0.191800, 0.065946],[0.040373, -0.158610, 0.152570, -0.312930, 0.110590, 0.012282, 0.345270, 0.204040, -0.176500, 0.064972], [0.043762, -0.166450, 0.179500, -0.317930, 0.117280, -0.004040, 0.304490, 0.201380, -0.182780, 0.044000]], # noqa: E231
|
| 964 |
+
device=torch_device,
|
| 965 |
+
)
|
| 966 |
+
# fmt: on
|
| 967 |
+
|
| 968 |
+
self.assertEqual(output.shape, torch.Size((1, 241, 768)))
|
| 969 |
+
self.assertTrue(torch.allclose(output[0, 64:78, 300:310], target, atol=0.0001))
|
tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/big_bird/test_modeling_flax_big_bird.py
ADDED
|
@@ -0,0 +1,224 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Copyright 2020 The HuggingFace Team. All rights reserved.
|
| 2 |
+
#
|
| 3 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
| 4 |
+
# you may not use this file except in compliance with the License.
|
| 5 |
+
# You may obtain a copy of the License at
|
| 6 |
+
#
|
| 7 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
| 8 |
+
#
|
| 9 |
+
# Unless required by applicable law or agreed to in writing, software
|
| 10 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
| 11 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
| 12 |
+
# See the License for the specific language governing permissions and
|
| 13 |
+
# limitations under the License.
|
| 14 |
+
|
| 15 |
+
import unittest
|
| 16 |
+
|
| 17 |
+
from transformers import BigBirdConfig, is_flax_available
|
| 18 |
+
from transformers.testing_utils import require_flax, slow
|
| 19 |
+
|
| 20 |
+
from ...test_modeling_flax_common import FlaxModelTesterMixin, ids_tensor, random_attention_mask
|
| 21 |
+
|
| 22 |
+
|
| 23 |
+
if is_flax_available():
|
| 24 |
+
import jax
|
| 25 |
+
|
| 26 |
+
from transformers.models.big_bird.modeling_flax_big_bird import (
|
| 27 |
+
FlaxBigBirdForCausalLM,
|
| 28 |
+
FlaxBigBirdForMaskedLM,
|
| 29 |
+
FlaxBigBirdForMultipleChoice,
|
| 30 |
+
FlaxBigBirdForPreTraining,
|
| 31 |
+
FlaxBigBirdForQuestionAnswering,
|
| 32 |
+
FlaxBigBirdForSequenceClassification,
|
| 33 |
+
FlaxBigBirdForTokenClassification,
|
| 34 |
+
FlaxBigBirdModel,
|
| 35 |
+
)
|
| 36 |
+
|
| 37 |
+
|
| 38 |
+
class FlaxBigBirdModelTester(unittest.TestCase):
|
| 39 |
+
def __init__(
|
| 40 |
+
self,
|
| 41 |
+
parent,
|
| 42 |
+
batch_size=2,
|
| 43 |
+
seq_length=56,
|
| 44 |
+
is_training=True,
|
| 45 |
+
use_attention_mask=True,
|
| 46 |
+
use_token_type_ids=True,
|
| 47 |
+
use_labels=True,
|
| 48 |
+
vocab_size=99,
|
| 49 |
+
hidden_size=32,
|
| 50 |
+
num_hidden_layers=2,
|
| 51 |
+
num_attention_heads=2,
|
| 52 |
+
intermediate_size=7,
|
| 53 |
+
hidden_act="gelu_new",
|
| 54 |
+
hidden_dropout_prob=0.1,
|
| 55 |
+
attention_probs_dropout_prob=0.1,
|
| 56 |
+
max_position_embeddings=512,
|
| 57 |
+
type_vocab_size=16,
|
| 58 |
+
type_sequence_label_size=2,
|
| 59 |
+
initializer_range=0.02,
|
| 60 |
+
num_choices=4,
|
| 61 |
+
attention_type="block_sparse",
|
| 62 |
+
use_bias=True,
|
| 63 |
+
rescale_embeddings=False,
|
| 64 |
+
block_size=2,
|
| 65 |
+
num_random_blocks=3,
|
| 66 |
+
):
|
| 67 |
+
self.parent = parent
|
| 68 |
+
self.batch_size = batch_size
|
| 69 |
+
self.seq_length = seq_length
|
| 70 |
+
self.is_training = is_training
|
| 71 |
+
self.use_attention_mask = use_attention_mask
|
| 72 |
+
self.use_token_type_ids = use_token_type_ids
|
| 73 |
+
self.use_labels = use_labels
|
| 74 |
+
self.vocab_size = vocab_size
|
| 75 |
+
self.hidden_size = hidden_size
|
| 76 |
+
self.num_hidden_layers = num_hidden_layers
|
| 77 |
+
self.num_attention_heads = num_attention_heads
|
| 78 |
+
self.intermediate_size = intermediate_size
|
| 79 |
+
self.hidden_act = hidden_act
|
| 80 |
+
self.hidden_dropout_prob = hidden_dropout_prob
|
| 81 |
+
self.attention_probs_dropout_prob = attention_probs_dropout_prob
|
| 82 |
+
self.max_position_embeddings = max_position_embeddings
|
| 83 |
+
self.type_vocab_size = type_vocab_size
|
| 84 |
+
self.type_sequence_label_size = type_sequence_label_size
|
| 85 |
+
self.initializer_range = initializer_range
|
| 86 |
+
self.num_choices = num_choices
|
| 87 |
+
|
| 88 |
+
self.rescale_embeddings = rescale_embeddings
|
| 89 |
+
self.attention_type = attention_type
|
| 90 |
+
self.use_bias = use_bias
|
| 91 |
+
self.block_size = block_size
|
| 92 |
+
self.num_random_blocks = num_random_blocks
|
| 93 |
+
super().__init__()
|
| 94 |
+
|
| 95 |
+
def prepare_config_and_inputs(self):
|
| 96 |
+
input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
|
| 97 |
+
|
| 98 |
+
attention_mask = None
|
| 99 |
+
if self.use_attention_mask:
|
| 100 |
+
attention_mask = random_attention_mask([self.batch_size, self.seq_length])
|
| 101 |
+
|
| 102 |
+
token_type_ids = None
|
| 103 |
+
if self.use_token_type_ids:
|
| 104 |
+
token_type_ids = ids_tensor([self.batch_size, self.seq_length], self.type_vocab_size)
|
| 105 |
+
|
| 106 |
+
config = BigBirdConfig(
|
| 107 |
+
vocab_size=self.vocab_size,
|
| 108 |
+
hidden_size=self.hidden_size,
|
| 109 |
+
num_hidden_layers=self.num_hidden_layers,
|
| 110 |
+
num_attention_heads=self.num_attention_heads,
|
| 111 |
+
intermediate_size=self.intermediate_size,
|
| 112 |
+
hidden_act=self.hidden_act,
|
| 113 |
+
hidden_dropout_prob=self.hidden_dropout_prob,
|
| 114 |
+
attention_probs_dropout_prob=self.attention_probs_dropout_prob,
|
| 115 |
+
max_position_embeddings=self.max_position_embeddings,
|
| 116 |
+
type_vocab_size=self.type_vocab_size,
|
| 117 |
+
is_decoder=False,
|
| 118 |
+
initializer_range=self.initializer_range,
|
| 119 |
+
attention_type=self.attention_type,
|
| 120 |
+
block_size=self.block_size,
|
| 121 |
+
num_random_blocks=self.num_random_blocks,
|
| 122 |
+
use_bias=self.use_bias,
|
| 123 |
+
rescale_embeddings=self.rescale_embeddings,
|
| 124 |
+
)
|
| 125 |
+
|
| 126 |
+
return config, input_ids, token_type_ids, attention_mask
|
| 127 |
+
|
| 128 |
+
def prepare_config_and_inputs_for_common(self):
|
| 129 |
+
config_and_inputs = self.prepare_config_and_inputs()
|
| 130 |
+
config, input_ids, token_type_ids, attention_mask = config_and_inputs
|
| 131 |
+
inputs_dict = {
|
| 132 |
+
"input_ids": input_ids,
|
| 133 |
+
"token_type_ids": token_type_ids,
|
| 134 |
+
"attention_mask": attention_mask,
|
| 135 |
+
}
|
| 136 |
+
return config, inputs_dict
|
| 137 |
+
|
| 138 |
+
|
| 139 |
+
@require_flax
|
| 140 |
+
class FlaxBigBirdModelTest(FlaxModelTesterMixin, unittest.TestCase):
|
| 141 |
+
all_model_classes = (
|
| 142 |
+
(
|
| 143 |
+
FlaxBigBirdForCausalLM,
|
| 144 |
+
FlaxBigBirdModel,
|
| 145 |
+
FlaxBigBirdForPreTraining,
|
| 146 |
+
FlaxBigBirdForMaskedLM,
|
| 147 |
+
FlaxBigBirdForMultipleChoice,
|
| 148 |
+
FlaxBigBirdForQuestionAnswering,
|
| 149 |
+
FlaxBigBirdForSequenceClassification,
|
| 150 |
+
FlaxBigBirdForTokenClassification,
|
| 151 |
+
)
|
| 152 |
+
if is_flax_available()
|
| 153 |
+
else ()
|
| 154 |
+
)
|
| 155 |
+
|
| 156 |
+
test_attn_probs = False
|
| 157 |
+
test_mismatched_shapes = False
|
| 158 |
+
|
| 159 |
+
def setUp(self):
|
| 160 |
+
self.model_tester = FlaxBigBirdModelTester(self)
|
| 161 |
+
|
| 162 |
+
@slow
|
| 163 |
+
# copied from `test_modeling_flax_common` because it takes much longer than other models
|
| 164 |
+
def test_from_pretrained_save_pretrained(self):
|
| 165 |
+
super().test_from_pretrained_save_pretrained()
|
| 166 |
+
|
| 167 |
+
@slow
|
| 168 |
+
# copied from `test_modeling_flax_common` because it takes much longer than other models
|
| 169 |
+
def test_from_pretrained_with_no_automatic_init(self):
|
| 170 |
+
super().test_from_pretrained_with_no_automatic_init()
|
| 171 |
+
|
| 172 |
+
@slow
|
| 173 |
+
# copied from `test_modeling_flax_common` because it takes much longer than other models
|
| 174 |
+
def test_no_automatic_init(self):
|
| 175 |
+
super().test_no_automatic_init()
|
| 176 |
+
|
| 177 |
+
@slow
|
| 178 |
+
# copied from `test_modeling_flax_common` because it takes much longer than other models
|
| 179 |
+
def test_hidden_states_output(self):
|
| 180 |
+
super().test_hidden_states_output()
|
| 181 |
+
|
| 182 |
+
@slow
|
| 183 |
+
def test_model_from_pretrained(self):
|
| 184 |
+
for model_class_name in self.all_model_classes:
|
| 185 |
+
model = model_class_name.from_pretrained("google/bigbird-roberta-base")
|
| 186 |
+
self.assertIsNotNone(model)
|
| 187 |
+
|
| 188 |
+
def test_attention_outputs(self):
|
| 189 |
+
if self.test_attn_probs:
|
| 190 |
+
super().test_attention_outputs()
|
| 191 |
+
|
| 192 |
+
@slow
|
| 193 |
+
# copied from `test_modeling_flax_common` because it takes much longer than other models
|
| 194 |
+
def test_jit_compilation(self):
|
| 195 |
+
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
| 196 |
+
|
| 197 |
+
for model_class in self.all_model_classes:
|
| 198 |
+
with self.subTest(model_class.__name__):
|
| 199 |
+
prepared_inputs_dict = self._prepare_for_class(inputs_dict, model_class)
|
| 200 |
+
model = model_class(config)
|
| 201 |
+
|
| 202 |
+
@jax.jit
|
| 203 |
+
def model_jitted(input_ids, attention_mask=None, **kwargs):
|
| 204 |
+
return model(input_ids=input_ids, attention_mask=attention_mask, **kwargs)
|
| 205 |
+
|
| 206 |
+
with self.subTest("JIT Enabled"):
|
| 207 |
+
jitted_outputs = model_jitted(**prepared_inputs_dict).to_tuple()
|
| 208 |
+
|
| 209 |
+
with self.subTest("JIT Disabled"):
|
| 210 |
+
with jax.disable_jit():
|
| 211 |
+
outputs = model_jitted(**prepared_inputs_dict).to_tuple()
|
| 212 |
+
|
| 213 |
+
self.assertEqual(len(outputs), len(jitted_outputs))
|
| 214 |
+
for jitted_output, output in zip(jitted_outputs, outputs):
|
| 215 |
+
self.assertEqual(jitted_output.shape, output.shape)
|
| 216 |
+
|
| 217 |
+
# overwrite from common in order to skip the check on `attentions`
|
| 218 |
+
def check_pt_flax_outputs(self, fx_outputs, pt_outputs, model_class, tol=1e-5, name="outputs", attributes=None):
|
| 219 |
+
# `bigbird_block_sparse_attention` in `FlaxBigBird` returns `attention_probs = None`, while in PyTorch version,
|
| 220 |
+
# an effort was done to return `attention_probs` (yet to be verified).
|
| 221 |
+
if name.startswith("outputs.attentions"):
|
| 222 |
+
return
|
| 223 |
+
else:
|
| 224 |
+
super().check_pt_flax_outputs(fx_outputs, pt_outputs, model_class, tol, name, attributes)
|
tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/big_bird/test_tokenization_big_bird.py
ADDED
|
@@ -0,0 +1,233 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# coding=utf-8
|
| 2 |
+
# Copyright 2020 The HuggingFace Team. All rights reserved.
|
| 3 |
+
#
|
| 4 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
| 5 |
+
# you may not use this file except in compliance with the License.
|
| 6 |
+
# You may obtain a copy of the License at
|
| 7 |
+
#
|
| 8 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
| 9 |
+
#
|
| 10 |
+
# Unless required by applicable law or agreed to in writing, software
|
| 11 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
| 12 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
| 13 |
+
# See the License for the specific language governing permissions and
|
| 14 |
+
# limitations under the License.
|
| 15 |
+
|
| 16 |
+
import unittest
|
| 17 |
+
|
| 18 |
+
from transformers import BigBirdTokenizer, BigBirdTokenizerFast
|
| 19 |
+
from transformers.testing_utils import get_tests_dir, require_sentencepiece, require_tokenizers, require_torch, slow
|
| 20 |
+
from transformers.utils import cached_property
|
| 21 |
+
|
| 22 |
+
from ...test_tokenization_common import TokenizerTesterMixin
|
| 23 |
+
|
| 24 |
+
|
| 25 |
+
SPIECE_UNDERLINE = "▁"
|
| 26 |
+
|
| 27 |
+
SAMPLE_VOCAB = get_tests_dir("fixtures/test_sentencepiece.model")
|
| 28 |
+
|
| 29 |
+
|
| 30 |
+
@require_sentencepiece
|
| 31 |
+
@require_tokenizers
|
| 32 |
+
class BigBirdTokenizationTest(TokenizerTesterMixin, unittest.TestCase):
|
| 33 |
+
from_pretrained_id = "google/bigbird-roberta-base"
|
| 34 |
+
tokenizer_class = BigBirdTokenizer
|
| 35 |
+
rust_tokenizer_class = BigBirdTokenizerFast
|
| 36 |
+
test_rust_tokenizer = True
|
| 37 |
+
test_sentencepiece = True
|
| 38 |
+
|
| 39 |
+
def setUp(self):
|
| 40 |
+
super().setUp()
|
| 41 |
+
|
| 42 |
+
tokenizer = self.tokenizer_class(SAMPLE_VOCAB, keep_accents=True)
|
| 43 |
+
tokenizer.save_pretrained(self.tmpdirname)
|
| 44 |
+
|
| 45 |
+
def test_convert_token_and_id(self):
|
| 46 |
+
"""Test ``_convert_token_to_id`` and ``_convert_id_to_token``."""
|
| 47 |
+
token = "<s>"
|
| 48 |
+
token_id = 1
|
| 49 |
+
|
| 50 |
+
self.assertEqual(self.get_tokenizer()._convert_token_to_id(token), token_id)
|
| 51 |
+
self.assertEqual(self.get_tokenizer()._convert_id_to_token(token_id), token)
|
| 52 |
+
|
| 53 |
+
def test_get_vocab(self):
|
| 54 |
+
vocab_keys = list(self.get_tokenizer().get_vocab().keys())
|
| 55 |
+
|
| 56 |
+
self.assertEqual(vocab_keys[0], "<unk>")
|
| 57 |
+
self.assertEqual(vocab_keys[1], "<s>")
|
| 58 |
+
self.assertEqual(vocab_keys[-1], "[MASK]")
|
| 59 |
+
self.assertEqual(len(vocab_keys), 1_004)
|
| 60 |
+
|
| 61 |
+
def test_vocab_size(self):
|
| 62 |
+
self.assertEqual(self.get_tokenizer().vocab_size, 1_000)
|
| 63 |
+
|
| 64 |
+
def test_rust_and_python_full_tokenizers(self):
|
| 65 |
+
if not self.test_rust_tokenizer:
|
| 66 |
+
self.skipTest(reason="test_rust_tokenizer is set to False")
|
| 67 |
+
|
| 68 |
+
tokenizer = self.get_tokenizer()
|
| 69 |
+
rust_tokenizer = self.get_rust_tokenizer()
|
| 70 |
+
|
| 71 |
+
sequence = "I was born in 92000, and this is falsé."
|
| 72 |
+
|
| 73 |
+
tokens = tokenizer.tokenize(sequence)
|
| 74 |
+
rust_tokens = rust_tokenizer.tokenize(sequence)
|
| 75 |
+
self.assertListEqual(tokens, rust_tokens)
|
| 76 |
+
|
| 77 |
+
ids = tokenizer.encode(sequence, add_special_tokens=False)
|
| 78 |
+
rust_ids = rust_tokenizer.encode(sequence, add_special_tokens=False)
|
| 79 |
+
self.assertListEqual(ids, rust_ids)
|
| 80 |
+
|
| 81 |
+
rust_tokenizer = self.get_rust_tokenizer()
|
| 82 |
+
ids = tokenizer.encode(sequence)
|
| 83 |
+
rust_ids = rust_tokenizer.encode(sequence)
|
| 84 |
+
self.assertListEqual(ids, rust_ids)
|
| 85 |
+
|
| 86 |
+
def test_full_tokenizer(self):
|
| 87 |
+
tokenizer = BigBirdTokenizer(SAMPLE_VOCAB, keep_accents=True)
|
| 88 |
+
|
| 89 |
+
tokens = tokenizer.tokenize("This is a test")
|
| 90 |
+
self.assertListEqual(tokens, ["▁This", "▁is", "▁a", "▁t", "est"])
|
| 91 |
+
|
| 92 |
+
self.assertListEqual(
|
| 93 |
+
tokenizer.convert_tokens_to_ids(tokens),
|
| 94 |
+
[285, 46, 10, 170, 382],
|
| 95 |
+
)
|
| 96 |
+
|
| 97 |
+
tokens = tokenizer.tokenize("I was born in 92000, and this is falsé.")
|
| 98 |
+
self.assertListEqual(
|
| 99 |
+
tokens,
|
| 100 |
+
[
|
| 101 |
+
SPIECE_UNDERLINE + "I",
|
| 102 |
+
SPIECE_UNDERLINE + "was",
|
| 103 |
+
SPIECE_UNDERLINE + "b",
|
| 104 |
+
"or",
|
| 105 |
+
"n",
|
| 106 |
+
SPIECE_UNDERLINE + "in",
|
| 107 |
+
SPIECE_UNDERLINE + "",
|
| 108 |
+
"9",
|
| 109 |
+
"2",
|
| 110 |
+
"0",
|
| 111 |
+
"0",
|
| 112 |
+
"0",
|
| 113 |
+
",",
|
| 114 |
+
SPIECE_UNDERLINE + "and",
|
| 115 |
+
SPIECE_UNDERLINE + "this",
|
| 116 |
+
SPIECE_UNDERLINE + "is",
|
| 117 |
+
SPIECE_UNDERLINE + "f",
|
| 118 |
+
"al",
|
| 119 |
+
"s",
|
| 120 |
+
"é",
|
| 121 |
+
".",
|
| 122 |
+
],
|
| 123 |
+
)
|
| 124 |
+
ids = tokenizer.convert_tokens_to_ids(tokens)
|
| 125 |
+
self.assertListEqual(
|
| 126 |
+
ids,
|
| 127 |
+
[8, 21, 84, 55, 24, 19, 7, 0, 602, 347, 347, 347, 3, 12, 66, 46, 72, 80, 6, 0, 4],
|
| 128 |
+
)
|
| 129 |
+
|
| 130 |
+
back_tokens = tokenizer.convert_ids_to_tokens(ids)
|
| 131 |
+
self.assertListEqual(
|
| 132 |
+
back_tokens,
|
| 133 |
+
[
|
| 134 |
+
SPIECE_UNDERLINE + "I",
|
| 135 |
+
SPIECE_UNDERLINE + "was",
|
| 136 |
+
SPIECE_UNDERLINE + "b",
|
| 137 |
+
"or",
|
| 138 |
+
"n",
|
| 139 |
+
SPIECE_UNDERLINE + "in",
|
| 140 |
+
SPIECE_UNDERLINE + "",
|
| 141 |
+
"<unk>",
|
| 142 |
+
"2",
|
| 143 |
+
"0",
|
| 144 |
+
"0",
|
| 145 |
+
"0",
|
| 146 |
+
",",
|
| 147 |
+
SPIECE_UNDERLINE + "and",
|
| 148 |
+
SPIECE_UNDERLINE + "this",
|
| 149 |
+
SPIECE_UNDERLINE + "is",
|
| 150 |
+
SPIECE_UNDERLINE + "f",
|
| 151 |
+
"al",
|
| 152 |
+
"s",
|
| 153 |
+
"<unk>",
|
| 154 |
+
".",
|
| 155 |
+
],
|
| 156 |
+
)
|
| 157 |
+
|
| 158 |
+
@cached_property
|
| 159 |
+
def big_tokenizer(self):
|
| 160 |
+
return BigBirdTokenizer.from_pretrained("google/bigbird-roberta-base")
|
| 161 |
+
|
| 162 |
+
@slow
|
| 163 |
+
def test_tokenization_base_easy_symbols(self):
|
| 164 |
+
symbols = "Hello World!"
|
| 165 |
+
original_tokenizer_encodings = [65, 18536, 2260, 101, 66]
|
| 166 |
+
|
| 167 |
+
self.assertListEqual(original_tokenizer_encodings, self.big_tokenizer.encode(symbols))
|
| 168 |
+
|
| 169 |
+
@slow
|
| 170 |
+
def test_tokenization_base_hard_symbols(self):
|
| 171 |
+
symbols = (
|
| 172 |
+
'This is a very long text with a lot of weird characters, such as: . , ~ ? ( ) " [ ] ! : - . Also we will'
|
| 173 |
+
" add words that should not exsist and be tokenized to <unk>, such as saoneuhaoesuth"
|
| 174 |
+
)
|
| 175 |
+
original_tokenizer_encodings = [65, 871, 419, 358, 946, 991, 2521, 452, 358, 1357, 387, 7751, 3536, 112, 985, 456, 126, 865, 938, 5400, 5734, 458, 1368, 467, 786, 2462, 5246, 1159, 633, 865, 4519, 457, 582, 852, 2557, 427, 916, 508, 405, 34324, 497, 391, 408, 11342, 1244, 385, 100, 938, 985, 456, 574, 362, 12597, 3200, 3129, 1172, 66] # fmt: skip
|
| 176 |
+
self.assertListEqual(original_tokenizer_encodings, self.big_tokenizer.encode(symbols))
|
| 177 |
+
|
| 178 |
+
@require_torch
|
| 179 |
+
@slow
|
| 180 |
+
def test_torch_encode_plus_sent_to_model(self):
|
| 181 |
+
import torch
|
| 182 |
+
|
| 183 |
+
from transformers import BigBirdConfig, BigBirdModel
|
| 184 |
+
|
| 185 |
+
# Build sequence
|
| 186 |
+
first_ten_tokens = list(self.big_tokenizer.get_vocab().keys())[:10]
|
| 187 |
+
sequence = " ".join(first_ten_tokens)
|
| 188 |
+
encoded_sequence = self.big_tokenizer.encode_plus(sequence, return_tensors="pt", return_token_type_ids=False)
|
| 189 |
+
batch_encoded_sequence = self.big_tokenizer.batch_encode_plus(
|
| 190 |
+
[sequence + " " + sequence], return_tensors="pt", return_token_type_ids=False
|
| 191 |
+
)
|
| 192 |
+
|
| 193 |
+
config = BigBirdConfig(attention_type="original_full")
|
| 194 |
+
model = BigBirdModel(config)
|
| 195 |
+
|
| 196 |
+
assert model.get_input_embeddings().weight.shape[0] >= self.big_tokenizer.vocab_size
|
| 197 |
+
|
| 198 |
+
with torch.no_grad():
|
| 199 |
+
model(**encoded_sequence)
|
| 200 |
+
model(**batch_encoded_sequence)
|
| 201 |
+
|
| 202 |
+
@slow
|
| 203 |
+
def test_special_tokens(self):
|
| 204 |
+
"""
|
| 205 |
+
To reproduce:
|
| 206 |
+
|
| 207 |
+
$ wget https://github.com/google-research/bigbird/blob/master/bigbird/vocab/gpt2.model?raw=true
|
| 208 |
+
$ mv gpt2.model?raw=true gpt2.model
|
| 209 |
+
|
| 210 |
+
```
|
| 211 |
+
import tensorflow_text as tft
|
| 212 |
+
import tensorflow as tf
|
| 213 |
+
|
| 214 |
+
vocab_model_file = "./gpt2.model"
|
| 215 |
+
tokenizer = tft.SentencepieceTokenizer(model=tf.io.gfile.GFile(vocab_model_file, "rb").read()))
|
| 216 |
+
ids = tokenizer.tokenize("Paris is the [MASK].")
|
| 217 |
+
ids = tf.concat([tf.constant([65]), ids, tf.constant([66])], axis=0)
|
| 218 |
+
detokenized = tokenizer.detokenize(ids) # should give [CLS] Paris is the [MASK].[SEP]
|
| 219 |
+
"""
|
| 220 |
+
tokenizer = BigBirdTokenizer.from_pretrained("google/bigbird-roberta-base")
|
| 221 |
+
decoded_text = tokenizer.decode(tokenizer("Paris is the [MASK].").input_ids)
|
| 222 |
+
|
| 223 |
+
self.assertTrue(decoded_text == "[CLS] Paris is the[MASK].[SEP]")
|
| 224 |
+
|
| 225 |
+
@slow
|
| 226 |
+
def test_tokenizer_integration(self):
|
| 227 |
+
expected_encoding = {'input_ids': [[65, 39286, 458, 36335, 2001, 456, 13073, 13266, 455, 113, 7746, 1741, 11157, 391, 13073, 13266, 455, 113, 3967, 35412, 113, 4936, 109, 3870, 2377, 113, 30084, 45720, 458, 134, 17496, 112, 503, 11672, 113, 118, 112, 5665, 13347, 38687, 112, 1496, 31389, 112, 3268, 47264, 134, 962, 112, 16377, 8035, 23130, 430, 12169, 15518, 28592, 458, 146, 41697, 109, 391, 12169, 15518, 16689, 458, 146, 41358, 109, 452, 726, 4034, 111, 763, 35412, 5082, 388, 1903, 111, 9051, 391, 2870, 48918, 1900, 1123, 550, 998, 112, 9586, 15985, 455, 391, 410, 22955, 37636, 114, 66], [65, 448, 17496, 419, 3663, 385, 763, 113, 27533, 2870, 3283, 13043, 1639, 24713, 523, 656, 24013, 18550, 2521, 517, 27014, 21244, 420, 1212, 1465, 391, 927, 4833, 388, 578, 11786, 114, 66, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [65, 484, 2169, 7687, 21932, 18146, 726, 363, 17032, 3391, 114, 66, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]], 'attention_mask': [[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]]} # fmt: skip
|
| 228 |
+
|
| 229 |
+
self.tokenizer_integration_test_util(
|
| 230 |
+
expected_encoding=expected_encoding,
|
| 231 |
+
model_name="google/bigbird-roberta-base",
|
| 232 |
+
revision="215c99f1600e06f83acce68422f2035b2b5c3510",
|
| 233 |
+
)
|
tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/bigbird_pegasus/__init__.py
ADDED
|
File without changes
|
tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/bigbird_pegasus/test_modeling_bigbird_pegasus.py
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/biogpt/__init__.py
ADDED
|
File without changes
|
tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/biogpt/test_modeling_biogpt.py
ADDED
|
@@ -0,0 +1,461 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# coding=utf-8
|
| 2 |
+
# Copyright 2022 The HuggingFace Inc. team. All rights reserved.
|
| 3 |
+
#
|
| 4 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
| 5 |
+
# you may not use this file except in compliance with the License.
|
| 6 |
+
# You may obtain a copy of the License at
|
| 7 |
+
#
|
| 8 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
| 9 |
+
#
|
| 10 |
+
# Unless required by applicable law or agreed to in writing, software
|
| 11 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
| 12 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
| 13 |
+
# See the License for the specific language governing permissions and
|
| 14 |
+
# limitations under the License.
|
| 15 |
+
"""Testing suite for the PyTorch BioGPT model."""
|
| 16 |
+
|
| 17 |
+
import math
|
| 18 |
+
import unittest
|
| 19 |
+
|
| 20 |
+
from transformers import BioGptConfig, is_sacremoses_available, is_torch_available
|
| 21 |
+
from transformers.testing_utils import require_torch, slow, torch_device
|
| 22 |
+
|
| 23 |
+
from ...generation.test_utils import GenerationTesterMixin
|
| 24 |
+
from ...test_configuration_common import ConfigTester
|
| 25 |
+
from ...test_modeling_common import ModelTesterMixin, ids_tensor, random_attention_mask
|
| 26 |
+
from ...test_pipeline_mixin import PipelineTesterMixin
|
| 27 |
+
|
| 28 |
+
|
| 29 |
+
if is_torch_available():
|
| 30 |
+
import torch
|
| 31 |
+
|
| 32 |
+
from transformers import (
|
| 33 |
+
BioGptForCausalLM,
|
| 34 |
+
BioGptForSequenceClassification,
|
| 35 |
+
BioGptForTokenClassification,
|
| 36 |
+
BioGptModel,
|
| 37 |
+
BioGptTokenizer,
|
| 38 |
+
)
|
| 39 |
+
|
| 40 |
+
|
| 41 |
+
class BioGptModelTester:
|
| 42 |
+
def __init__(
|
| 43 |
+
self,
|
| 44 |
+
parent,
|
| 45 |
+
batch_size=13,
|
| 46 |
+
seq_length=7,
|
| 47 |
+
is_training=True,
|
| 48 |
+
use_input_mask=True,
|
| 49 |
+
use_token_type_ids=False,
|
| 50 |
+
use_labels=True,
|
| 51 |
+
vocab_size=99,
|
| 52 |
+
hidden_size=32,
|
| 53 |
+
num_hidden_layers=2,
|
| 54 |
+
num_attention_heads=4,
|
| 55 |
+
intermediate_size=37,
|
| 56 |
+
hidden_act="gelu",
|
| 57 |
+
hidden_dropout_prob=0.1,
|
| 58 |
+
attention_probs_dropout_prob=0.1,
|
| 59 |
+
max_position_embeddings=512,
|
| 60 |
+
type_vocab_size=16,
|
| 61 |
+
type_sequence_label_size=2,
|
| 62 |
+
initializer_range=0.02,
|
| 63 |
+
num_labels=3,
|
| 64 |
+
num_choices=4,
|
| 65 |
+
scope=None,
|
| 66 |
+
):
|
| 67 |
+
self.parent = parent
|
| 68 |
+
self.batch_size = batch_size
|
| 69 |
+
self.seq_length = seq_length
|
| 70 |
+
self.is_training = is_training
|
| 71 |
+
self.use_input_mask = use_input_mask
|
| 72 |
+
self.use_token_type_ids = use_token_type_ids
|
| 73 |
+
self.use_labels = use_labels
|
| 74 |
+
self.vocab_size = vocab_size
|
| 75 |
+
self.hidden_size = hidden_size
|
| 76 |
+
self.num_hidden_layers = num_hidden_layers
|
| 77 |
+
self.num_attention_heads = num_attention_heads
|
| 78 |
+
self.intermediate_size = intermediate_size
|
| 79 |
+
self.hidden_act = hidden_act
|
| 80 |
+
self.hidden_dropout_prob = hidden_dropout_prob
|
| 81 |
+
self.attention_probs_dropout_prob = attention_probs_dropout_prob
|
| 82 |
+
self.max_position_embeddings = max_position_embeddings
|
| 83 |
+
self.type_vocab_size = type_vocab_size
|
| 84 |
+
self.type_sequence_label_size = type_sequence_label_size
|
| 85 |
+
self.initializer_range = initializer_range
|
| 86 |
+
self.num_labels = num_labels
|
| 87 |
+
self.num_choices = num_choices
|
| 88 |
+
self.scope = scope
|
| 89 |
+
|
| 90 |
+
def prepare_config_and_inputs(self):
|
| 91 |
+
input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
|
| 92 |
+
|
| 93 |
+
input_mask = None
|
| 94 |
+
if self.use_input_mask:
|
| 95 |
+
input_mask = random_attention_mask([self.batch_size, self.seq_length])
|
| 96 |
+
|
| 97 |
+
token_type_ids = None
|
| 98 |
+
if self.use_token_type_ids:
|
| 99 |
+
token_type_ids = ids_tensor([self.batch_size, self.seq_length], self.type_vocab_size)
|
| 100 |
+
|
| 101 |
+
sequence_labels = None
|
| 102 |
+
token_labels = None
|
| 103 |
+
choice_labels = None
|
| 104 |
+
if self.use_labels:
|
| 105 |
+
sequence_labels = ids_tensor([self.batch_size], self.type_sequence_label_size)
|
| 106 |
+
token_labels = ids_tensor([self.batch_size, self.seq_length], self.num_labels)
|
| 107 |
+
choice_labels = ids_tensor([self.batch_size], self.num_choices)
|
| 108 |
+
|
| 109 |
+
config = self.get_config()
|
| 110 |
+
|
| 111 |
+
return config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
|
| 112 |
+
|
| 113 |
+
def get_config(self):
|
| 114 |
+
return BioGptConfig(
|
| 115 |
+
vocab_size=self.vocab_size,
|
| 116 |
+
hidden_size=self.hidden_size,
|
| 117 |
+
num_hidden_layers=self.num_hidden_layers,
|
| 118 |
+
num_attention_heads=self.num_attention_heads,
|
| 119 |
+
intermediate_size=self.intermediate_size,
|
| 120 |
+
hidden_act=self.hidden_act,
|
| 121 |
+
hidden_dropout_prob=self.hidden_dropout_prob,
|
| 122 |
+
attention_probs_dropout_prob=self.attention_probs_dropout_prob,
|
| 123 |
+
max_position_embeddings=self.max_position_embeddings,
|
| 124 |
+
type_vocab_size=self.type_vocab_size,
|
| 125 |
+
is_decoder=False,
|
| 126 |
+
initializer_range=self.initializer_range,
|
| 127 |
+
)
|
| 128 |
+
|
| 129 |
+
def create_and_check_model(
|
| 130 |
+
self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
|
| 131 |
+
):
|
| 132 |
+
model = BioGptModel(config=config)
|
| 133 |
+
model.to(torch_device)
|
| 134 |
+
model.eval()
|
| 135 |
+
result = model(input_ids, attention_mask=input_mask)
|
| 136 |
+
result = model(input_ids)
|
| 137 |
+
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
|
| 138 |
+
|
| 139 |
+
def create_and_check_for_causal_lm(
|
| 140 |
+
self,
|
| 141 |
+
config,
|
| 142 |
+
input_ids,
|
| 143 |
+
token_type_ids,
|
| 144 |
+
input_mask,
|
| 145 |
+
sequence_labels,
|
| 146 |
+
token_labels,
|
| 147 |
+
choice_labels,
|
| 148 |
+
encoder_hidden_states,
|
| 149 |
+
encoder_attention_mask,
|
| 150 |
+
):
|
| 151 |
+
model = BioGptForCausalLM(config=config)
|
| 152 |
+
model.to(torch_device)
|
| 153 |
+
model.eval()
|
| 154 |
+
result = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids, labels=token_labels)
|
| 155 |
+
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
|
| 156 |
+
|
| 157 |
+
def create_and_check_biogpt_model_attention_mask_past(
|
| 158 |
+
self, config, input_ids, input_mask, head_mask, token_type_ids, *args
|
| 159 |
+
):
|
| 160 |
+
model = BioGptModel(config=config)
|
| 161 |
+
model.to(torch_device)
|
| 162 |
+
model.eval()
|
| 163 |
+
|
| 164 |
+
# create attention mask
|
| 165 |
+
attn_mask = torch.ones(input_ids.shape, dtype=torch.long, device=torch_device)
|
| 166 |
+
half_seq_length = self.seq_length // 2
|
| 167 |
+
attn_mask[:, half_seq_length:] = 0
|
| 168 |
+
|
| 169 |
+
# first forward pass
|
| 170 |
+
output, past = model(input_ids, attention_mask=attn_mask).to_tuple()
|
| 171 |
+
|
| 172 |
+
# create hypothetical next token and extent to next_input_ids
|
| 173 |
+
next_tokens = ids_tensor((self.batch_size, 1), config.vocab_size)
|
| 174 |
+
|
| 175 |
+
# change a random masked slice from input_ids
|
| 176 |
+
random_seq_idx_to_change = ids_tensor((1,), half_seq_length).item() + 1
|
| 177 |
+
random_other_next_tokens = ids_tensor((self.batch_size, 1), config.vocab_size).squeeze(-1)
|
| 178 |
+
input_ids[:, -random_seq_idx_to_change] = random_other_next_tokens
|
| 179 |
+
|
| 180 |
+
# append to next input_ids and attn_mask
|
| 181 |
+
next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
|
| 182 |
+
attn_mask = torch.cat(
|
| 183 |
+
[attn_mask, torch.ones((attn_mask.shape[0], 1), dtype=torch.long, device=torch_device)],
|
| 184 |
+
dim=1,
|
| 185 |
+
)
|
| 186 |
+
|
| 187 |
+
# get two different outputs
|
| 188 |
+
output_from_no_past = model(next_input_ids, attention_mask=attn_mask)["last_hidden_state"]
|
| 189 |
+
output_from_past = model(next_tokens, past_key_values=past, attention_mask=attn_mask)["last_hidden_state"]
|
| 190 |
+
|
| 191 |
+
# select random slice
|
| 192 |
+
random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()
|
| 193 |
+
output_from_no_past_slice = output_from_no_past[:, -1, random_slice_idx].detach()
|
| 194 |
+
output_from_past_slice = output_from_past[:, 0, random_slice_idx].detach()
|
| 195 |
+
|
| 196 |
+
# test that outputs are equal for slice
|
| 197 |
+
self.parent.assertTrue(torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3))
|
| 198 |
+
|
| 199 |
+
def create_and_check_biogpt_model_past_large_inputs(
|
| 200 |
+
self, config, input_ids, input_mask, head_mask, token_type_ids, *args
|
| 201 |
+
):
|
| 202 |
+
model = BioGptModel(config=config).to(torch_device).eval()
|
| 203 |
+
|
| 204 |
+
attention_mask = torch.ones(input_ids.shape, dtype=torch.long, device=torch_device)
|
| 205 |
+
|
| 206 |
+
# first forward pass
|
| 207 |
+
outputs = model(input_ids, attention_mask=attention_mask, use_cache=True)
|
| 208 |
+
|
| 209 |
+
output, past_key_values = outputs.to_tuple()
|
| 210 |
+
|
| 211 |
+
# create hypothetical multiple next token and extent to next_input_ids
|
| 212 |
+
next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size)
|
| 213 |
+
next_attn_mask = ids_tensor((self.batch_size, 3), 2)
|
| 214 |
+
|
| 215 |
+
# append to next input_ids and
|
| 216 |
+
next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
|
| 217 |
+
next_attention_mask = torch.cat([attention_mask, next_attn_mask], dim=-1)
|
| 218 |
+
|
| 219 |
+
output_from_no_past = model(next_input_ids, attention_mask=next_attention_mask)["last_hidden_state"]
|
| 220 |
+
output_from_past = model(next_tokens, attention_mask=next_attention_mask, past_key_values=past_key_values)[
|
| 221 |
+
"last_hidden_state"
|
| 222 |
+
]
|
| 223 |
+
|
| 224 |
+
# select random slice
|
| 225 |
+
random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()
|
| 226 |
+
output_from_no_past_slice = output_from_no_past[:, -3:, random_slice_idx].detach()
|
| 227 |
+
output_from_past_slice = output_from_past[:, :, random_slice_idx].detach()
|
| 228 |
+
|
| 229 |
+
self.parent.assertTrue(output_from_past_slice.shape[1] == next_tokens.shape[1])
|
| 230 |
+
|
| 231 |
+
# test that outputs are equal for slice
|
| 232 |
+
self.parent.assertTrue(torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3))
|
| 233 |
+
|
| 234 |
+
def create_and_check_forward_and_backwards(
|
| 235 |
+
self, config, input_ids, input_mask, head_mask, token_type_ids, *args, gradient_checkpointing=False
|
| 236 |
+
):
|
| 237 |
+
model = BioGptForCausalLM(config)
|
| 238 |
+
model.to(torch_device)
|
| 239 |
+
if gradient_checkpointing:
|
| 240 |
+
model.gradient_checkpointing_enable()
|
| 241 |
+
|
| 242 |
+
result = model(input_ids, labels=input_ids)
|
| 243 |
+
self.parent.assertEqual(result.loss.shape, ())
|
| 244 |
+
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
|
| 245 |
+
result.loss.backward()
|
| 246 |
+
|
| 247 |
+
def create_and_check_biogpt_weight_initialization(self, config, *args):
|
| 248 |
+
model = BioGptModel(config)
|
| 249 |
+
model_std = model.config.initializer_range / math.sqrt(2 * model.config.num_hidden_layers)
|
| 250 |
+
for key in model.state_dict().keys():
|
| 251 |
+
if "c_proj" in key and "weight" in key:
|
| 252 |
+
self.parent.assertLessEqual(abs(torch.std(model.state_dict()[key]) - model_std), 0.001)
|
| 253 |
+
self.parent.assertLessEqual(abs(torch.mean(model.state_dict()[key]) - 0.0), 0.01)
|
| 254 |
+
|
| 255 |
+
def create_and_check_biogpt_for_token_classification(
|
| 256 |
+
self, config, input_ids, input_mask, head_mask, token_type_ids, *args
|
| 257 |
+
):
|
| 258 |
+
config.num_labels = self.num_labels
|
| 259 |
+
model = BioGptForTokenClassification(config)
|
| 260 |
+
model.to(torch_device)
|
| 261 |
+
model.eval()
|
| 262 |
+
result = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids)
|
| 263 |
+
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.num_labels))
|
| 264 |
+
|
| 265 |
+
def prepare_config_and_inputs_for_common(self):
|
| 266 |
+
config_and_inputs = self.prepare_config_and_inputs()
|
| 267 |
+
(
|
| 268 |
+
config,
|
| 269 |
+
input_ids,
|
| 270 |
+
token_type_ids,
|
| 271 |
+
input_mask,
|
| 272 |
+
sequence_labels,
|
| 273 |
+
token_labels,
|
| 274 |
+
choice_labels,
|
| 275 |
+
) = config_and_inputs
|
| 276 |
+
inputs_dict = {"input_ids": input_ids, "attention_mask": input_mask}
|
| 277 |
+
return config, inputs_dict
|
| 278 |
+
|
| 279 |
+
|
| 280 |
+
@require_torch
|
| 281 |
+
class BioGptModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixin, unittest.TestCase):
|
| 282 |
+
all_model_classes = (
|
| 283 |
+
(BioGptModel, BioGptForCausalLM, BioGptForSequenceClassification, BioGptForTokenClassification)
|
| 284 |
+
if is_torch_available()
|
| 285 |
+
else ()
|
| 286 |
+
)
|
| 287 |
+
all_generative_model_classes = (BioGptForCausalLM,) if is_torch_available() else ()
|
| 288 |
+
pipeline_model_mapping = (
|
| 289 |
+
{
|
| 290 |
+
"feature-extraction": BioGptModel,
|
| 291 |
+
"text-classification": BioGptForSequenceClassification,
|
| 292 |
+
"text-generation": BioGptForCausalLM,
|
| 293 |
+
"token-classification": BioGptForTokenClassification,
|
| 294 |
+
"zero-shot": BioGptForSequenceClassification,
|
| 295 |
+
}
|
| 296 |
+
if is_torch_available() and is_sacremoses_available()
|
| 297 |
+
else {}
|
| 298 |
+
)
|
| 299 |
+
test_pruning = False
|
| 300 |
+
|
| 301 |
+
def setUp(self):
|
| 302 |
+
self.model_tester = BioGptModelTester(self)
|
| 303 |
+
self.config_tester = ConfigTester(self, config_class=BioGptConfig, hidden_size=37)
|
| 304 |
+
|
| 305 |
+
def test_config(self):
|
| 306 |
+
self.config_tester.run_common_tests()
|
| 307 |
+
|
| 308 |
+
def test_model(self):
|
| 309 |
+
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
| 310 |
+
self.model_tester.create_and_check_model(*config_and_inputs)
|
| 311 |
+
|
| 312 |
+
def test_model_various_embeddings(self):
|
| 313 |
+
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
| 314 |
+
for type in ["absolute", "relative_key", "relative_key_query"]:
|
| 315 |
+
config_and_inputs[0].position_embedding_type = type
|
| 316 |
+
self.model_tester.create_and_check_model(*config_and_inputs)
|
| 317 |
+
|
| 318 |
+
def test_biogpt_model_att_mask_past(self):
|
| 319 |
+
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
| 320 |
+
self.model_tester.create_and_check_biogpt_model_attention_mask_past(*config_and_inputs)
|
| 321 |
+
|
| 322 |
+
def test_biogpt_gradient_checkpointing(self):
|
| 323 |
+
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
| 324 |
+
self.model_tester.create_and_check_forward_and_backwards(*config_and_inputs, gradient_checkpointing=True)
|
| 325 |
+
|
| 326 |
+
def test_biogpt_model_past_with_large_inputs(self):
|
| 327 |
+
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
| 328 |
+
self.model_tester.create_and_check_biogpt_model_past_large_inputs(*config_and_inputs)
|
| 329 |
+
|
| 330 |
+
def test_biogpt_weight_initialization(self):
|
| 331 |
+
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
| 332 |
+
self.model_tester.create_and_check_biogpt_weight_initialization(*config_and_inputs)
|
| 333 |
+
|
| 334 |
+
def test_biogpt_token_classification_model(self):
|
| 335 |
+
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
| 336 |
+
self.model_tester.create_and_check_biogpt_for_token_classification(*config_and_inputs)
|
| 337 |
+
|
| 338 |
+
@slow
|
| 339 |
+
def test_batch_generation(self):
|
| 340 |
+
model = BioGptForCausalLM.from_pretrained("microsoft/biogpt")
|
| 341 |
+
model.to(torch_device)
|
| 342 |
+
tokenizer = BioGptTokenizer.from_pretrained("microsoft/biogpt")
|
| 343 |
+
|
| 344 |
+
tokenizer.padding_side = "left"
|
| 345 |
+
|
| 346 |
+
# Define PAD Token = EOS Token = 50256
|
| 347 |
+
tokenizer.pad_token = tokenizer.eos_token
|
| 348 |
+
model.config.pad_token_id = model.config.eos_token_id
|
| 349 |
+
|
| 350 |
+
# use different length sentences to test batching
|
| 351 |
+
sentences = [
|
| 352 |
+
"Hello, my dog is a little",
|
| 353 |
+
"Today, I",
|
| 354 |
+
]
|
| 355 |
+
|
| 356 |
+
inputs = tokenizer(sentences, return_tensors="pt", padding=True)
|
| 357 |
+
input_ids = inputs["input_ids"].to(torch_device)
|
| 358 |
+
|
| 359 |
+
outputs = model.generate(
|
| 360 |
+
input_ids=input_ids,
|
| 361 |
+
attention_mask=inputs["attention_mask"].to(torch_device),
|
| 362 |
+
)
|
| 363 |
+
|
| 364 |
+
inputs_non_padded = tokenizer(sentences[0], return_tensors="pt").input_ids.to(torch_device)
|
| 365 |
+
output_non_padded = model.generate(input_ids=inputs_non_padded)
|
| 366 |
+
|
| 367 |
+
num_paddings = inputs_non_padded.shape[-1] - inputs["attention_mask"][-1].long().sum().cpu().item()
|
| 368 |
+
inputs_padded = tokenizer(sentences[1], return_tensors="pt").input_ids.to(torch_device)
|
| 369 |
+
output_padded = model.generate(input_ids=inputs_padded, max_length=model.config.max_length - num_paddings)
|
| 370 |
+
|
| 371 |
+
batch_out_sentence = tokenizer.batch_decode(outputs, skip_special_tokens=True)
|
| 372 |
+
non_padded_sentence = tokenizer.decode(output_non_padded[0], skip_special_tokens=True)
|
| 373 |
+
padded_sentence = tokenizer.decode(output_padded[0], skip_special_tokens=True)
|
| 374 |
+
|
| 375 |
+
expected_output_sentence = [
|
| 376 |
+
"Hello, my dog is a little bit bigger than a little bit.",
|
| 377 |
+
"Today, I have a good idea of how to use the information",
|
| 378 |
+
]
|
| 379 |
+
self.assertListEqual(expected_output_sentence, batch_out_sentence)
|
| 380 |
+
self.assertListEqual(expected_output_sentence, [non_padded_sentence, padded_sentence])
|
| 381 |
+
|
| 382 |
+
@slow
|
| 383 |
+
def test_model_from_pretrained(self):
|
| 384 |
+
model_name = "microsoft/biogpt"
|
| 385 |
+
model = BioGptModel.from_pretrained(model_name)
|
| 386 |
+
self.assertIsNotNone(model)
|
| 387 |
+
|
| 388 |
+
# Copied from tests.models.opt.test_modeling_opt.OPTModelTest.test_opt_sequence_classification_model with OPT->BioGpt,opt->biogpt,prepare_config_and_inputs->prepare_config_and_inputs_for_common
|
| 389 |
+
def test_biogpt_sequence_classification_model(self):
|
| 390 |
+
config, input_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
| 391 |
+
config.num_labels = 3
|
| 392 |
+
input_ids = input_dict["input_ids"]
|
| 393 |
+
attention_mask = input_ids.ne(1).to(torch_device)
|
| 394 |
+
sequence_labels = ids_tensor([self.model_tester.batch_size], self.model_tester.type_sequence_label_size)
|
| 395 |
+
model = BioGptForSequenceClassification(config)
|
| 396 |
+
model.to(torch_device)
|
| 397 |
+
model.eval()
|
| 398 |
+
result = model(input_ids, attention_mask=attention_mask, labels=sequence_labels)
|
| 399 |
+
self.assertEqual(result.logits.shape, (self.model_tester.batch_size, self.model_tester.num_labels))
|
| 400 |
+
|
| 401 |
+
# Copied from tests.models.opt.test_modeling_opt.OPTModelTest.test_opt_sequence_classification_model_for_multi_label with OPT->BioGpt,opt->biogpt,prepare_config_and_inputs->prepare_config_and_inputs_for_common
|
| 402 |
+
def test_biogpt_sequence_classification_model_for_multi_label(self):
|
| 403 |
+
config, input_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
| 404 |
+
config.num_labels = 3
|
| 405 |
+
config.problem_type = "multi_label_classification"
|
| 406 |
+
input_ids = input_dict["input_ids"]
|
| 407 |
+
attention_mask = input_ids.ne(1).to(torch_device)
|
| 408 |
+
sequence_labels = ids_tensor(
|
| 409 |
+
[self.model_tester.batch_size, config.num_labels], self.model_tester.type_sequence_label_size
|
| 410 |
+
).to(torch.float)
|
| 411 |
+
model = BioGptForSequenceClassification(config)
|
| 412 |
+
model.to(torch_device)
|
| 413 |
+
model.eval()
|
| 414 |
+
result = model(input_ids, attention_mask=attention_mask, labels=sequence_labels)
|
| 415 |
+
self.assertEqual(result.logits.shape, (self.model_tester.batch_size, self.model_tester.num_labels))
|
| 416 |
+
|
| 417 |
+
|
| 418 |
+
@require_torch
|
| 419 |
+
class BioGptModelIntegrationTest(unittest.TestCase):
|
| 420 |
+
@slow
|
| 421 |
+
def test_inference_lm_head_model(self):
|
| 422 |
+
model = BioGptForCausalLM.from_pretrained("microsoft/biogpt")
|
| 423 |
+
input_ids = torch.tensor([[2, 4805, 9, 656, 21]])
|
| 424 |
+
output = model(input_ids)[0]
|
| 425 |
+
|
| 426 |
+
vocab_size = 42384
|
| 427 |
+
|
| 428 |
+
expected_shape = torch.Size((1, 5, vocab_size))
|
| 429 |
+
self.assertEqual(output.shape, expected_shape)
|
| 430 |
+
|
| 431 |
+
expected_slice = torch.tensor(
|
| 432 |
+
[[[-9.5236, -9.8918, 10.4557], [-11.0469, -9.6423, 8.1022], [-8.8664, -7.8826, 5.5325]]]
|
| 433 |
+
)
|
| 434 |
+
|
| 435 |
+
self.assertTrue(torch.allclose(output[:, :3, :3], expected_slice, atol=1e-4))
|
| 436 |
+
|
| 437 |
+
@slow
|
| 438 |
+
def test_biogpt_generation(self):
|
| 439 |
+
tokenizer = BioGptTokenizer.from_pretrained("microsoft/biogpt")
|
| 440 |
+
model = BioGptForCausalLM.from_pretrained("microsoft/biogpt")
|
| 441 |
+
model.to(torch_device)
|
| 442 |
+
|
| 443 |
+
torch.manual_seed(0)
|
| 444 |
+
tokenized = tokenizer("COVID-19 is", return_tensors="pt").to(torch_device)
|
| 445 |
+
output_ids = model.generate(
|
| 446 |
+
**tokenized,
|
| 447 |
+
min_length=100,
|
| 448 |
+
max_length=1024,
|
| 449 |
+
num_beams=5,
|
| 450 |
+
early_stopping=True,
|
| 451 |
+
)
|
| 452 |
+
output_str = tokenizer.decode(output_ids[0], skip_special_tokens=True)
|
| 453 |
+
|
| 454 |
+
EXPECTED_OUTPUT_STR = (
|
| 455 |
+
"COVID-19 is a global pandemic caused by severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2), the"
|
| 456 |
+
" causative agent of coronavirus disease 2019 (COVID-19), which has spread to more than 200 countries and"
|
| 457 |
+
" territories, including the United States (US), Canada, Australia, New Zealand, the United Kingdom (UK),"
|
| 458 |
+
" and the United States of America (USA), as of March 11, 2020, with more than 800,000 confirmed cases and"
|
| 459 |
+
" more than 800,000 deaths."
|
| 460 |
+
)
|
| 461 |
+
self.assertEqual(output_str, EXPECTED_OUTPUT_STR)
|
tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/biogpt/test_tokenization_biogpt.py
ADDED
|
@@ -0,0 +1,99 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# coding=utf-8
|
| 2 |
+
# Copyright 2022 The HuggingFace Team. All rights reserved.
|
| 3 |
+
#
|
| 4 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
| 5 |
+
# you may not use this file except in compliance with the License.
|
| 6 |
+
# You may obtain a copy of the License at
|
| 7 |
+
#
|
| 8 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
| 9 |
+
#
|
| 10 |
+
# Unless required by applicable law or agreed to in writing, software
|
| 11 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
| 12 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
| 13 |
+
# See the License for the specific language governing permissions and
|
| 14 |
+
# limitations under the License.
|
| 15 |
+
|
| 16 |
+
|
| 17 |
+
import json
|
| 18 |
+
import os
|
| 19 |
+
import unittest
|
| 20 |
+
|
| 21 |
+
from transformers.models.biogpt.tokenization_biogpt import VOCAB_FILES_NAMES, BioGptTokenizer
|
| 22 |
+
from transformers.testing_utils import require_sacremoses, slow
|
| 23 |
+
|
| 24 |
+
from ...test_tokenization_common import TokenizerTesterMixin
|
| 25 |
+
|
| 26 |
+
|
| 27 |
+
@require_sacremoses
|
| 28 |
+
class BioGptTokenizationTest(TokenizerTesterMixin, unittest.TestCase):
|
| 29 |
+
from_pretrained_id = "microsoft/biogpt"
|
| 30 |
+
tokenizer_class = BioGptTokenizer
|
| 31 |
+
test_rust_tokenizer = False
|
| 32 |
+
|
| 33 |
+
def setUp(self):
|
| 34 |
+
super().setUp()
|
| 35 |
+
|
| 36 |
+
# Adapted from Sennrich et al. 2015 and https://github.com/rsennrich/subword-nmt
|
| 37 |
+
vocab = [
|
| 38 |
+
"l",
|
| 39 |
+
"o",
|
| 40 |
+
"w",
|
| 41 |
+
"e",
|
| 42 |
+
"r",
|
| 43 |
+
"s",
|
| 44 |
+
"t",
|
| 45 |
+
"i",
|
| 46 |
+
"d",
|
| 47 |
+
"n",
|
| 48 |
+
"w</w>",
|
| 49 |
+
"r</w>",
|
| 50 |
+
"t</w>",
|
| 51 |
+
"lo",
|
| 52 |
+
"low",
|
| 53 |
+
"er</w>",
|
| 54 |
+
"low</w>",
|
| 55 |
+
"lowest</w>",
|
| 56 |
+
"newer</w>",
|
| 57 |
+
"wider</w>",
|
| 58 |
+
"<unk>",
|
| 59 |
+
]
|
| 60 |
+
vocab_tokens = dict(zip(vocab, range(len(vocab))))
|
| 61 |
+
merges = ["l o 123", "lo w 1456", "e r</w> 1789", ""]
|
| 62 |
+
|
| 63 |
+
self.vocab_file = os.path.join(self.tmpdirname, VOCAB_FILES_NAMES["vocab_file"])
|
| 64 |
+
self.merges_file = os.path.join(self.tmpdirname, VOCAB_FILES_NAMES["merges_file"])
|
| 65 |
+
with open(self.vocab_file, "w") as fp:
|
| 66 |
+
fp.write(json.dumps(vocab_tokens))
|
| 67 |
+
with open(self.merges_file, "w") as fp:
|
| 68 |
+
fp.write("\n".join(merges))
|
| 69 |
+
|
| 70 |
+
def get_input_output_texts(self, tokenizer):
|
| 71 |
+
input_text = "lower newer"
|
| 72 |
+
output_text = "lower newer"
|
| 73 |
+
return input_text, output_text
|
| 74 |
+
|
| 75 |
+
def test_full_tokenizer(self):
|
| 76 |
+
"""Adapted from Sennrich et al. 2015 and https://github.com/rsennrich/subword-nmt"""
|
| 77 |
+
tokenizer = BioGptTokenizer(self.vocab_file, self.merges_file)
|
| 78 |
+
|
| 79 |
+
text = "lower"
|
| 80 |
+
bpe_tokens = ["low", "er</w>"]
|
| 81 |
+
tokens = tokenizer.tokenize(text)
|
| 82 |
+
self.assertListEqual(tokens, bpe_tokens)
|
| 83 |
+
|
| 84 |
+
input_tokens = tokens + ["<unk>"]
|
| 85 |
+
input_bpe_tokens = [14, 15, 20]
|
| 86 |
+
self.assertListEqual(tokenizer.convert_tokens_to_ids(input_tokens), input_bpe_tokens)
|
| 87 |
+
|
| 88 |
+
@slow
|
| 89 |
+
def test_sequence_builders(self):
|
| 90 |
+
tokenizer = BioGptTokenizer.from_pretrained("microsoft/biogpt")
|
| 91 |
+
|
| 92 |
+
text = tokenizer.encode("sequence builders", add_special_tokens=False)
|
| 93 |
+
text_2 = tokenizer.encode("multi-sequence build", add_special_tokens=False)
|
| 94 |
+
|
| 95 |
+
encoded_sentence = tokenizer.build_inputs_with_special_tokens(text)
|
| 96 |
+
encoded_pair = tokenizer.build_inputs_with_special_tokens(text, text_2)
|
| 97 |
+
|
| 98 |
+
self.assertTrue(encoded_sentence == [2] + text)
|
| 99 |
+
self.assertTrue(encoded_pair == [2] + text + [2] + text_2)
|
tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/bit/__init__.py
ADDED
|
File without changes
|
tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/bit/test_modeling_bit.py
ADDED
|
@@ -0,0 +1,310 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# coding=utf-8
|
| 2 |
+
# Copyright 2022 The HuggingFace Inc. team. All rights reserved.
|
| 3 |
+
#
|
| 4 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
| 5 |
+
# you may not use this file except in compliance with the License.
|
| 6 |
+
# You may obtain a copy of the License at
|
| 7 |
+
#
|
| 8 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
| 9 |
+
#
|
| 10 |
+
# Unless required by applicable law or agreed to in writing, software
|
| 11 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
| 12 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
| 13 |
+
# See the License for the specific language governing permissions and
|
| 14 |
+
# limitations under the License.
|
| 15 |
+
"""Testing suite for the PyTorch Bit model."""
|
| 16 |
+
|
| 17 |
+
import unittest
|
| 18 |
+
|
| 19 |
+
from transformers import BitConfig
|
| 20 |
+
from transformers.testing_utils import require_torch, require_vision, slow, torch_device
|
| 21 |
+
from transformers.utils import cached_property, is_torch_available, is_vision_available
|
| 22 |
+
|
| 23 |
+
from ...test_backbone_common import BackboneTesterMixin
|
| 24 |
+
from ...test_configuration_common import ConfigTester
|
| 25 |
+
from ...test_modeling_common import ModelTesterMixin, floats_tensor, ids_tensor
|
| 26 |
+
from ...test_pipeline_mixin import PipelineTesterMixin
|
| 27 |
+
|
| 28 |
+
|
| 29 |
+
if is_torch_available():
|
| 30 |
+
import torch
|
| 31 |
+
from torch import nn
|
| 32 |
+
|
| 33 |
+
from transformers import BitBackbone, BitForImageClassification, BitImageProcessor, BitModel
|
| 34 |
+
|
| 35 |
+
|
| 36 |
+
if is_vision_available():
|
| 37 |
+
from PIL import Image
|
| 38 |
+
|
| 39 |
+
|
| 40 |
+
class BitModelTester:
|
| 41 |
+
def __init__(
|
| 42 |
+
self,
|
| 43 |
+
parent,
|
| 44 |
+
batch_size=3,
|
| 45 |
+
image_size=32,
|
| 46 |
+
num_channels=3,
|
| 47 |
+
embeddings_size=10,
|
| 48 |
+
hidden_sizes=[8, 16, 32, 64],
|
| 49 |
+
depths=[1, 1, 2, 1],
|
| 50 |
+
is_training=True,
|
| 51 |
+
use_labels=True,
|
| 52 |
+
hidden_act="relu",
|
| 53 |
+
num_labels=3,
|
| 54 |
+
scope=None,
|
| 55 |
+
out_features=["stage2", "stage3", "stage4"],
|
| 56 |
+
out_indices=[2, 3, 4],
|
| 57 |
+
num_groups=1,
|
| 58 |
+
):
|
| 59 |
+
self.parent = parent
|
| 60 |
+
self.batch_size = batch_size
|
| 61 |
+
self.image_size = image_size
|
| 62 |
+
self.num_channels = num_channels
|
| 63 |
+
self.embeddings_size = embeddings_size
|
| 64 |
+
self.hidden_sizes = hidden_sizes
|
| 65 |
+
self.depths = depths
|
| 66 |
+
self.is_training = is_training
|
| 67 |
+
self.use_labels = use_labels
|
| 68 |
+
self.hidden_act = hidden_act
|
| 69 |
+
self.num_labels = num_labels
|
| 70 |
+
self.scope = scope
|
| 71 |
+
self.num_stages = len(hidden_sizes)
|
| 72 |
+
self.out_features = out_features
|
| 73 |
+
self.out_indices = out_indices
|
| 74 |
+
self.num_groups = num_groups
|
| 75 |
+
|
| 76 |
+
def prepare_config_and_inputs(self):
|
| 77 |
+
pixel_values = floats_tensor([self.batch_size, self.num_channels, self.image_size, self.image_size])
|
| 78 |
+
|
| 79 |
+
labels = None
|
| 80 |
+
if self.use_labels:
|
| 81 |
+
labels = ids_tensor([self.batch_size], self.num_labels)
|
| 82 |
+
|
| 83 |
+
config = self.get_config()
|
| 84 |
+
|
| 85 |
+
return config, pixel_values, labels
|
| 86 |
+
|
| 87 |
+
def get_config(self):
|
| 88 |
+
return BitConfig(
|
| 89 |
+
num_channels=self.num_channels,
|
| 90 |
+
embeddings_size=self.embeddings_size,
|
| 91 |
+
hidden_sizes=self.hidden_sizes,
|
| 92 |
+
depths=self.depths,
|
| 93 |
+
hidden_act=self.hidden_act,
|
| 94 |
+
num_labels=self.num_labels,
|
| 95 |
+
out_features=self.out_features,
|
| 96 |
+
out_indices=self.out_indices,
|
| 97 |
+
num_groups=self.num_groups,
|
| 98 |
+
)
|
| 99 |
+
|
| 100 |
+
def create_and_check_model(self, config, pixel_values, labels):
|
| 101 |
+
model = BitModel(config=config)
|
| 102 |
+
model.to(torch_device)
|
| 103 |
+
model.eval()
|
| 104 |
+
result = model(pixel_values)
|
| 105 |
+
self.parent.assertEqual(
|
| 106 |
+
result.last_hidden_state.shape,
|
| 107 |
+
(self.batch_size, self.hidden_sizes[-1], self.image_size // 32, self.image_size // 32),
|
| 108 |
+
)
|
| 109 |
+
|
| 110 |
+
def create_and_check_for_image_classification(self, config, pixel_values, labels):
|
| 111 |
+
config.num_labels = self.num_labels
|
| 112 |
+
model = BitForImageClassification(config)
|
| 113 |
+
model.to(torch_device)
|
| 114 |
+
model.eval()
|
| 115 |
+
result = model(pixel_values, labels=labels)
|
| 116 |
+
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.num_labels))
|
| 117 |
+
|
| 118 |
+
def create_and_check_backbone(self, config, pixel_values, labels):
|
| 119 |
+
model = BitBackbone(config=config)
|
| 120 |
+
model.to(torch_device)
|
| 121 |
+
model.eval()
|
| 122 |
+
result = model(pixel_values)
|
| 123 |
+
|
| 124 |
+
# verify feature maps
|
| 125 |
+
self.parent.assertEqual(len(result.feature_maps), len(config.out_features))
|
| 126 |
+
self.parent.assertListEqual(list(result.feature_maps[0].shape), [self.batch_size, self.hidden_sizes[1], 4, 4])
|
| 127 |
+
|
| 128 |
+
# verify channels
|
| 129 |
+
self.parent.assertEqual(len(model.channels), len(config.out_features))
|
| 130 |
+
self.parent.assertListEqual(model.channels, config.hidden_sizes[1:])
|
| 131 |
+
|
| 132 |
+
# verify backbone works with out_features=None
|
| 133 |
+
config.out_features = None
|
| 134 |
+
model = BitBackbone(config=config)
|
| 135 |
+
model.to(torch_device)
|
| 136 |
+
model.eval()
|
| 137 |
+
result = model(pixel_values)
|
| 138 |
+
|
| 139 |
+
# verify feature maps
|
| 140 |
+
self.parent.assertEqual(len(result.feature_maps), 1)
|
| 141 |
+
self.parent.assertListEqual(list(result.feature_maps[0].shape), [self.batch_size, self.hidden_sizes[-1], 1, 1])
|
| 142 |
+
|
| 143 |
+
# verify channels
|
| 144 |
+
self.parent.assertEqual(len(model.channels), 1)
|
| 145 |
+
self.parent.assertListEqual(model.channels, [config.hidden_sizes[-1]])
|
| 146 |
+
|
| 147 |
+
def prepare_config_and_inputs_for_common(self):
|
| 148 |
+
config_and_inputs = self.prepare_config_and_inputs()
|
| 149 |
+
config, pixel_values, labels = config_and_inputs
|
| 150 |
+
inputs_dict = {"pixel_values": pixel_values}
|
| 151 |
+
return config, inputs_dict
|
| 152 |
+
|
| 153 |
+
|
| 154 |
+
@require_torch
|
| 155 |
+
class BitModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase):
|
| 156 |
+
"""
|
| 157 |
+
Here we also overwrite some of the tests of test_modeling_common.py, as Bit does not use input_ids, inputs_embeds,
|
| 158 |
+
attention_mask and seq_length.
|
| 159 |
+
"""
|
| 160 |
+
|
| 161 |
+
all_model_classes = (BitModel, BitForImageClassification, BitBackbone) if is_torch_available() else ()
|
| 162 |
+
pipeline_model_mapping = (
|
| 163 |
+
{"image-feature-extraction": BitModel, "image-classification": BitForImageClassification}
|
| 164 |
+
if is_torch_available()
|
| 165 |
+
else {}
|
| 166 |
+
)
|
| 167 |
+
|
| 168 |
+
fx_compatible = False
|
| 169 |
+
test_pruning = False
|
| 170 |
+
test_resize_embeddings = False
|
| 171 |
+
test_head_masking = False
|
| 172 |
+
has_attentions = False
|
| 173 |
+
|
| 174 |
+
def setUp(self):
|
| 175 |
+
self.model_tester = BitModelTester(self)
|
| 176 |
+
self.config_tester = ConfigTester(
|
| 177 |
+
self, config_class=BitConfig, has_text_modality=False, common_properties=["num_channels"]
|
| 178 |
+
)
|
| 179 |
+
|
| 180 |
+
def test_config(self):
|
| 181 |
+
self.config_tester.run_common_tests()
|
| 182 |
+
|
| 183 |
+
@unittest.skip(reason="Bit does not output attentions")
|
| 184 |
+
def test_attention_outputs(self):
|
| 185 |
+
pass
|
| 186 |
+
|
| 187 |
+
@unittest.skip(reason="Bit does not use inputs_embeds")
|
| 188 |
+
def test_inputs_embeds(self):
|
| 189 |
+
pass
|
| 190 |
+
|
| 191 |
+
@unittest.skip(reason="Bit does not support input and output embeddings")
|
| 192 |
+
def test_model_get_set_embeddings(self):
|
| 193 |
+
pass
|
| 194 |
+
|
| 195 |
+
def test_model(self):
|
| 196 |
+
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
| 197 |
+
self.model_tester.create_and_check_model(*config_and_inputs)
|
| 198 |
+
|
| 199 |
+
def test_backbone(self):
|
| 200 |
+
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
| 201 |
+
self.model_tester.create_and_check_backbone(*config_and_inputs)
|
| 202 |
+
|
| 203 |
+
def test_initialization(self):
|
| 204 |
+
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
| 205 |
+
|
| 206 |
+
for model_class in self.all_model_classes:
|
| 207 |
+
model = model_class(config=config)
|
| 208 |
+
for name, module in model.named_modules():
|
| 209 |
+
if isinstance(module, (nn.BatchNorm2d, nn.GroupNorm)):
|
| 210 |
+
self.assertTrue(
|
| 211 |
+
torch.all(module.weight == 1),
|
| 212 |
+
msg=f"Parameter {name} of model {model_class} seems not properly initialized",
|
| 213 |
+
)
|
| 214 |
+
self.assertTrue(
|
| 215 |
+
torch.all(module.bias == 0),
|
| 216 |
+
msg=f"Parameter {name} of model {model_class} seems not properly initialized",
|
| 217 |
+
)
|
| 218 |
+
|
| 219 |
+
def test_hidden_states_output(self):
|
| 220 |
+
def check_hidden_states_output(inputs_dict, config, model_class):
|
| 221 |
+
model = model_class(config)
|
| 222 |
+
model.to(torch_device)
|
| 223 |
+
model.eval()
|
| 224 |
+
|
| 225 |
+
with torch.no_grad():
|
| 226 |
+
outputs = model(**self._prepare_for_class(inputs_dict, model_class))
|
| 227 |
+
|
| 228 |
+
hidden_states = outputs.encoder_hidden_states if config.is_encoder_decoder else outputs.hidden_states
|
| 229 |
+
|
| 230 |
+
expected_num_stages = self.model_tester.num_stages
|
| 231 |
+
self.assertEqual(len(hidden_states), expected_num_stages + 1)
|
| 232 |
+
|
| 233 |
+
# Bit's feature maps are of shape (batch_size, num_channels, height, width)
|
| 234 |
+
self.assertListEqual(
|
| 235 |
+
list(hidden_states[0].shape[-2:]),
|
| 236 |
+
[self.model_tester.image_size // 4, self.model_tester.image_size // 4],
|
| 237 |
+
)
|
| 238 |
+
|
| 239 |
+
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
| 240 |
+
layers_type = ["preactivation", "bottleneck"]
|
| 241 |
+
for model_class in self.all_model_classes:
|
| 242 |
+
for layer_type in layers_type:
|
| 243 |
+
config.layer_type = layer_type
|
| 244 |
+
inputs_dict["output_hidden_states"] = True
|
| 245 |
+
check_hidden_states_output(inputs_dict, config, model_class)
|
| 246 |
+
|
| 247 |
+
# check that output_hidden_states also work using config
|
| 248 |
+
del inputs_dict["output_hidden_states"]
|
| 249 |
+
config.output_hidden_states = True
|
| 250 |
+
|
| 251 |
+
check_hidden_states_output(inputs_dict, config, model_class)
|
| 252 |
+
|
| 253 |
+
@unittest.skip(reason="Bit does not use feedforward chunking")
|
| 254 |
+
def test_feed_forward_chunking(self):
|
| 255 |
+
pass
|
| 256 |
+
|
| 257 |
+
def test_for_image_classification(self):
|
| 258 |
+
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
| 259 |
+
self.model_tester.create_and_check_for_image_classification(*config_and_inputs)
|
| 260 |
+
|
| 261 |
+
@slow
|
| 262 |
+
def test_model_from_pretrained(self):
|
| 263 |
+
model_name = "google/bit-50"
|
| 264 |
+
model = BitModel.from_pretrained(model_name)
|
| 265 |
+
self.assertIsNotNone(model)
|
| 266 |
+
|
| 267 |
+
|
| 268 |
+
# We will verify our results on an image of cute cats
|
| 269 |
+
def prepare_img():
|
| 270 |
+
image = Image.open("./tests/fixtures/tests_samples/COCO/000000039769.png")
|
| 271 |
+
return image
|
| 272 |
+
|
| 273 |
+
|
| 274 |
+
@require_torch
|
| 275 |
+
@require_vision
|
| 276 |
+
class BitModelIntegrationTest(unittest.TestCase):
|
| 277 |
+
@cached_property
|
| 278 |
+
def default_image_processor(self):
|
| 279 |
+
return BitImageProcessor.from_pretrained("google/bit-50") if is_vision_available() else None
|
| 280 |
+
|
| 281 |
+
@slow
|
| 282 |
+
def test_inference_image_classification_head(self):
|
| 283 |
+
model = BitForImageClassification.from_pretrained("google/bit-50").to(torch_device)
|
| 284 |
+
|
| 285 |
+
image_processor = self.default_image_processor
|
| 286 |
+
image = prepare_img()
|
| 287 |
+
inputs = image_processor(images=image, return_tensors="pt").to(torch_device)
|
| 288 |
+
|
| 289 |
+
# forward pass
|
| 290 |
+
with torch.no_grad():
|
| 291 |
+
outputs = model(**inputs)
|
| 292 |
+
|
| 293 |
+
# verify the logits
|
| 294 |
+
expected_shape = torch.Size((1, 1000))
|
| 295 |
+
self.assertEqual(outputs.logits.shape, expected_shape)
|
| 296 |
+
|
| 297 |
+
expected_slice = torch.tensor([[-0.6526, -0.5263, -1.4398]]).to(torch_device)
|
| 298 |
+
|
| 299 |
+
self.assertTrue(torch.allclose(outputs.logits[0, :3], expected_slice, atol=1e-4))
|
| 300 |
+
|
| 301 |
+
|
| 302 |
+
@require_torch
|
| 303 |
+
class BitBackboneTest(BackboneTesterMixin, unittest.TestCase):
|
| 304 |
+
all_model_classes = (BitBackbone,) if is_torch_available() else ()
|
| 305 |
+
config_class = BitConfig
|
| 306 |
+
|
| 307 |
+
has_attentions = False
|
| 308 |
+
|
| 309 |
+
def setUp(self):
|
| 310 |
+
self.model_tester = BitModelTester(self)
|
tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/blenderbot/__init__.py
ADDED
|
File without changes
|
tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/blenderbot/test_modeling_blenderbot.py
ADDED
|
@@ -0,0 +1,559 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# coding=utf-8
|
| 2 |
+
# Copyright 2021, The HuggingFace Inc. team. All rights reserved.
|
| 3 |
+
#
|
| 4 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
| 5 |
+
# you may not use this file except in compliance with the License.
|
| 6 |
+
# You may obtain a copy of the License at
|
| 7 |
+
#
|
| 8 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
| 9 |
+
#
|
| 10 |
+
# Unless required by applicable law or agreed to in writing, software
|
| 11 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
| 12 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
| 13 |
+
# See the License for the specific language governing permissions and
|
| 14 |
+
# limitations under the License.
|
| 15 |
+
"""Testing suite for the PyTorch Blenderbot model."""
|
| 16 |
+
|
| 17 |
+
import tempfile
|
| 18 |
+
import unittest
|
| 19 |
+
|
| 20 |
+
from transformers import BlenderbotConfig, is_torch_available
|
| 21 |
+
from transformers.testing_utils import (
|
| 22 |
+
backend_empty_cache,
|
| 23 |
+
require_sentencepiece,
|
| 24 |
+
require_tokenizers,
|
| 25 |
+
require_torch,
|
| 26 |
+
require_torch_fp16,
|
| 27 |
+
slow,
|
| 28 |
+
torch_device,
|
| 29 |
+
)
|
| 30 |
+
from transformers.utils import cached_property
|
| 31 |
+
|
| 32 |
+
from ...generation.test_utils import GenerationTesterMixin
|
| 33 |
+
from ...test_configuration_common import ConfigTester
|
| 34 |
+
from ...test_modeling_common import ModelTesterMixin, ids_tensor
|
| 35 |
+
from ...test_pipeline_mixin import PipelineTesterMixin
|
| 36 |
+
|
| 37 |
+
|
| 38 |
+
if is_torch_available():
|
| 39 |
+
import torch
|
| 40 |
+
|
| 41 |
+
from transformers import BlenderbotForConditionalGeneration, BlenderbotModel, BlenderbotTokenizer
|
| 42 |
+
from transformers.models.blenderbot.modeling_blenderbot import (
|
| 43 |
+
BlenderbotDecoder,
|
| 44 |
+
BlenderbotEncoder,
|
| 45 |
+
BlenderbotForCausalLM,
|
| 46 |
+
)
|
| 47 |
+
|
| 48 |
+
|
| 49 |
+
def prepare_blenderbot_inputs_dict(
|
| 50 |
+
config,
|
| 51 |
+
input_ids,
|
| 52 |
+
decoder_input_ids,
|
| 53 |
+
attention_mask=None,
|
| 54 |
+
decoder_attention_mask=None,
|
| 55 |
+
head_mask=None,
|
| 56 |
+
decoder_head_mask=None,
|
| 57 |
+
cross_attn_head_mask=None,
|
| 58 |
+
):
|
| 59 |
+
if attention_mask is None:
|
| 60 |
+
attention_mask = input_ids.ne(config.pad_token_id)
|
| 61 |
+
if decoder_attention_mask is None:
|
| 62 |
+
decoder_attention_mask = decoder_input_ids.ne(config.pad_token_id)
|
| 63 |
+
if head_mask is None:
|
| 64 |
+
head_mask = torch.ones(config.encoder_layers, config.encoder_attention_heads, device=torch_device)
|
| 65 |
+
if decoder_head_mask is None:
|
| 66 |
+
decoder_head_mask = torch.ones(config.decoder_layers, config.decoder_attention_heads, device=torch_device)
|
| 67 |
+
if cross_attn_head_mask is None:
|
| 68 |
+
cross_attn_head_mask = torch.ones(config.decoder_layers, config.decoder_attention_heads, device=torch_device)
|
| 69 |
+
return {
|
| 70 |
+
"input_ids": input_ids,
|
| 71 |
+
"decoder_input_ids": decoder_input_ids,
|
| 72 |
+
"attention_mask": attention_mask,
|
| 73 |
+
"decoder_attention_mask": attention_mask,
|
| 74 |
+
"head_mask": head_mask,
|
| 75 |
+
"decoder_head_mask": decoder_head_mask,
|
| 76 |
+
"cross_attn_head_mask": cross_attn_head_mask,
|
| 77 |
+
}
|
| 78 |
+
|
| 79 |
+
|
| 80 |
+
class BlenderbotModelTester:
|
| 81 |
+
def __init__(
|
| 82 |
+
self,
|
| 83 |
+
parent,
|
| 84 |
+
batch_size=13,
|
| 85 |
+
seq_length=7,
|
| 86 |
+
is_training=True,
|
| 87 |
+
use_labels=False,
|
| 88 |
+
vocab_size=99,
|
| 89 |
+
hidden_size=16,
|
| 90 |
+
num_hidden_layers=2,
|
| 91 |
+
num_attention_heads=4,
|
| 92 |
+
intermediate_size=4,
|
| 93 |
+
hidden_act="gelu",
|
| 94 |
+
hidden_dropout_prob=0.1,
|
| 95 |
+
attention_probs_dropout_prob=0.1,
|
| 96 |
+
max_position_embeddings=50,
|
| 97 |
+
eos_token_id=2,
|
| 98 |
+
pad_token_id=1,
|
| 99 |
+
bos_token_id=0,
|
| 100 |
+
):
|
| 101 |
+
self.parent = parent
|
| 102 |
+
self.batch_size = batch_size
|
| 103 |
+
self.seq_length = seq_length
|
| 104 |
+
self.is_training = is_training
|
| 105 |
+
self.use_labels = use_labels
|
| 106 |
+
self.vocab_size = vocab_size
|
| 107 |
+
self.hidden_size = hidden_size
|
| 108 |
+
self.num_hidden_layers = num_hidden_layers
|
| 109 |
+
self.num_attention_heads = num_attention_heads
|
| 110 |
+
self.intermediate_size = intermediate_size
|
| 111 |
+
self.hidden_act = hidden_act
|
| 112 |
+
self.hidden_dropout_prob = hidden_dropout_prob
|
| 113 |
+
self.attention_probs_dropout_prob = attention_probs_dropout_prob
|
| 114 |
+
self.max_position_embeddings = max_position_embeddings
|
| 115 |
+
self.eos_token_id = eos_token_id
|
| 116 |
+
self.pad_token_id = pad_token_id
|
| 117 |
+
self.bos_token_id = bos_token_id
|
| 118 |
+
|
| 119 |
+
def prepare_config_and_inputs(self):
|
| 120 |
+
input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size).clamp(
|
| 121 |
+
3,
|
| 122 |
+
)
|
| 123 |
+
input_ids[:, -1] = self.eos_token_id # Eos Token
|
| 124 |
+
|
| 125 |
+
decoder_input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
|
| 126 |
+
|
| 127 |
+
config = self.get_config()
|
| 128 |
+
inputs_dict = prepare_blenderbot_inputs_dict(config, input_ids, decoder_input_ids)
|
| 129 |
+
return config, inputs_dict
|
| 130 |
+
|
| 131 |
+
def get_config(self):
|
| 132 |
+
return BlenderbotConfig(
|
| 133 |
+
vocab_size=self.vocab_size,
|
| 134 |
+
d_model=self.hidden_size,
|
| 135 |
+
encoder_layers=self.num_hidden_layers,
|
| 136 |
+
decoder_layers=self.num_hidden_layers,
|
| 137 |
+
encoder_attention_heads=self.num_attention_heads,
|
| 138 |
+
decoder_attention_heads=self.num_attention_heads,
|
| 139 |
+
encoder_ffn_dim=self.intermediate_size,
|
| 140 |
+
decoder_ffn_dim=self.intermediate_size,
|
| 141 |
+
dropout=self.hidden_dropout_prob,
|
| 142 |
+
attention_dropout=self.attention_probs_dropout_prob,
|
| 143 |
+
max_position_embeddings=self.max_position_embeddings,
|
| 144 |
+
eos_token_id=self.eos_token_id,
|
| 145 |
+
bos_token_id=self.bos_token_id,
|
| 146 |
+
pad_token_id=self.pad_token_id,
|
| 147 |
+
)
|
| 148 |
+
|
| 149 |
+
def get_pipeline_config(self):
|
| 150 |
+
config = self.get_config()
|
| 151 |
+
config.max_position_embeddings = 100
|
| 152 |
+
config.vocab_size = 300
|
| 153 |
+
return config
|
| 154 |
+
|
| 155 |
+
def prepare_config_and_inputs_for_common(self):
|
| 156 |
+
config, inputs_dict = self.prepare_config_and_inputs()
|
| 157 |
+
return config, inputs_dict
|
| 158 |
+
|
| 159 |
+
def create_and_check_decoder_model_past_large_inputs(self, config, inputs_dict):
|
| 160 |
+
model = BlenderbotModel(config=config).get_decoder().to(torch_device).eval()
|
| 161 |
+
input_ids = inputs_dict["input_ids"]
|
| 162 |
+
attention_mask = inputs_dict["attention_mask"]
|
| 163 |
+
head_mask = inputs_dict["head_mask"]
|
| 164 |
+
|
| 165 |
+
# first forward pass
|
| 166 |
+
outputs = model(input_ids, attention_mask=attention_mask, head_mask=head_mask, use_cache=True)
|
| 167 |
+
|
| 168 |
+
output, past_key_values = outputs.to_tuple()
|
| 169 |
+
|
| 170 |
+
# create hypothetical multiple next token and extent to next_input_ids
|
| 171 |
+
next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size)
|
| 172 |
+
next_attn_mask = ids_tensor((self.batch_size, 3), 2)
|
| 173 |
+
|
| 174 |
+
# append to next input_ids and
|
| 175 |
+
next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
|
| 176 |
+
next_attention_mask = torch.cat([attention_mask, next_attn_mask], dim=-1)
|
| 177 |
+
|
| 178 |
+
output_from_no_past = model(next_input_ids, attention_mask=next_attention_mask)["last_hidden_state"]
|
| 179 |
+
output_from_past = model(next_tokens, attention_mask=next_attention_mask, past_key_values=past_key_values)[
|
| 180 |
+
"last_hidden_state"
|
| 181 |
+
]
|
| 182 |
+
|
| 183 |
+
# select random slice
|
| 184 |
+
random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()
|
| 185 |
+
output_from_no_past_slice = output_from_no_past[:, -3:, random_slice_idx].detach()
|
| 186 |
+
output_from_past_slice = output_from_past[:, :, random_slice_idx].detach()
|
| 187 |
+
|
| 188 |
+
self.parent.assertTrue(output_from_past_slice.shape[1] == next_tokens.shape[1])
|
| 189 |
+
|
| 190 |
+
# test that outputs are equal for slice
|
| 191 |
+
self.parent.assertTrue(torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3))
|
| 192 |
+
|
| 193 |
+
def check_encoder_decoder_model_standalone(self, config, inputs_dict):
|
| 194 |
+
model = BlenderbotModel(config=config).to(torch_device).eval()
|
| 195 |
+
outputs = model(**inputs_dict)
|
| 196 |
+
|
| 197 |
+
encoder_last_hidden_state = outputs.encoder_last_hidden_state
|
| 198 |
+
last_hidden_state = outputs.last_hidden_state
|
| 199 |
+
|
| 200 |
+
with tempfile.TemporaryDirectory() as tmpdirname:
|
| 201 |
+
encoder = model.get_encoder()
|
| 202 |
+
encoder.save_pretrained(tmpdirname)
|
| 203 |
+
encoder = BlenderbotEncoder.from_pretrained(tmpdirname).to(torch_device)
|
| 204 |
+
|
| 205 |
+
encoder_last_hidden_state_2 = encoder(inputs_dict["input_ids"], attention_mask=inputs_dict["attention_mask"])[
|
| 206 |
+
0
|
| 207 |
+
]
|
| 208 |
+
|
| 209 |
+
self.parent.assertTrue((encoder_last_hidden_state_2 - encoder_last_hidden_state).abs().max().item() < 1e-3)
|
| 210 |
+
|
| 211 |
+
with tempfile.TemporaryDirectory() as tmpdirname:
|
| 212 |
+
decoder = model.get_decoder()
|
| 213 |
+
decoder.save_pretrained(tmpdirname)
|
| 214 |
+
decoder = BlenderbotDecoder.from_pretrained(tmpdirname).to(torch_device)
|
| 215 |
+
|
| 216 |
+
last_hidden_state_2 = decoder(
|
| 217 |
+
input_ids=inputs_dict["decoder_input_ids"],
|
| 218 |
+
attention_mask=inputs_dict["decoder_attention_mask"],
|
| 219 |
+
encoder_hidden_states=encoder_last_hidden_state,
|
| 220 |
+
encoder_attention_mask=inputs_dict["attention_mask"],
|
| 221 |
+
)[0]
|
| 222 |
+
|
| 223 |
+
self.parent.assertTrue((last_hidden_state_2 - last_hidden_state).abs().max().item() < 1e-3)
|
| 224 |
+
|
| 225 |
+
|
| 226 |
+
@require_torch
|
| 227 |
+
class BlenderbotModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixin, unittest.TestCase):
|
| 228 |
+
all_model_classes = (BlenderbotModel, BlenderbotForConditionalGeneration) if is_torch_available() else ()
|
| 229 |
+
all_generative_model_classes = (BlenderbotForConditionalGeneration,) if is_torch_available() else ()
|
| 230 |
+
pipeline_model_mapping = (
|
| 231 |
+
{
|
| 232 |
+
"feature-extraction": BlenderbotModel,
|
| 233 |
+
"summarization": BlenderbotForConditionalGeneration,
|
| 234 |
+
"text-generation": BlenderbotForCausalLM,
|
| 235 |
+
"text2text-generation": BlenderbotForConditionalGeneration,
|
| 236 |
+
"translation": BlenderbotForConditionalGeneration,
|
| 237 |
+
}
|
| 238 |
+
if is_torch_available()
|
| 239 |
+
else {}
|
| 240 |
+
)
|
| 241 |
+
is_encoder_decoder = True
|
| 242 |
+
fx_compatible = True
|
| 243 |
+
test_pruning = False
|
| 244 |
+
test_missing_keys = False
|
| 245 |
+
|
| 246 |
+
def setUp(self):
|
| 247 |
+
self.model_tester = BlenderbotModelTester(self)
|
| 248 |
+
self.config_tester = ConfigTester(self, config_class=BlenderbotConfig)
|
| 249 |
+
|
| 250 |
+
def test_config(self):
|
| 251 |
+
self.config_tester.run_common_tests()
|
| 252 |
+
|
| 253 |
+
def test_save_load_strict(self):
|
| 254 |
+
config, inputs_dict = self.model_tester.prepare_config_and_inputs()
|
| 255 |
+
for model_class in self.all_model_classes:
|
| 256 |
+
model = model_class(config)
|
| 257 |
+
|
| 258 |
+
with tempfile.TemporaryDirectory() as tmpdirname:
|
| 259 |
+
model.save_pretrained(tmpdirname)
|
| 260 |
+
model2, info = model_class.from_pretrained(tmpdirname, output_loading_info=True)
|
| 261 |
+
self.assertEqual(info["missing_keys"], [])
|
| 262 |
+
|
| 263 |
+
def test_decoder_model_past_with_large_inputs(self):
|
| 264 |
+
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
| 265 |
+
self.model_tester.create_and_check_decoder_model_past_large_inputs(*config_and_inputs)
|
| 266 |
+
|
| 267 |
+
def test_encoder_decoder_model_standalone(self):
|
| 268 |
+
config_and_inputs = self.model_tester.prepare_config_and_inputs_for_common()
|
| 269 |
+
self.model_tester.check_encoder_decoder_model_standalone(*config_and_inputs)
|
| 270 |
+
|
| 271 |
+
@require_torch_fp16
|
| 272 |
+
def test_generate_fp16(self):
|
| 273 |
+
config, input_dict = self.model_tester.prepare_config_and_inputs()
|
| 274 |
+
input_ids = input_dict["input_ids"]
|
| 275 |
+
attention_mask = input_ids.ne(1).to(torch_device)
|
| 276 |
+
model = BlenderbotForConditionalGeneration(config).eval().to(torch_device)
|
| 277 |
+
model.half()
|
| 278 |
+
model.generate(input_ids, attention_mask=attention_mask)
|
| 279 |
+
model.generate(num_beams=4, do_sample=True, early_stopping=False, num_return_sequences=3)
|
| 280 |
+
|
| 281 |
+
|
| 282 |
+
def assert_tensors_close(a, b, atol=1e-12, prefix=""):
|
| 283 |
+
"""If tensors have different shapes, different values or a and b are not both tensors, raise a nice Assertion error."""
|
| 284 |
+
if a is None and b is None:
|
| 285 |
+
return True
|
| 286 |
+
try:
|
| 287 |
+
if torch.allclose(a, b, atol=atol):
|
| 288 |
+
return True
|
| 289 |
+
raise
|
| 290 |
+
except Exception:
|
| 291 |
+
pct_different = (torch.gt((a - b).abs(), atol)).float().mean().item()
|
| 292 |
+
if a.numel() > 100:
|
| 293 |
+
msg = f"tensor values are {pct_different:.1%} percent different."
|
| 294 |
+
else:
|
| 295 |
+
msg = f"{a} != {b}"
|
| 296 |
+
if prefix:
|
| 297 |
+
msg = prefix + ": " + msg
|
| 298 |
+
raise AssertionError(msg)
|
| 299 |
+
|
| 300 |
+
|
| 301 |
+
@unittest.skipUnless(torch_device != "cpu", "3B test too slow on CPU.")
|
| 302 |
+
@require_torch
|
| 303 |
+
@require_sentencepiece
|
| 304 |
+
@require_tokenizers
|
| 305 |
+
class Blenderbot3BIntegrationTests(unittest.TestCase):
|
| 306 |
+
ckpt = "facebook/blenderbot-3B"
|
| 307 |
+
|
| 308 |
+
@cached_property
|
| 309 |
+
def tokenizer(self):
|
| 310 |
+
return BlenderbotTokenizer.from_pretrained(self.ckpt)
|
| 311 |
+
|
| 312 |
+
@slow
|
| 313 |
+
def test_generation_from_short_input_same_as_parlai_3B(self):
|
| 314 |
+
FASTER_GEN_KWARGS = {"num_beams": 1, "early_stopping": True, "min_length": 15, "max_length": 25}
|
| 315 |
+
TOK_DECODE_KW = {"skip_special_tokens": True, "clean_up_tokenization_spaces": True}
|
| 316 |
+
|
| 317 |
+
backend_empty_cache(torch_device)
|
| 318 |
+
model = BlenderbotForConditionalGeneration.from_pretrained(self.ckpt).half().to(torch_device)
|
| 319 |
+
|
| 320 |
+
src_text = ["Sam"]
|
| 321 |
+
model_inputs = self.tokenizer(src_text, return_tensors="pt").to(torch_device)
|
| 322 |
+
|
| 323 |
+
generated_utterances = model.generate(**model_inputs, **FASTER_GEN_KWARGS)
|
| 324 |
+
tgt_text = 'Sam is a great name. It means "sun" in Gaelic.'
|
| 325 |
+
|
| 326 |
+
generated_txt = self.tokenizer.batch_decode(generated_utterances, **TOK_DECODE_KW)
|
| 327 |
+
assert generated_txt[0].strip() == tgt_text
|
| 328 |
+
|
| 329 |
+
src_text = (
|
| 330 |
+
"Social anxiety\nWow, I am never shy. Do you have anxiety?\nYes. I end up sweating and blushing and feel"
|
| 331 |
+
" like i'm going to throw up.\nand why is that?"
|
| 332 |
+
)
|
| 333 |
+
|
| 334 |
+
model_inputs = self.tokenizer([src_text], return_tensors="pt").to(torch_device)
|
| 335 |
+
|
| 336 |
+
generated_ids = model.generate(**model_inputs, **FASTER_GEN_KWARGS)[0]
|
| 337 |
+
reply = self.tokenizer.decode(generated_ids, **TOK_DECODE_KW)
|
| 338 |
+
|
| 339 |
+
assert "I think it's because we are so worried about what people think of us." == reply.strip()
|
| 340 |
+
del model
|
| 341 |
+
|
| 342 |
+
|
| 343 |
+
class BlenderbotStandaloneDecoderModelTester:
|
| 344 |
+
def __init__(
|
| 345 |
+
self,
|
| 346 |
+
parent,
|
| 347 |
+
vocab_size=99,
|
| 348 |
+
batch_size=13,
|
| 349 |
+
d_model=16,
|
| 350 |
+
decoder_seq_length=7,
|
| 351 |
+
is_training=True,
|
| 352 |
+
is_decoder=True,
|
| 353 |
+
use_attention_mask=True,
|
| 354 |
+
use_cache=False,
|
| 355 |
+
use_labels=True,
|
| 356 |
+
decoder_start_token_id=2,
|
| 357 |
+
decoder_ffn_dim=32,
|
| 358 |
+
decoder_layers=2,
|
| 359 |
+
encoder_attention_heads=4,
|
| 360 |
+
decoder_attention_heads=4,
|
| 361 |
+
max_position_embeddings=30,
|
| 362 |
+
is_encoder_decoder=False,
|
| 363 |
+
pad_token_id=0,
|
| 364 |
+
bos_token_id=1,
|
| 365 |
+
eos_token_id=2,
|
| 366 |
+
scope=None,
|
| 367 |
+
):
|
| 368 |
+
self.parent = parent
|
| 369 |
+
self.batch_size = batch_size
|
| 370 |
+
self.decoder_seq_length = decoder_seq_length
|
| 371 |
+
# For common tests
|
| 372 |
+
self.seq_length = self.decoder_seq_length
|
| 373 |
+
self.is_training = is_training
|
| 374 |
+
self.use_attention_mask = use_attention_mask
|
| 375 |
+
self.use_labels = use_labels
|
| 376 |
+
|
| 377 |
+
self.vocab_size = vocab_size
|
| 378 |
+
self.d_model = d_model
|
| 379 |
+
self.hidden_size = d_model
|
| 380 |
+
self.num_hidden_layers = decoder_layers
|
| 381 |
+
self.decoder_layers = decoder_layers
|
| 382 |
+
self.decoder_ffn_dim = decoder_ffn_dim
|
| 383 |
+
self.encoder_attention_heads = encoder_attention_heads
|
| 384 |
+
self.decoder_attention_heads = decoder_attention_heads
|
| 385 |
+
self.num_attention_heads = decoder_attention_heads
|
| 386 |
+
self.eos_token_id = eos_token_id
|
| 387 |
+
self.bos_token_id = bos_token_id
|
| 388 |
+
self.pad_token_id = pad_token_id
|
| 389 |
+
self.decoder_start_token_id = decoder_start_token_id
|
| 390 |
+
self.use_cache = use_cache
|
| 391 |
+
self.max_position_embeddings = max_position_embeddings
|
| 392 |
+
self.is_encoder_decoder = is_encoder_decoder
|
| 393 |
+
|
| 394 |
+
self.scope = None
|
| 395 |
+
self.decoder_key_length = decoder_seq_length
|
| 396 |
+
self.base_model_out_len = 2
|
| 397 |
+
self.decoder_attention_idx = 1
|
| 398 |
+
|
| 399 |
+
def prepare_config_and_inputs(self):
|
| 400 |
+
input_ids = ids_tensor([self.batch_size, self.decoder_seq_length], self.vocab_size)
|
| 401 |
+
|
| 402 |
+
attention_mask = None
|
| 403 |
+
if self.use_attention_mask:
|
| 404 |
+
attention_mask = ids_tensor([self.batch_size, self.decoder_seq_length], vocab_size=2)
|
| 405 |
+
|
| 406 |
+
lm_labels = None
|
| 407 |
+
if self.use_labels:
|
| 408 |
+
lm_labels = ids_tensor([self.batch_size, self.decoder_seq_length], self.vocab_size)
|
| 409 |
+
|
| 410 |
+
config = BlenderbotConfig(
|
| 411 |
+
vocab_size=self.vocab_size,
|
| 412 |
+
d_model=self.d_model,
|
| 413 |
+
decoder_layers=self.decoder_layers,
|
| 414 |
+
decoder_ffn_dim=self.decoder_ffn_dim,
|
| 415 |
+
encoder_attention_heads=self.encoder_attention_heads,
|
| 416 |
+
decoder_attention_heads=self.decoder_attention_heads,
|
| 417 |
+
eos_token_id=self.eos_token_id,
|
| 418 |
+
bos_token_id=self.bos_token_id,
|
| 419 |
+
use_cache=self.use_cache,
|
| 420 |
+
pad_token_id=self.pad_token_id,
|
| 421 |
+
decoder_start_token_id=self.decoder_start_token_id,
|
| 422 |
+
max_position_embeddings=self.max_position_embeddings,
|
| 423 |
+
is_encoder_decoder=self.is_encoder_decoder,
|
| 424 |
+
)
|
| 425 |
+
|
| 426 |
+
return (
|
| 427 |
+
config,
|
| 428 |
+
input_ids,
|
| 429 |
+
attention_mask,
|
| 430 |
+
lm_labels,
|
| 431 |
+
)
|
| 432 |
+
|
| 433 |
+
def create_and_check_decoder_model_past(
|
| 434 |
+
self,
|
| 435 |
+
config,
|
| 436 |
+
input_ids,
|
| 437 |
+
attention_mask,
|
| 438 |
+
lm_labels,
|
| 439 |
+
):
|
| 440 |
+
config.use_cache = True
|
| 441 |
+
model = BlenderbotDecoder(config=config).to(torch_device).eval()
|
| 442 |
+
# first forward pass
|
| 443 |
+
outputs = model(input_ids, use_cache=True)
|
| 444 |
+
outputs_use_cache_conf = model(input_ids)
|
| 445 |
+
outputs_no_past = model(input_ids, use_cache=False)
|
| 446 |
+
|
| 447 |
+
self.parent.assertTrue(len(outputs) == len(outputs_use_cache_conf))
|
| 448 |
+
self.parent.assertTrue(len(outputs) == len(outputs_no_past) + 1)
|
| 449 |
+
|
| 450 |
+
past_key_values = outputs["past_key_values"]
|
| 451 |
+
|
| 452 |
+
# create hypothetical next token and extent to next_input_ids
|
| 453 |
+
next_tokens = ids_tensor((self.batch_size, 1), config.vocab_size)
|
| 454 |
+
|
| 455 |
+
# append to next input_ids and
|
| 456 |
+
next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
|
| 457 |
+
|
| 458 |
+
output_from_no_past = model(next_input_ids)["last_hidden_state"]
|
| 459 |
+
output_from_past = model(next_tokens, past_key_values=past_key_values)["last_hidden_state"]
|
| 460 |
+
|
| 461 |
+
# select random slice
|
| 462 |
+
random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()
|
| 463 |
+
output_from_no_past_slice = output_from_no_past[:, next_input_ids.shape[-1] - 1, random_slice_idx].detach()
|
| 464 |
+
output_from_past_slice = output_from_past[:, 0, random_slice_idx].detach()
|
| 465 |
+
|
| 466 |
+
# test that outputs are equal for slice
|
| 467 |
+
assert torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3)
|
| 468 |
+
|
| 469 |
+
def create_and_check_decoder_model_attention_mask_past(
|
| 470 |
+
self,
|
| 471 |
+
config,
|
| 472 |
+
input_ids,
|
| 473 |
+
attention_mask,
|
| 474 |
+
lm_labels,
|
| 475 |
+
):
|
| 476 |
+
model = BlenderbotDecoder(config=config).to(torch_device).eval()
|
| 477 |
+
|
| 478 |
+
# create attention mask
|
| 479 |
+
attn_mask = torch.ones(input_ids.shape, dtype=torch.long, device=torch_device)
|
| 480 |
+
|
| 481 |
+
half_seq_length = input_ids.shape[-1] // 2
|
| 482 |
+
attn_mask[:, half_seq_length:] = 0
|
| 483 |
+
|
| 484 |
+
# first forward pass
|
| 485 |
+
past_key_values = model(input_ids, attention_mask=attn_mask, use_cache=True)["past_key_values"]
|
| 486 |
+
# past_key_values = model(input_ids, use_cache=True)["past_key_values"]
|
| 487 |
+
|
| 488 |
+
# create hypothetical next token and extent to next_input_ids
|
| 489 |
+
next_tokens = ids_tensor((self.batch_size, 1), config.vocab_size)
|
| 490 |
+
|
| 491 |
+
# change a random masked slice from input_ids
|
| 492 |
+
random_seq_idx_to_change = ids_tensor((1,), half_seq_length).item() + 1
|
| 493 |
+
random_other_next_tokens = ids_tensor((self.batch_size, 1), config.vocab_size).squeeze(-1)
|
| 494 |
+
input_ids[:, -random_seq_idx_to_change] = random_other_next_tokens
|
| 495 |
+
|
| 496 |
+
# append to next input_ids and attn_mask
|
| 497 |
+
next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
|
| 498 |
+
attn_mask = torch.cat(
|
| 499 |
+
[attn_mask, torch.ones((attn_mask.shape[0], 1), dtype=torch.long, device=torch_device)],
|
| 500 |
+
dim=1,
|
| 501 |
+
)
|
| 502 |
+
|
| 503 |
+
# get two different outputs
|
| 504 |
+
output_from_no_past = model(next_input_ids, attention_mask=attn_mask)["last_hidden_state"]
|
| 505 |
+
output_from_past = model(next_tokens, past_key_values=past_key_values, attention_mask=attn_mask)[
|
| 506 |
+
"last_hidden_state"
|
| 507 |
+
]
|
| 508 |
+
|
| 509 |
+
# select random slice
|
| 510 |
+
random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()
|
| 511 |
+
output_from_no_past_slice = output_from_no_past[:, next_input_ids.shape[-1] - 1, random_slice_idx].detach()
|
| 512 |
+
output_from_past_slice = output_from_past[:, 0, random_slice_idx].detach()
|
| 513 |
+
|
| 514 |
+
# test that outputs are equal for slice
|
| 515 |
+
assert torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3)
|
| 516 |
+
|
| 517 |
+
def prepare_config_and_inputs_for_common(self):
|
| 518 |
+
config_and_inputs = self.prepare_config_and_inputs()
|
| 519 |
+
(
|
| 520 |
+
config,
|
| 521 |
+
input_ids,
|
| 522 |
+
attention_mask,
|
| 523 |
+
lm_labels,
|
| 524 |
+
) = config_and_inputs
|
| 525 |
+
|
| 526 |
+
inputs_dict = {
|
| 527 |
+
"input_ids": input_ids,
|
| 528 |
+
"attention_mask": attention_mask,
|
| 529 |
+
}
|
| 530 |
+
return config, inputs_dict
|
| 531 |
+
|
| 532 |
+
|
| 533 |
+
@require_torch
|
| 534 |
+
class BlenderbotStandaloneDecoderModelTest(ModelTesterMixin, GenerationTesterMixin, unittest.TestCase):
|
| 535 |
+
all_model_classes = (BlenderbotDecoder, BlenderbotForCausalLM) if is_torch_available() else ()
|
| 536 |
+
all_generative_model_classes = (BlenderbotForCausalLM,) if is_torch_available() else ()
|
| 537 |
+
test_pruning = False
|
| 538 |
+
is_encoder_decoder = False
|
| 539 |
+
|
| 540 |
+
def setUp(
|
| 541 |
+
self,
|
| 542 |
+
):
|
| 543 |
+
self.model_tester = BlenderbotStandaloneDecoderModelTester(self, is_training=False)
|
| 544 |
+
self.config_tester = ConfigTester(self, config_class=BlenderbotConfig)
|
| 545 |
+
|
| 546 |
+
def test_config(self):
|
| 547 |
+
self.config_tester.run_common_tests()
|
| 548 |
+
|
| 549 |
+
def test_decoder_model_past(self):
|
| 550 |
+
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
| 551 |
+
self.model_tester.create_and_check_decoder_model_past(*config_and_inputs)
|
| 552 |
+
|
| 553 |
+
def test_decoder_model_attn_mask_past(self):
|
| 554 |
+
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
| 555 |
+
self.model_tester.create_and_check_decoder_model_attention_mask_past(*config_and_inputs)
|
| 556 |
+
|
| 557 |
+
@unittest.skip(reason="decoder cannot keep gradients")
|
| 558 |
+
def test_retain_grad_hidden_states_attentions(self):
|
| 559 |
+
return
|
tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/blenderbot/test_modeling_flax_blenderbot.py
ADDED
|
@@ -0,0 +1,418 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Copyright 2021 The HuggingFace Team. All rights reserved.
|
| 2 |
+
#
|
| 3 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
| 4 |
+
# you may not use this file except in compliance with the License.
|
| 5 |
+
# You may obtain a copy of the License at
|
| 6 |
+
#
|
| 7 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
| 8 |
+
#
|
| 9 |
+
# Unless required by applicable law or agreed to in writing, software
|
| 10 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
| 11 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
| 12 |
+
# See the License for the specific language governing permissions and
|
| 13 |
+
# limitations under the License.
|
| 14 |
+
|
| 15 |
+
import unittest
|
| 16 |
+
|
| 17 |
+
import numpy as np
|
| 18 |
+
import timeout_decorator # noqa
|
| 19 |
+
|
| 20 |
+
from transformers import BlenderbotConfig, is_flax_available
|
| 21 |
+
from transformers.testing_utils import jax_device, require_flax, slow
|
| 22 |
+
|
| 23 |
+
from ...generation.test_flax_utils import FlaxGenerationTesterMixin
|
| 24 |
+
from ...test_modeling_flax_common import FlaxModelTesterMixin, ids_tensor
|
| 25 |
+
|
| 26 |
+
|
| 27 |
+
if is_flax_available():
|
| 28 |
+
import os
|
| 29 |
+
|
| 30 |
+
# The slow tests are often failing with OOM error on GPU
|
| 31 |
+
# This makes JAX allocate exactly what is needed on demand, and deallocate memory that is no longer needed
|
| 32 |
+
# but will be slower as stated here https://jax.readthedocs.io/en/latest/gpu_memory_allocation.html
|
| 33 |
+
os.environ["XLA_PYTHON_CLIENT_ALLOCATOR"] = "platform"
|
| 34 |
+
|
| 35 |
+
import jax
|
| 36 |
+
import jax.numpy as jnp
|
| 37 |
+
|
| 38 |
+
from transformers import BlenderbotTokenizer
|
| 39 |
+
from transformers.models.blenderbot.modeling_flax_blenderbot import (
|
| 40 |
+
FlaxBlenderbotForConditionalGeneration,
|
| 41 |
+
FlaxBlenderbotModel,
|
| 42 |
+
shift_tokens_right,
|
| 43 |
+
)
|
| 44 |
+
|
| 45 |
+
|
| 46 |
+
def prepare_blenderbot_inputs_dict(
|
| 47 |
+
config,
|
| 48 |
+
input_ids,
|
| 49 |
+
decoder_input_ids=None,
|
| 50 |
+
attention_mask=None,
|
| 51 |
+
decoder_attention_mask=None,
|
| 52 |
+
head_mask=None,
|
| 53 |
+
decoder_head_mask=None,
|
| 54 |
+
cross_attn_head_mask=None,
|
| 55 |
+
):
|
| 56 |
+
if attention_mask is None:
|
| 57 |
+
attention_mask = np.where(input_ids != config.pad_token_id, 1, 0)
|
| 58 |
+
if decoder_attention_mask is None:
|
| 59 |
+
decoder_attention_mask = np.where(decoder_input_ids != config.pad_token_id, 1, 0)
|
| 60 |
+
if head_mask is None:
|
| 61 |
+
head_mask = np.ones((config.encoder_layers, config.encoder_attention_heads))
|
| 62 |
+
if decoder_head_mask is None:
|
| 63 |
+
decoder_head_mask = np.ones((config.decoder_layers, config.decoder_attention_heads))
|
| 64 |
+
if cross_attn_head_mask is None:
|
| 65 |
+
cross_attn_head_mask = np.ones((config.decoder_layers, config.decoder_attention_heads))
|
| 66 |
+
return {
|
| 67 |
+
"input_ids": input_ids,
|
| 68 |
+
"decoder_input_ids": decoder_input_ids,
|
| 69 |
+
"attention_mask": attention_mask,
|
| 70 |
+
"decoder_attention_mask": attention_mask,
|
| 71 |
+
}
|
| 72 |
+
|
| 73 |
+
|
| 74 |
+
class FlaxBlenderbotModelTester:
|
| 75 |
+
def __init__(
|
| 76 |
+
self,
|
| 77 |
+
parent,
|
| 78 |
+
batch_size=13,
|
| 79 |
+
seq_length=7,
|
| 80 |
+
is_training=True,
|
| 81 |
+
use_labels=False,
|
| 82 |
+
vocab_size=99,
|
| 83 |
+
hidden_size=16,
|
| 84 |
+
num_hidden_layers=2,
|
| 85 |
+
num_attention_heads=4,
|
| 86 |
+
intermediate_size=4,
|
| 87 |
+
hidden_act="gelu",
|
| 88 |
+
hidden_dropout_prob=0.1,
|
| 89 |
+
attention_probs_dropout_prob=0.1,
|
| 90 |
+
max_position_embeddings=50,
|
| 91 |
+
eos_token_id=2,
|
| 92 |
+
pad_token_id=1,
|
| 93 |
+
bos_token_id=0,
|
| 94 |
+
initializer_range=0.02,
|
| 95 |
+
):
|
| 96 |
+
self.parent = parent
|
| 97 |
+
self.batch_size = batch_size
|
| 98 |
+
self.seq_length = seq_length
|
| 99 |
+
self.is_training = is_training
|
| 100 |
+
self.use_labels = use_labels
|
| 101 |
+
self.vocab_size = vocab_size
|
| 102 |
+
self.hidden_size = hidden_size
|
| 103 |
+
self.num_hidden_layers = num_hidden_layers
|
| 104 |
+
self.num_attention_heads = num_attention_heads
|
| 105 |
+
self.intermediate_size = intermediate_size
|
| 106 |
+
self.hidden_act = hidden_act
|
| 107 |
+
self.hidden_dropout_prob = hidden_dropout_prob
|
| 108 |
+
self.attention_probs_dropout_prob = attention_probs_dropout_prob
|
| 109 |
+
self.max_position_embeddings = max_position_embeddings
|
| 110 |
+
self.eos_token_id = eos_token_id
|
| 111 |
+
self.pad_token_id = pad_token_id
|
| 112 |
+
self.bos_token_id = bos_token_id
|
| 113 |
+
self.initializer_range = initializer_range
|
| 114 |
+
|
| 115 |
+
def prepare_config_and_inputs(self):
|
| 116 |
+
input_ids = np.clip(ids_tensor([self.batch_size, self.seq_length - 1], self.vocab_size), 3, self.vocab_size)
|
| 117 |
+
input_ids = np.concatenate((input_ids, 2 * np.ones((self.batch_size, 1), dtype=np.int64)), -1)
|
| 118 |
+
|
| 119 |
+
decoder_input_ids = shift_tokens_right(input_ids, 1, 2)
|
| 120 |
+
|
| 121 |
+
config = BlenderbotConfig(
|
| 122 |
+
vocab_size=self.vocab_size,
|
| 123 |
+
d_model=self.hidden_size,
|
| 124 |
+
encoder_layers=self.num_hidden_layers,
|
| 125 |
+
decoder_layers=self.num_hidden_layers,
|
| 126 |
+
encoder_attention_heads=self.num_attention_heads,
|
| 127 |
+
decoder_attention_heads=self.num_attention_heads,
|
| 128 |
+
encoder_ffn_dim=self.intermediate_size,
|
| 129 |
+
decoder_ffn_dim=self.intermediate_size,
|
| 130 |
+
dropout=self.hidden_dropout_prob,
|
| 131 |
+
attention_dropout=self.attention_probs_dropout_prob,
|
| 132 |
+
max_position_embeddings=self.max_position_embeddings,
|
| 133 |
+
eos_token_id=self.eos_token_id,
|
| 134 |
+
bos_token_id=self.bos_token_id,
|
| 135 |
+
pad_token_id=self.pad_token_id,
|
| 136 |
+
initializer_range=self.initializer_range,
|
| 137 |
+
use_cache=False,
|
| 138 |
+
)
|
| 139 |
+
inputs_dict = prepare_blenderbot_inputs_dict(config, input_ids, decoder_input_ids)
|
| 140 |
+
return config, inputs_dict
|
| 141 |
+
|
| 142 |
+
def prepare_config_and_inputs_for_common(self):
|
| 143 |
+
config, inputs_dict = self.prepare_config_and_inputs()
|
| 144 |
+
return config, inputs_dict
|
| 145 |
+
|
| 146 |
+
def check_use_cache_forward(self, model_class_name, config, inputs_dict):
|
| 147 |
+
max_decoder_length = 20
|
| 148 |
+
model = model_class_name(config)
|
| 149 |
+
|
| 150 |
+
encoder_outputs = model.encode(inputs_dict["input_ids"])
|
| 151 |
+
|
| 152 |
+
decoder_input_ids, decoder_attention_mask = (
|
| 153 |
+
inputs_dict["decoder_input_ids"],
|
| 154 |
+
inputs_dict["decoder_attention_mask"],
|
| 155 |
+
)
|
| 156 |
+
|
| 157 |
+
past_key_values = model.init_cache(decoder_input_ids.shape[0], max_decoder_length, encoder_outputs)
|
| 158 |
+
decoder_attention_mask = jnp.ones((decoder_input_ids.shape[0], max_decoder_length), dtype="i4")
|
| 159 |
+
|
| 160 |
+
decoder_position_ids = jnp.broadcast_to(
|
| 161 |
+
jnp.arange(decoder_input_ids.shape[-1] - 1)[None, :],
|
| 162 |
+
(decoder_input_ids.shape[0], decoder_input_ids.shape[-1] - 1),
|
| 163 |
+
)
|
| 164 |
+
outputs_cache = model.decode(
|
| 165 |
+
decoder_input_ids[:, :-1],
|
| 166 |
+
encoder_outputs,
|
| 167 |
+
decoder_attention_mask=decoder_attention_mask,
|
| 168 |
+
past_key_values=past_key_values,
|
| 169 |
+
decoder_position_ids=decoder_position_ids,
|
| 170 |
+
)
|
| 171 |
+
|
| 172 |
+
decoder_position_ids = jnp.array(decoder_input_ids.shape[0] * [[decoder_input_ids.shape[-1] - 1]], dtype="i4")
|
| 173 |
+
outputs_cache_next = model.decode(
|
| 174 |
+
decoder_input_ids[:, -1:],
|
| 175 |
+
encoder_outputs,
|
| 176 |
+
decoder_attention_mask=decoder_attention_mask,
|
| 177 |
+
past_key_values=outputs_cache.past_key_values,
|
| 178 |
+
decoder_position_ids=decoder_position_ids,
|
| 179 |
+
)
|
| 180 |
+
|
| 181 |
+
outputs = model.decode(decoder_input_ids, encoder_outputs)
|
| 182 |
+
|
| 183 |
+
diff = np.max(np.abs((outputs_cache_next[0][:, -1, :5] - outputs[0][:, -1, :5])))
|
| 184 |
+
self.parent.assertTrue(diff < 1e-3, msg=f"Max diff is {diff}")
|
| 185 |
+
|
| 186 |
+
def check_use_cache_forward_with_attn_mask(self, model_class_name, config, inputs_dict):
|
| 187 |
+
max_decoder_length = 20
|
| 188 |
+
model = model_class_name(config)
|
| 189 |
+
|
| 190 |
+
encoder_outputs = model.encode(inputs_dict["input_ids"])
|
| 191 |
+
|
| 192 |
+
decoder_input_ids, decoder_attention_mask = (
|
| 193 |
+
inputs_dict["decoder_input_ids"],
|
| 194 |
+
inputs_dict["decoder_attention_mask"],
|
| 195 |
+
)
|
| 196 |
+
|
| 197 |
+
decoder_attention_mask_cache = jnp.concatenate(
|
| 198 |
+
[
|
| 199 |
+
decoder_attention_mask,
|
| 200 |
+
jnp.zeros((decoder_attention_mask.shape[0], max_decoder_length - decoder_attention_mask.shape[1])),
|
| 201 |
+
],
|
| 202 |
+
axis=-1,
|
| 203 |
+
)
|
| 204 |
+
|
| 205 |
+
past_key_values = model.init_cache(decoder_input_ids.shape[0], max_decoder_length, encoder_outputs)
|
| 206 |
+
decoder_position_ids = jnp.broadcast_to(
|
| 207 |
+
jnp.arange(decoder_input_ids.shape[-1] - 1)[None, :],
|
| 208 |
+
(decoder_input_ids.shape[0], decoder_input_ids.shape[-1] - 1),
|
| 209 |
+
)
|
| 210 |
+
|
| 211 |
+
outputs_cache = model.decode(
|
| 212 |
+
decoder_input_ids[:, :-1],
|
| 213 |
+
encoder_outputs,
|
| 214 |
+
decoder_attention_mask=decoder_attention_mask_cache,
|
| 215 |
+
past_key_values=past_key_values,
|
| 216 |
+
decoder_position_ids=decoder_position_ids,
|
| 217 |
+
)
|
| 218 |
+
decoder_position_ids = jnp.array(decoder_input_ids.shape[0] * [[decoder_input_ids.shape[-1] - 1]], dtype="i4")
|
| 219 |
+
outputs_cache_next = model.decode(
|
| 220 |
+
decoder_input_ids[:, -1:],
|
| 221 |
+
encoder_outputs,
|
| 222 |
+
past_key_values=outputs_cache.past_key_values,
|
| 223 |
+
decoder_attention_mask=decoder_attention_mask_cache,
|
| 224 |
+
decoder_position_ids=decoder_position_ids,
|
| 225 |
+
)
|
| 226 |
+
|
| 227 |
+
outputs = model.decode(decoder_input_ids, encoder_outputs, decoder_attention_mask=decoder_attention_mask)
|
| 228 |
+
|
| 229 |
+
diff = np.max(np.abs((outputs_cache_next[0][:, -1, :5] - outputs[0][:, -1, :5])))
|
| 230 |
+
self.parent.assertTrue(diff < 1e-3, msg=f"Max diff is {diff}")
|
| 231 |
+
|
| 232 |
+
|
| 233 |
+
@require_flax
|
| 234 |
+
class BlenderbotHeadTests(unittest.TestCase):
|
| 235 |
+
vocab_size = 99
|
| 236 |
+
|
| 237 |
+
def _get_config_and_data(self):
|
| 238 |
+
input_ids = np.array(
|
| 239 |
+
[
|
| 240 |
+
[71, 82, 18, 33, 46, 91, 2],
|
| 241 |
+
[68, 34, 26, 58, 30, 82, 2],
|
| 242 |
+
[5, 97, 17, 39, 94, 40, 2],
|
| 243 |
+
[76, 83, 94, 25, 70, 78, 2],
|
| 244 |
+
[87, 59, 41, 35, 48, 66, 2],
|
| 245 |
+
[55, 13, 16, 58, 5, 2, 1], # note padding
|
| 246 |
+
[64, 27, 31, 51, 12, 75, 2],
|
| 247 |
+
[52, 64, 86, 17, 83, 39, 2],
|
| 248 |
+
[48, 61, 9, 24, 71, 82, 2],
|
| 249 |
+
[26, 1, 60, 48, 22, 13, 2],
|
| 250 |
+
[21, 5, 62, 28, 14, 76, 2],
|
| 251 |
+
[45, 98, 37, 86, 59, 48, 2],
|
| 252 |
+
[70, 70, 50, 9, 28, 0, 2],
|
| 253 |
+
],
|
| 254 |
+
dtype=np.int64,
|
| 255 |
+
)
|
| 256 |
+
|
| 257 |
+
batch_size = input_ids.shape[0]
|
| 258 |
+
config = BlenderbotConfig(
|
| 259 |
+
vocab_size=self.vocab_size,
|
| 260 |
+
d_model=24,
|
| 261 |
+
encoder_layers=2,
|
| 262 |
+
decoder_layers=2,
|
| 263 |
+
encoder_attention_heads=2,
|
| 264 |
+
decoder_attention_heads=2,
|
| 265 |
+
encoder_ffn_dim=32,
|
| 266 |
+
decoder_ffn_dim=32,
|
| 267 |
+
max_position_embeddings=48,
|
| 268 |
+
eos_token_id=2,
|
| 269 |
+
pad_token_id=1,
|
| 270 |
+
bos_token_id=0,
|
| 271 |
+
)
|
| 272 |
+
return config, input_ids, batch_size
|
| 273 |
+
|
| 274 |
+
# @timeout_decorator.timeout(1) # not working with the decorator so far
|
| 275 |
+
def test_lm_forward(self):
|
| 276 |
+
config, input_ids, batch_size = self._get_config_and_data()
|
| 277 |
+
lm_model = FlaxBlenderbotForConditionalGeneration(config)
|
| 278 |
+
outputs = lm_model(input_ids=input_ids)
|
| 279 |
+
expected_shape = (batch_size, input_ids.shape[1], config.vocab_size)
|
| 280 |
+
self.assertEqual(outputs["logits"].shape, expected_shape)
|
| 281 |
+
|
| 282 |
+
def test_lm_uneven_forward(self):
|
| 283 |
+
config = BlenderbotConfig(
|
| 284 |
+
vocab_size=self.vocab_size,
|
| 285 |
+
d_model=14,
|
| 286 |
+
encoder_layers=2,
|
| 287 |
+
decoder_layers=2,
|
| 288 |
+
encoder_attention_heads=2,
|
| 289 |
+
decoder_attention_heads=2,
|
| 290 |
+
encoder_ffn_dim=8,
|
| 291 |
+
decoder_ffn_dim=8,
|
| 292 |
+
max_position_embeddings=48,
|
| 293 |
+
)
|
| 294 |
+
lm_model = FlaxBlenderbotForConditionalGeneration(config)
|
| 295 |
+
context = np.array([[71, 82, 18, 33, 46, 91, 2], [68, 34, 26, 58, 30, 2, 1]], dtype=np.int64)
|
| 296 |
+
summary = np.array([[82, 71, 82, 18, 2], [58, 68, 2, 1, 1]], dtype=np.int64)
|
| 297 |
+
outputs = lm_model(input_ids=context, decoder_input_ids=summary)
|
| 298 |
+
expected_shape = (*summary.shape, config.vocab_size)
|
| 299 |
+
self.assertEqual(outputs["logits"].shape, expected_shape)
|
| 300 |
+
|
| 301 |
+
def test_shift_tokens_right(self):
|
| 302 |
+
input_ids = np.array([[71, 82, 18, 33, 2, 1, 1], [68, 34, 26, 58, 30, 82, 2]], dtype=np.int64)
|
| 303 |
+
shifted = shift_tokens_right(input_ids, 1, 2)
|
| 304 |
+
n_pad_before = np.equal(input_ids, 1).astype(np.float32).sum()
|
| 305 |
+
n_pad_after = np.equal(shifted, 1).astype(np.float32).sum()
|
| 306 |
+
self.assertEqual(shifted.shape, input_ids.shape)
|
| 307 |
+
self.assertEqual(n_pad_after, n_pad_before - 1)
|
| 308 |
+
self.assertTrue(np.equal(shifted[:, 0], 2).all())
|
| 309 |
+
|
| 310 |
+
|
| 311 |
+
@require_flax
|
| 312 |
+
class FlaxBlenderbotModelTest(FlaxModelTesterMixin, unittest.TestCase, FlaxGenerationTesterMixin):
|
| 313 |
+
is_encoder_decoder = True
|
| 314 |
+
all_model_classes = (
|
| 315 |
+
(
|
| 316 |
+
FlaxBlenderbotModel,
|
| 317 |
+
FlaxBlenderbotForConditionalGeneration,
|
| 318 |
+
)
|
| 319 |
+
if is_flax_available()
|
| 320 |
+
else ()
|
| 321 |
+
)
|
| 322 |
+
all_generative_model_classes = (FlaxBlenderbotForConditionalGeneration,) if is_flax_available() else ()
|
| 323 |
+
|
| 324 |
+
def setUp(self):
|
| 325 |
+
self.model_tester = FlaxBlenderbotModelTester(self)
|
| 326 |
+
|
| 327 |
+
def test_use_cache_forward(self):
|
| 328 |
+
config, inputs_dict = self.model_tester.prepare_config_and_inputs()
|
| 329 |
+
for model_class in self.all_model_classes:
|
| 330 |
+
self.model_tester.check_use_cache_forward(model_class, config, inputs_dict)
|
| 331 |
+
|
| 332 |
+
def test_use_cache_forward_with_attn_mask(self):
|
| 333 |
+
config, inputs_dict = self.model_tester.prepare_config_and_inputs()
|
| 334 |
+
for model_class in self.all_model_classes:
|
| 335 |
+
self.model_tester.check_use_cache_forward_with_attn_mask(model_class, config, inputs_dict)
|
| 336 |
+
|
| 337 |
+
def test_encode(self):
|
| 338 |
+
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
| 339 |
+
|
| 340 |
+
for model_class in self.all_model_classes:
|
| 341 |
+
with self.subTest(model_class.__name__):
|
| 342 |
+
prepared_inputs_dict = self._prepare_for_class(inputs_dict, model_class)
|
| 343 |
+
model = model_class(config)
|
| 344 |
+
|
| 345 |
+
@jax.jit
|
| 346 |
+
def encode_jitted(input_ids, attention_mask=None, **kwargs):
|
| 347 |
+
return model.encode(input_ids=input_ids, attention_mask=attention_mask)
|
| 348 |
+
|
| 349 |
+
with self.subTest("JIT Enabled"):
|
| 350 |
+
jitted_outputs = encode_jitted(**prepared_inputs_dict).to_tuple()
|
| 351 |
+
|
| 352 |
+
with self.subTest("JIT Disabled"):
|
| 353 |
+
with jax.disable_jit():
|
| 354 |
+
outputs = encode_jitted(**prepared_inputs_dict).to_tuple()
|
| 355 |
+
|
| 356 |
+
self.assertEqual(len(outputs), len(jitted_outputs))
|
| 357 |
+
for jitted_output, output in zip(jitted_outputs, outputs):
|
| 358 |
+
self.assertEqual(jitted_output.shape, output.shape)
|
| 359 |
+
|
| 360 |
+
def test_decode(self):
|
| 361 |
+
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
| 362 |
+
|
| 363 |
+
for model_class in self.all_model_classes:
|
| 364 |
+
with self.subTest(model_class.__name__):
|
| 365 |
+
model = model_class(config)
|
| 366 |
+
encoder_outputs = model.encode(inputs_dict["input_ids"], inputs_dict["attention_mask"])
|
| 367 |
+
|
| 368 |
+
prepared_inputs_dict = {
|
| 369 |
+
"decoder_input_ids": inputs_dict["decoder_input_ids"],
|
| 370 |
+
"decoder_attention_mask": inputs_dict["decoder_attention_mask"],
|
| 371 |
+
"encoder_outputs": encoder_outputs,
|
| 372 |
+
}
|
| 373 |
+
|
| 374 |
+
@jax.jit
|
| 375 |
+
def decode_jitted(decoder_input_ids, decoder_attention_mask, encoder_outputs):
|
| 376 |
+
return model.decode(
|
| 377 |
+
decoder_input_ids=decoder_input_ids,
|
| 378 |
+
decoder_attention_mask=decoder_attention_mask,
|
| 379 |
+
encoder_outputs=encoder_outputs,
|
| 380 |
+
)
|
| 381 |
+
|
| 382 |
+
with self.subTest("JIT Enabled"):
|
| 383 |
+
jitted_outputs = decode_jitted(**prepared_inputs_dict).to_tuple()
|
| 384 |
+
|
| 385 |
+
with self.subTest("JIT Disabled"):
|
| 386 |
+
with jax.disable_jit():
|
| 387 |
+
outputs = decode_jitted(**prepared_inputs_dict).to_tuple()
|
| 388 |
+
|
| 389 |
+
self.assertEqual(len(outputs), len(jitted_outputs))
|
| 390 |
+
for jitted_output, output in zip(jitted_outputs, outputs):
|
| 391 |
+
self.assertEqual(jitted_output.shape, output.shape)
|
| 392 |
+
|
| 393 |
+
@slow
|
| 394 |
+
def test_model_from_pretrained(self):
|
| 395 |
+
for model_class_name in self.all_model_classes:
|
| 396 |
+
model = model_class_name.from_pretrained("facebook/blenderbot-400M-distill")
|
| 397 |
+
# FlaxBlenderbotForSequenceClassification expects eos token in input_ids
|
| 398 |
+
input_ids = np.ones((1, 1)) * model.config.eos_token_id
|
| 399 |
+
outputs = model(input_ids)
|
| 400 |
+
self.assertIsNotNone(outputs)
|
| 401 |
+
|
| 402 |
+
@unittest.skipUnless(jax_device != "cpu", "3B test too slow on CPU.")
|
| 403 |
+
@slow
|
| 404 |
+
def test_generation_from_short_input_same_as_parlai_3B(self):
|
| 405 |
+
FASTER_GEN_KWARGS = {"num_beams": 1, "early_stopping": True, "min_length": 15, "max_length": 25}
|
| 406 |
+
TOK_DECODE_KW = {"skip_special_tokens": True, "clean_up_tokenization_spaces": True}
|
| 407 |
+
|
| 408 |
+
model = FlaxBlenderbotForConditionalGeneration.from_pretrained("facebook/blenderbot-3B", from_pt=True)
|
| 409 |
+
tokenizer = BlenderbotTokenizer.from_pretrained("facebook/blenderbot-3B")
|
| 410 |
+
|
| 411 |
+
src_text = ["Sam"]
|
| 412 |
+
model_inputs = tokenizer(src_text, return_tensors="jax")
|
| 413 |
+
|
| 414 |
+
generated_utterances = model.generate(**model_inputs, **FASTER_GEN_KWARGS)
|
| 415 |
+
tgt_text = 'Sam is a great name. It means "sun" in Gaelic.'
|
| 416 |
+
|
| 417 |
+
generated_txt = tokenizer.batch_decode(generated_utterances, **TOK_DECODE_KW)
|
| 418 |
+
assert generated_txt[0].strip() == tgt_text
|
tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/blenderbot/test_modeling_tf_blenderbot.py
ADDED
|
@@ -0,0 +1,235 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# coding=utf-8
|
| 2 |
+
# Copyright 2021 The HuggingFace Inc. team. All rights reserved.
|
| 3 |
+
#
|
| 4 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
| 5 |
+
# you may not use this file except in compliance with the License.
|
| 6 |
+
# You may obtain a copy of the License at
|
| 7 |
+
#
|
| 8 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
| 9 |
+
#
|
| 10 |
+
# Unless required by applicable law or agreed to in writing, software
|
| 11 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
| 12 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
| 13 |
+
# See the License for the specific language governing permissions and
|
| 14 |
+
# limitations under the License.
|
| 15 |
+
|
| 16 |
+
|
| 17 |
+
from __future__ import annotations
|
| 18 |
+
|
| 19 |
+
import unittest
|
| 20 |
+
|
| 21 |
+
from transformers import BlenderbotConfig, BlenderbotTokenizer, is_tf_available
|
| 22 |
+
from transformers.testing_utils import require_tf, require_tokenizers, slow
|
| 23 |
+
from transformers.utils import cached_property
|
| 24 |
+
|
| 25 |
+
from ...test_configuration_common import ConfigTester
|
| 26 |
+
from ...test_modeling_tf_common import TFModelTesterMixin, ids_tensor
|
| 27 |
+
from ...test_pipeline_mixin import PipelineTesterMixin
|
| 28 |
+
|
| 29 |
+
|
| 30 |
+
if is_tf_available():
|
| 31 |
+
import tensorflow as tf
|
| 32 |
+
|
| 33 |
+
from transformers import TFAutoModelForSeq2SeqLM, TFBlenderbotForConditionalGeneration, TFBlenderbotModel
|
| 34 |
+
|
| 35 |
+
|
| 36 |
+
@require_tf
|
| 37 |
+
class TFBlenderbotModelTester:
|
| 38 |
+
config_cls = BlenderbotConfig
|
| 39 |
+
config_updates = {}
|
| 40 |
+
hidden_act = "gelu"
|
| 41 |
+
|
| 42 |
+
def __init__(
|
| 43 |
+
self,
|
| 44 |
+
parent,
|
| 45 |
+
batch_size=13,
|
| 46 |
+
seq_length=7,
|
| 47 |
+
is_training=True,
|
| 48 |
+
use_labels=False,
|
| 49 |
+
vocab_size=99,
|
| 50 |
+
hidden_size=32,
|
| 51 |
+
num_hidden_layers=2,
|
| 52 |
+
num_attention_heads=4,
|
| 53 |
+
intermediate_size=37,
|
| 54 |
+
hidden_dropout_prob=0.1,
|
| 55 |
+
attention_probs_dropout_prob=0.1,
|
| 56 |
+
max_position_embeddings=50,
|
| 57 |
+
eos_token_id=2,
|
| 58 |
+
pad_token_id=1,
|
| 59 |
+
bos_token_id=0,
|
| 60 |
+
):
|
| 61 |
+
self.parent = parent
|
| 62 |
+
self.batch_size = batch_size
|
| 63 |
+
self.seq_length = seq_length
|
| 64 |
+
self.is_training = is_training
|
| 65 |
+
self.use_labels = use_labels
|
| 66 |
+
self.vocab_size = vocab_size
|
| 67 |
+
self.hidden_size = hidden_size
|
| 68 |
+
self.num_hidden_layers = num_hidden_layers
|
| 69 |
+
self.num_attention_heads = num_attention_heads
|
| 70 |
+
self.intermediate_size = intermediate_size
|
| 71 |
+
|
| 72 |
+
self.hidden_dropout_prob = hidden_dropout_prob
|
| 73 |
+
self.attention_probs_dropout_prob = attention_probs_dropout_prob
|
| 74 |
+
self.max_position_embeddings = max_position_embeddings
|
| 75 |
+
self.eos_token_id = eos_token_id
|
| 76 |
+
self.pad_token_id = pad_token_id
|
| 77 |
+
self.bos_token_id = bos_token_id
|
| 78 |
+
|
| 79 |
+
def prepare_config_and_inputs_for_common(self):
|
| 80 |
+
input_ids = ids_tensor([self.batch_size, self.seq_length - 1], self.vocab_size)
|
| 81 |
+
eos_tensor = tf.expand_dims(tf.constant([self.eos_token_id] * self.batch_size), 1)
|
| 82 |
+
input_ids = tf.concat([input_ids, eos_tensor], axis=1)
|
| 83 |
+
|
| 84 |
+
decoder_input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
|
| 85 |
+
|
| 86 |
+
config = self.config_cls(
|
| 87 |
+
vocab_size=self.vocab_size,
|
| 88 |
+
d_model=self.hidden_size,
|
| 89 |
+
encoder_layers=self.num_hidden_layers,
|
| 90 |
+
decoder_layers=self.num_hidden_layers,
|
| 91 |
+
encoder_attention_heads=self.num_attention_heads,
|
| 92 |
+
decoder_attention_heads=self.num_attention_heads,
|
| 93 |
+
encoder_ffn_dim=self.intermediate_size,
|
| 94 |
+
decoder_ffn_dim=self.intermediate_size,
|
| 95 |
+
dropout=self.hidden_dropout_prob,
|
| 96 |
+
attention_dropout=self.attention_probs_dropout_prob,
|
| 97 |
+
max_position_embeddings=self.max_position_embeddings,
|
| 98 |
+
eos_token_ids=[2],
|
| 99 |
+
bos_token_id=self.bos_token_id,
|
| 100 |
+
pad_token_id=self.pad_token_id,
|
| 101 |
+
decoder_start_token_id=self.pad_token_id,
|
| 102 |
+
**self.config_updates,
|
| 103 |
+
)
|
| 104 |
+
inputs_dict = prepare_blenderbot_inputs_dict(config, input_ids, decoder_input_ids)
|
| 105 |
+
return config, inputs_dict
|
| 106 |
+
|
| 107 |
+
def check_decoder_model_past_large_inputs(self, config, inputs_dict):
|
| 108 |
+
model = TFBlenderbotModel(config=config).get_decoder()
|
| 109 |
+
input_ids = inputs_dict["input_ids"]
|
| 110 |
+
|
| 111 |
+
input_ids = input_ids[:1, :]
|
| 112 |
+
attention_mask = inputs_dict["attention_mask"][:1, :]
|
| 113 |
+
head_mask = inputs_dict["head_mask"]
|
| 114 |
+
self.batch_size = 1
|
| 115 |
+
|
| 116 |
+
# first forward pass
|
| 117 |
+
outputs = model(input_ids, attention_mask=attention_mask, head_mask=head_mask, use_cache=True)
|
| 118 |
+
|
| 119 |
+
output, past_key_values = outputs.to_tuple()
|
| 120 |
+
|
| 121 |
+
# create hypothetical next token and extent to next_input_ids
|
| 122 |
+
next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size)
|
| 123 |
+
next_attn_mask = tf.cast(ids_tensor((self.batch_size, 3), 2), tf.int8)
|
| 124 |
+
|
| 125 |
+
# append to next input_ids and
|
| 126 |
+
next_input_ids = tf.concat([input_ids, next_tokens], axis=-1)
|
| 127 |
+
next_attention_mask = tf.concat([attention_mask, next_attn_mask], axis=-1)
|
| 128 |
+
|
| 129 |
+
output_from_no_past = model(next_input_ids, attention_mask=next_attention_mask)[0]
|
| 130 |
+
output_from_past = model(next_tokens, attention_mask=next_attention_mask, past_key_values=past_key_values)[0]
|
| 131 |
+
|
| 132 |
+
self.parent.assertEqual(next_tokens.shape[1], output_from_past.shape[1])
|
| 133 |
+
|
| 134 |
+
# select random slice
|
| 135 |
+
random_slice_idx = int(ids_tensor((1,), output_from_past.shape[-1]))
|
| 136 |
+
output_from_no_past_slice = output_from_no_past[:, -3:, random_slice_idx]
|
| 137 |
+
output_from_past_slice = output_from_past[:, :, random_slice_idx]
|
| 138 |
+
|
| 139 |
+
# test that outputs are equal for slice
|
| 140 |
+
tf.debugging.assert_near(output_from_past_slice, output_from_no_past_slice, rtol=1e-3)
|
| 141 |
+
|
| 142 |
+
|
| 143 |
+
def prepare_blenderbot_inputs_dict(
|
| 144 |
+
config,
|
| 145 |
+
input_ids,
|
| 146 |
+
decoder_input_ids,
|
| 147 |
+
attention_mask=None,
|
| 148 |
+
decoder_attention_mask=None,
|
| 149 |
+
head_mask=None,
|
| 150 |
+
decoder_head_mask=None,
|
| 151 |
+
cross_attn_head_mask=None,
|
| 152 |
+
):
|
| 153 |
+
if attention_mask is None:
|
| 154 |
+
attention_mask = tf.cast(tf.math.not_equal(input_ids, config.pad_token_id), tf.int8)
|
| 155 |
+
if decoder_attention_mask is None:
|
| 156 |
+
decoder_attention_mask = tf.concat(
|
| 157 |
+
[
|
| 158 |
+
tf.ones(decoder_input_ids[:, :1].shape, dtype=tf.int8),
|
| 159 |
+
tf.cast(tf.math.not_equal(decoder_input_ids[:, 1:], config.pad_token_id), tf.int8),
|
| 160 |
+
],
|
| 161 |
+
axis=-1,
|
| 162 |
+
)
|
| 163 |
+
if head_mask is None:
|
| 164 |
+
head_mask = tf.ones((config.encoder_layers, config.encoder_attention_heads))
|
| 165 |
+
if decoder_head_mask is None:
|
| 166 |
+
decoder_head_mask = tf.ones((config.decoder_layers, config.decoder_attention_heads))
|
| 167 |
+
if cross_attn_head_mask is None:
|
| 168 |
+
cross_attn_head_mask = tf.ones((config.decoder_layers, config.decoder_attention_heads))
|
| 169 |
+
return {
|
| 170 |
+
"input_ids": input_ids,
|
| 171 |
+
"decoder_input_ids": decoder_input_ids,
|
| 172 |
+
"attention_mask": attention_mask,
|
| 173 |
+
"decoder_attention_mask": decoder_attention_mask,
|
| 174 |
+
"head_mask": head_mask,
|
| 175 |
+
"decoder_head_mask": decoder_head_mask,
|
| 176 |
+
"cross_attn_head_mask": cross_attn_head_mask,
|
| 177 |
+
}
|
| 178 |
+
|
| 179 |
+
|
| 180 |
+
@require_tf
|
| 181 |
+
class TFBlenderbotModelTest(TFModelTesterMixin, PipelineTesterMixin, unittest.TestCase):
|
| 182 |
+
all_model_classes = (TFBlenderbotForConditionalGeneration, TFBlenderbotModel) if is_tf_available() else ()
|
| 183 |
+
all_generative_model_classes = (TFBlenderbotForConditionalGeneration,) if is_tf_available() else ()
|
| 184 |
+
pipeline_model_mapping = (
|
| 185 |
+
{
|
| 186 |
+
"feature-extraction": TFBlenderbotModel,
|
| 187 |
+
"summarization": TFBlenderbotForConditionalGeneration,
|
| 188 |
+
"text2text-generation": TFBlenderbotForConditionalGeneration,
|
| 189 |
+
"translation": TFBlenderbotForConditionalGeneration,
|
| 190 |
+
}
|
| 191 |
+
if is_tf_available()
|
| 192 |
+
else {}
|
| 193 |
+
)
|
| 194 |
+
is_encoder_decoder = True
|
| 195 |
+
test_pruning = False
|
| 196 |
+
test_onnx = False
|
| 197 |
+
|
| 198 |
+
def setUp(self):
|
| 199 |
+
self.model_tester = TFBlenderbotModelTester(self)
|
| 200 |
+
self.config_tester = ConfigTester(self, config_class=BlenderbotConfig)
|
| 201 |
+
|
| 202 |
+
def test_config(self):
|
| 203 |
+
self.config_tester.run_common_tests()
|
| 204 |
+
|
| 205 |
+
def test_decoder_model_past_large_inputs(self):
|
| 206 |
+
config_and_inputs = self.model_tester.prepare_config_and_inputs_for_common()
|
| 207 |
+
self.model_tester.check_decoder_model_past_large_inputs(*config_and_inputs)
|
| 208 |
+
|
| 209 |
+
|
| 210 |
+
@require_tokenizers
|
| 211 |
+
@require_tf
|
| 212 |
+
class TFBlenderbot400MIntegrationTests(unittest.TestCase):
|
| 213 |
+
src_text = ["My friends are cool but they eat too many carbs."]
|
| 214 |
+
model_name = "facebook/blenderbot-400M-distill"
|
| 215 |
+
|
| 216 |
+
@cached_property
|
| 217 |
+
def tokenizer(self):
|
| 218 |
+
return BlenderbotTokenizer.from_pretrained(self.model_name)
|
| 219 |
+
|
| 220 |
+
@cached_property
|
| 221 |
+
def model(self):
|
| 222 |
+
model = TFAutoModelForSeq2SeqLM.from_pretrained(self.model_name)
|
| 223 |
+
return model
|
| 224 |
+
|
| 225 |
+
@slow
|
| 226 |
+
def test_generation_from_long_input(self):
|
| 227 |
+
model_inputs = self.tokenizer(self.src_text, return_tensors="tf")
|
| 228 |
+
generated_ids = self.model.generate(
|
| 229 |
+
model_inputs.input_ids,
|
| 230 |
+
)
|
| 231 |
+
generated_words = self.tokenizer.batch_decode(generated_ids.numpy(), skip_special_tokens=True)[0]
|
| 232 |
+
assert (
|
| 233 |
+
generated_words
|
| 234 |
+
== " That's unfortunate. Are they trying to lose weight or are they just trying to be healthier?"
|
| 235 |
+
)
|
tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/blenderbot/test_tokenization_blenderbot.py
ADDED
|
@@ -0,0 +1,75 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/usr/bin/env python3
|
| 2 |
+
# coding=utf-8
|
| 3 |
+
# Copyright 2020 The HuggingFace Team. All rights reserved.
|
| 4 |
+
#
|
| 5 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
| 6 |
+
# you may not use this file except in compliance with the License.
|
| 7 |
+
# You may obtain a copy of the License at
|
| 8 |
+
#
|
| 9 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
| 10 |
+
#
|
| 11 |
+
# Unless required by applicable law or agreed to in writing, software
|
| 12 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
| 13 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
| 14 |
+
# See the License for the specific language governing permissions and
|
| 15 |
+
# limitations under the License.
|
| 16 |
+
"""Tests for Blenderbot Tokenizers, including common tests for BlenderbotSmallTokenizer."""
|
| 17 |
+
|
| 18 |
+
import unittest
|
| 19 |
+
|
| 20 |
+
from transformers import BlenderbotTokenizer, BlenderbotTokenizerFast
|
| 21 |
+
from transformers.testing_utils import require_jinja
|
| 22 |
+
from transformers.utils import cached_property
|
| 23 |
+
|
| 24 |
+
|
| 25 |
+
class Blenderbot3BTokenizerTests(unittest.TestCase):
|
| 26 |
+
@cached_property
|
| 27 |
+
def tokenizer_3b(self):
|
| 28 |
+
return BlenderbotTokenizer.from_pretrained("facebook/blenderbot-3B")
|
| 29 |
+
|
| 30 |
+
@cached_property
|
| 31 |
+
def rust_tokenizer_3b(self):
|
| 32 |
+
return BlenderbotTokenizerFast.from_pretrained("facebook/blenderbot-3B")
|
| 33 |
+
|
| 34 |
+
def test_encode_decode_cycle(self):
|
| 35 |
+
tok = self.tokenizer_3b
|
| 36 |
+
src_text = " I am a small frog."
|
| 37 |
+
encoded = tok([src_text], padding=False, truncation=False)["input_ids"]
|
| 38 |
+
decoded = tok.batch_decode(encoded, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
|
| 39 |
+
assert src_text == decoded
|
| 40 |
+
|
| 41 |
+
def test_encode_decode_cycle_rust_tokenizer(self):
|
| 42 |
+
tok = self.rust_tokenizer_3b
|
| 43 |
+
src_text = " I am a small frog."
|
| 44 |
+
encoded = tok([src_text], padding=False, truncation=False)["input_ids"]
|
| 45 |
+
decoded = tok.batch_decode(encoded, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
|
| 46 |
+
assert src_text == decoded
|
| 47 |
+
|
| 48 |
+
def test_3B_tokenization_same_as_parlai(self):
|
| 49 |
+
assert self.tokenizer_3b.add_prefix_space
|
| 50 |
+
assert self.tokenizer_3b([" Sam", "Sam"]).input_ids == [[5502, 2], [5502, 2]]
|
| 51 |
+
|
| 52 |
+
def test_3B_tokenization_same_as_parlai_rust_tokenizer(self):
|
| 53 |
+
assert self.rust_tokenizer_3b.add_prefix_space
|
| 54 |
+
assert self.rust_tokenizer_3b([" Sam", "Sam"]).input_ids == [[5502, 2], [5502, 2]]
|
| 55 |
+
|
| 56 |
+
@require_jinja
|
| 57 |
+
def test_tokenization_for_chat(self):
|
| 58 |
+
tok = self.tokenizer_3b
|
| 59 |
+
test_chats = [
|
| 60 |
+
[{"role": "system", "content": "You are a helpful chatbot."}, {"role": "user", "content": "Hello!"}],
|
| 61 |
+
[
|
| 62 |
+
{"role": "system", "content": "You are a helpful chatbot."},
|
| 63 |
+
{"role": "user", "content": "Hello!"},
|
| 64 |
+
{"role": "assistant", "content": "Nice to meet you."},
|
| 65 |
+
],
|
| 66 |
+
[{"role": "assistant", "content": "Nice to meet you."}, {"role": "user", "content": "Hello!"}],
|
| 67 |
+
]
|
| 68 |
+
tokenized_chats = [tok.apply_chat_template(test_chat) for test_chat in test_chats]
|
| 69 |
+
expected_tokens = [
|
| 70 |
+
[553, 366, 265, 4792, 3879, 73, 311, 21, 228, 228, 6950, 8, 2],
|
| 71 |
+
[553, 366, 265, 4792, 3879, 73, 311, 21, 228, 228, 6950, 8, 228, 3490, 287, 2273, 304, 21, 2],
|
| 72 |
+
[3490, 287, 2273, 304, 21, 228, 228, 6950, 8, 2],
|
| 73 |
+
]
|
| 74 |
+
for tokenized_chat, expected_tokens in zip(tokenized_chats, expected_tokens):
|
| 75 |
+
self.assertListEqual(tokenized_chat, expected_tokens)
|
tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/blenderbot_small/__init__.py
ADDED
|
File without changes
|
tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/blenderbot_small/test_modeling_blenderbot_small.py
ADDED
|
@@ -0,0 +1,568 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# coding=utf-8
|
| 2 |
+
# Copyright 2021, The HuggingFace Inc. team. All rights reserved.
|
| 3 |
+
#
|
| 4 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
| 5 |
+
# you may not use this file except in compliance with the License.
|
| 6 |
+
# You may obtain a copy of the License at
|
| 7 |
+
#
|
| 8 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
| 9 |
+
#
|
| 10 |
+
# Unless required by applicable law or agreed to in writing, software
|
| 11 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
| 12 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
| 13 |
+
# See the License for the specific language governing permissions and
|
| 14 |
+
# limitations under the License.
|
| 15 |
+
"""Testing suite for the PyTorch BlenderbotSmall model."""
|
| 16 |
+
|
| 17 |
+
import tempfile
|
| 18 |
+
import unittest
|
| 19 |
+
|
| 20 |
+
from transformers import BlenderbotSmallConfig, is_torch_available
|
| 21 |
+
from transformers.testing_utils import (
|
| 22 |
+
require_torch,
|
| 23 |
+
require_torch_fp16,
|
| 24 |
+
slow,
|
| 25 |
+
torch_device,
|
| 26 |
+
)
|
| 27 |
+
from transformers.utils import cached_property
|
| 28 |
+
|
| 29 |
+
from ...generation.test_utils import GenerationTesterMixin
|
| 30 |
+
from ...test_configuration_common import ConfigTester
|
| 31 |
+
from ...test_modeling_common import ModelTesterMixin, ids_tensor
|
| 32 |
+
from ...test_pipeline_mixin import PipelineTesterMixin
|
| 33 |
+
|
| 34 |
+
|
| 35 |
+
if is_torch_available():
|
| 36 |
+
import torch
|
| 37 |
+
|
| 38 |
+
from transformers import BlenderbotSmallForConditionalGeneration, BlenderbotSmallModel, BlenderbotSmallTokenizer
|
| 39 |
+
from transformers.models.blenderbot_small.modeling_blenderbot_small import (
|
| 40 |
+
BlenderbotSmallDecoder,
|
| 41 |
+
BlenderbotSmallEncoder,
|
| 42 |
+
BlenderbotSmallForCausalLM,
|
| 43 |
+
)
|
| 44 |
+
|
| 45 |
+
|
| 46 |
+
def prepare_blenderbot_small_inputs_dict(
|
| 47 |
+
config,
|
| 48 |
+
input_ids,
|
| 49 |
+
decoder_input_ids,
|
| 50 |
+
attention_mask=None,
|
| 51 |
+
decoder_attention_mask=None,
|
| 52 |
+
head_mask=None,
|
| 53 |
+
decoder_head_mask=None,
|
| 54 |
+
cross_attn_head_mask=None,
|
| 55 |
+
):
|
| 56 |
+
if attention_mask is None:
|
| 57 |
+
attention_mask = input_ids.ne(config.pad_token_id)
|
| 58 |
+
if decoder_attention_mask is None:
|
| 59 |
+
decoder_attention_mask = decoder_input_ids.ne(config.pad_token_id)
|
| 60 |
+
if head_mask is None:
|
| 61 |
+
head_mask = torch.ones(config.encoder_layers, config.encoder_attention_heads, device=torch_device)
|
| 62 |
+
if decoder_head_mask is None:
|
| 63 |
+
decoder_head_mask = torch.ones(config.decoder_layers, config.decoder_attention_heads, device=torch_device)
|
| 64 |
+
if cross_attn_head_mask is None:
|
| 65 |
+
cross_attn_head_mask = torch.ones(config.decoder_layers, config.decoder_attention_heads, device=torch_device)
|
| 66 |
+
return {
|
| 67 |
+
"input_ids": input_ids,
|
| 68 |
+
"decoder_input_ids": decoder_input_ids,
|
| 69 |
+
"attention_mask": attention_mask,
|
| 70 |
+
"decoder_attention_mask": attention_mask,
|
| 71 |
+
"head_mask": head_mask,
|
| 72 |
+
"decoder_head_mask": decoder_head_mask,
|
| 73 |
+
"cross_attn_head_mask": cross_attn_head_mask,
|
| 74 |
+
}
|
| 75 |
+
|
| 76 |
+
|
| 77 |
+
class BlenderbotSmallModelTester:
|
| 78 |
+
def __init__(
|
| 79 |
+
self,
|
| 80 |
+
parent,
|
| 81 |
+
batch_size=13,
|
| 82 |
+
seq_length=7,
|
| 83 |
+
is_training=True,
|
| 84 |
+
use_labels=False,
|
| 85 |
+
vocab_size=99,
|
| 86 |
+
hidden_size=16,
|
| 87 |
+
num_hidden_layers=2,
|
| 88 |
+
num_attention_heads=4,
|
| 89 |
+
intermediate_size=4,
|
| 90 |
+
hidden_act="gelu",
|
| 91 |
+
hidden_dropout_prob=0.1,
|
| 92 |
+
attention_probs_dropout_prob=0.1,
|
| 93 |
+
max_position_embeddings=50,
|
| 94 |
+
eos_token_id=2,
|
| 95 |
+
pad_token_id=1,
|
| 96 |
+
bos_token_id=0,
|
| 97 |
+
):
|
| 98 |
+
self.parent = parent
|
| 99 |
+
self.batch_size = batch_size
|
| 100 |
+
self.seq_length = seq_length
|
| 101 |
+
self.is_training = is_training
|
| 102 |
+
self.use_labels = use_labels
|
| 103 |
+
self.vocab_size = vocab_size
|
| 104 |
+
self.hidden_size = hidden_size
|
| 105 |
+
self.num_hidden_layers = num_hidden_layers
|
| 106 |
+
self.num_attention_heads = num_attention_heads
|
| 107 |
+
self.intermediate_size = intermediate_size
|
| 108 |
+
self.hidden_act = hidden_act
|
| 109 |
+
self.hidden_dropout_prob = hidden_dropout_prob
|
| 110 |
+
self.attention_probs_dropout_prob = attention_probs_dropout_prob
|
| 111 |
+
self.max_position_embeddings = max_position_embeddings
|
| 112 |
+
self.eos_token_id = eos_token_id
|
| 113 |
+
self.pad_token_id = pad_token_id
|
| 114 |
+
self.bos_token_id = bos_token_id
|
| 115 |
+
|
| 116 |
+
def prepare_config_and_inputs(self):
|
| 117 |
+
input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size).clamp(
|
| 118 |
+
3,
|
| 119 |
+
)
|
| 120 |
+
input_ids[:, -1] = self.eos_token_id # Eos Token
|
| 121 |
+
|
| 122 |
+
decoder_input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
|
| 123 |
+
|
| 124 |
+
config = self.get_config()
|
| 125 |
+
inputs_dict = prepare_blenderbot_small_inputs_dict(config, input_ids, decoder_input_ids)
|
| 126 |
+
return config, inputs_dict
|
| 127 |
+
|
| 128 |
+
def get_config(self):
|
| 129 |
+
return BlenderbotSmallConfig(
|
| 130 |
+
vocab_size=self.vocab_size,
|
| 131 |
+
d_model=self.hidden_size,
|
| 132 |
+
encoder_layers=self.num_hidden_layers,
|
| 133 |
+
decoder_layers=self.num_hidden_layers,
|
| 134 |
+
encoder_attention_heads=self.num_attention_heads,
|
| 135 |
+
decoder_attention_heads=self.num_attention_heads,
|
| 136 |
+
encoder_ffn_dim=self.intermediate_size,
|
| 137 |
+
decoder_ffn_dim=self.intermediate_size,
|
| 138 |
+
dropout=self.hidden_dropout_prob,
|
| 139 |
+
attention_dropout=self.attention_probs_dropout_prob,
|
| 140 |
+
max_position_embeddings=self.max_position_embeddings,
|
| 141 |
+
eos_token_id=self.eos_token_id,
|
| 142 |
+
bos_token_id=self.bos_token_id,
|
| 143 |
+
pad_token_id=self.pad_token_id,
|
| 144 |
+
)
|
| 145 |
+
|
| 146 |
+
def prepare_config_and_inputs_for_common(self):
|
| 147 |
+
config, inputs_dict = self.prepare_config_and_inputs()
|
| 148 |
+
return config, inputs_dict
|
| 149 |
+
|
| 150 |
+
def create_and_check_decoder_model_past_large_inputs(self, config, inputs_dict):
|
| 151 |
+
model = BlenderbotSmallModel(config=config).get_decoder().to(torch_device).eval()
|
| 152 |
+
input_ids = inputs_dict["input_ids"]
|
| 153 |
+
attention_mask = inputs_dict["attention_mask"]
|
| 154 |
+
head_mask = inputs_dict["head_mask"]
|
| 155 |
+
|
| 156 |
+
# first forward pass
|
| 157 |
+
outputs = model(input_ids, attention_mask=attention_mask, head_mask=head_mask, use_cache=True)
|
| 158 |
+
|
| 159 |
+
output, past_key_values = outputs.to_tuple()
|
| 160 |
+
|
| 161 |
+
# create hypothetical multiple next token and extent to next_input_ids
|
| 162 |
+
next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size)
|
| 163 |
+
next_attn_mask = ids_tensor((self.batch_size, 3), 2)
|
| 164 |
+
|
| 165 |
+
# append to next input_ids and
|
| 166 |
+
next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
|
| 167 |
+
next_attention_mask = torch.cat([attention_mask, next_attn_mask], dim=-1)
|
| 168 |
+
|
| 169 |
+
output_from_no_past = model(next_input_ids, attention_mask=next_attention_mask)["last_hidden_state"]
|
| 170 |
+
output_from_past = model(next_tokens, attention_mask=next_attention_mask, past_key_values=past_key_values)[
|
| 171 |
+
"last_hidden_state"
|
| 172 |
+
]
|
| 173 |
+
|
| 174 |
+
# select random slice
|
| 175 |
+
random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()
|
| 176 |
+
output_from_no_past_slice = output_from_no_past[:, -3:, random_slice_idx].detach()
|
| 177 |
+
output_from_past_slice = output_from_past[:, :, random_slice_idx].detach()
|
| 178 |
+
|
| 179 |
+
self.parent.assertTrue(output_from_past_slice.shape[1] == next_tokens.shape[1])
|
| 180 |
+
|
| 181 |
+
# test that outputs are equal for slice
|
| 182 |
+
self.parent.assertTrue(torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3))
|
| 183 |
+
|
| 184 |
+
def check_encoder_decoder_model_standalone(self, config, inputs_dict):
|
| 185 |
+
model = BlenderbotSmallModel(config=config).to(torch_device).eval()
|
| 186 |
+
outputs = model(**inputs_dict)
|
| 187 |
+
|
| 188 |
+
encoder_last_hidden_state = outputs.encoder_last_hidden_state
|
| 189 |
+
last_hidden_state = outputs.last_hidden_state
|
| 190 |
+
|
| 191 |
+
with tempfile.TemporaryDirectory() as tmpdirname:
|
| 192 |
+
encoder = model.get_encoder()
|
| 193 |
+
encoder.save_pretrained(tmpdirname)
|
| 194 |
+
encoder = BlenderbotSmallEncoder.from_pretrained(tmpdirname).to(torch_device)
|
| 195 |
+
|
| 196 |
+
encoder_last_hidden_state_2 = encoder(inputs_dict["input_ids"], attention_mask=inputs_dict["attention_mask"])[
|
| 197 |
+
0
|
| 198 |
+
]
|
| 199 |
+
|
| 200 |
+
self.parent.assertTrue((encoder_last_hidden_state_2 - encoder_last_hidden_state).abs().max().item() < 1e-3)
|
| 201 |
+
|
| 202 |
+
with tempfile.TemporaryDirectory() as tmpdirname:
|
| 203 |
+
decoder = model.get_decoder()
|
| 204 |
+
decoder.save_pretrained(tmpdirname)
|
| 205 |
+
decoder = BlenderbotSmallDecoder.from_pretrained(tmpdirname).to(torch_device)
|
| 206 |
+
|
| 207 |
+
last_hidden_state_2 = decoder(
|
| 208 |
+
input_ids=inputs_dict["decoder_input_ids"],
|
| 209 |
+
attention_mask=inputs_dict["decoder_attention_mask"],
|
| 210 |
+
encoder_hidden_states=encoder_last_hidden_state,
|
| 211 |
+
encoder_attention_mask=inputs_dict["attention_mask"],
|
| 212 |
+
)[0]
|
| 213 |
+
|
| 214 |
+
self.parent.assertTrue((last_hidden_state_2 - last_hidden_state).abs().max().item() < 1e-3)
|
| 215 |
+
|
| 216 |
+
|
| 217 |
+
@require_torch
|
| 218 |
+
class BlenderbotSmallModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixin, unittest.TestCase):
|
| 219 |
+
all_model_classes = (BlenderbotSmallModel, BlenderbotSmallForConditionalGeneration) if is_torch_available() else ()
|
| 220 |
+
all_generative_model_classes = (BlenderbotSmallForConditionalGeneration,) if is_torch_available() else ()
|
| 221 |
+
pipeline_model_mapping = (
|
| 222 |
+
{
|
| 223 |
+
"feature-extraction": BlenderbotSmallModel,
|
| 224 |
+
"summarization": BlenderbotSmallForConditionalGeneration,
|
| 225 |
+
"text-generation": BlenderbotSmallForCausalLM,
|
| 226 |
+
"text2text-generation": BlenderbotSmallForConditionalGeneration,
|
| 227 |
+
"translation": BlenderbotSmallForConditionalGeneration,
|
| 228 |
+
}
|
| 229 |
+
if is_torch_available()
|
| 230 |
+
else {}
|
| 231 |
+
)
|
| 232 |
+
is_encoder_decoder = True
|
| 233 |
+
fx_compatible = True
|
| 234 |
+
test_pruning = False
|
| 235 |
+
test_missing_keys = False
|
| 236 |
+
|
| 237 |
+
# TODO: Fix the failed tests when this model gets more usage
|
| 238 |
+
def is_pipeline_test_to_skip(
|
| 239 |
+
self,
|
| 240 |
+
pipeline_test_case_name,
|
| 241 |
+
config_class,
|
| 242 |
+
model_architecture,
|
| 243 |
+
tokenizer_name,
|
| 244 |
+
image_processor_name,
|
| 245 |
+
feature_extractor_name,
|
| 246 |
+
processor_name,
|
| 247 |
+
):
|
| 248 |
+
return pipeline_test_case_name == "TextGenerationPipelineTests"
|
| 249 |
+
|
| 250 |
+
def setUp(self):
|
| 251 |
+
self.model_tester = BlenderbotSmallModelTester(self)
|
| 252 |
+
self.config_tester = ConfigTester(self, config_class=BlenderbotSmallConfig)
|
| 253 |
+
|
| 254 |
+
def test_config(self):
|
| 255 |
+
self.config_tester.run_common_tests()
|
| 256 |
+
|
| 257 |
+
def test_save_load_strict(self):
|
| 258 |
+
config, inputs_dict = self.model_tester.prepare_config_and_inputs()
|
| 259 |
+
for model_class in self.all_model_classes:
|
| 260 |
+
model = model_class(config)
|
| 261 |
+
|
| 262 |
+
with tempfile.TemporaryDirectory() as tmpdirname:
|
| 263 |
+
model.save_pretrained(tmpdirname)
|
| 264 |
+
model2, info = model_class.from_pretrained(tmpdirname, output_loading_info=True)
|
| 265 |
+
self.assertEqual(info["missing_keys"], [])
|
| 266 |
+
|
| 267 |
+
def test_decoder_model_past_with_large_inputs(self):
|
| 268 |
+
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
| 269 |
+
self.model_tester.create_and_check_decoder_model_past_large_inputs(*config_and_inputs)
|
| 270 |
+
|
| 271 |
+
def test_encoder_decoder_model_standalone(self):
|
| 272 |
+
config_and_inputs = self.model_tester.prepare_config_and_inputs_for_common()
|
| 273 |
+
self.model_tester.check_encoder_decoder_model_standalone(*config_and_inputs)
|
| 274 |
+
|
| 275 |
+
@require_torch_fp16
|
| 276 |
+
def test_generate_fp16(self):
|
| 277 |
+
config, input_dict = self.model_tester.prepare_config_and_inputs()
|
| 278 |
+
input_ids = input_dict["input_ids"]
|
| 279 |
+
attention_mask = input_ids.ne(1).to(torch_device)
|
| 280 |
+
model = BlenderbotSmallForConditionalGeneration(config).eval().to(torch_device)
|
| 281 |
+
model.half()
|
| 282 |
+
model.generate(input_ids, attention_mask=attention_mask)
|
| 283 |
+
model.generate(num_beams=4, do_sample=True, early_stopping=False, num_return_sequences=3)
|
| 284 |
+
|
| 285 |
+
|
| 286 |
+
def assert_tensors_close(a, b, atol=1e-12, prefix=""):
|
| 287 |
+
"""If tensors have different shapes, different values or a and b are not both tensors, raise a nice Assertion error."""
|
| 288 |
+
if a is None and b is None:
|
| 289 |
+
return True
|
| 290 |
+
try:
|
| 291 |
+
if torch.allclose(a, b, atol=atol):
|
| 292 |
+
return True
|
| 293 |
+
raise
|
| 294 |
+
except Exception:
|
| 295 |
+
pct_different = (torch.gt((a - b).abs(), atol)).float().mean().item()
|
| 296 |
+
if a.numel() > 100:
|
| 297 |
+
msg = f"tensor values are {pct_different:.1%} percent different."
|
| 298 |
+
else:
|
| 299 |
+
msg = f"{a} != {b}"
|
| 300 |
+
if prefix:
|
| 301 |
+
msg = prefix + ": " + msg
|
| 302 |
+
raise AssertionError(msg)
|
| 303 |
+
|
| 304 |
+
|
| 305 |
+
@require_torch
|
| 306 |
+
class Blenderbot90MIntegrationTests(unittest.TestCase):
|
| 307 |
+
ckpt = "facebook/blenderbot-90M"
|
| 308 |
+
|
| 309 |
+
@cached_property
|
| 310 |
+
def model(self):
|
| 311 |
+
model = BlenderbotSmallForConditionalGeneration.from_pretrained(self.ckpt).to(torch_device)
|
| 312 |
+
if torch_device == "cuda":
|
| 313 |
+
model = model.half()
|
| 314 |
+
return model
|
| 315 |
+
|
| 316 |
+
@cached_property
|
| 317 |
+
def tokenizer(self):
|
| 318 |
+
return BlenderbotSmallTokenizer.from_pretrained(self.ckpt)
|
| 319 |
+
|
| 320 |
+
@slow
|
| 321 |
+
def test_90_generation_from_long_input(self):
|
| 322 |
+
src_text = [
|
| 323 |
+
"Social anxiety\nWow, I am never shy. Do you have anxiety?\nYes. I end up sweating and blushing and feel"
|
| 324 |
+
" like i'm going to throw up.\nand why is that?"
|
| 325 |
+
]
|
| 326 |
+
|
| 327 |
+
model_inputs = self.tokenizer(src_text, return_tensors="pt").to(torch_device)
|
| 328 |
+
|
| 329 |
+
assert isinstance(self.tokenizer, BlenderbotSmallTokenizer)
|
| 330 |
+
generated_ids = self.model.generate(**model_inputs)[0]
|
| 331 |
+
reply = self.tokenizer.decode(generated_ids, skip_special_tokens=True, clean_up_tokenization_spaces=True)
|
| 332 |
+
|
| 333 |
+
assert reply in (
|
| 334 |
+
"i don't know. i just feel like i'm going to throw up. it's not fun.",
|
| 335 |
+
"i'm not sure. i just feel like i've been feeling like i have to be in a certain place",
|
| 336 |
+
)
|
| 337 |
+
|
| 338 |
+
@slow
|
| 339 |
+
def test_90_generation_from_short_input(self):
|
| 340 |
+
model_inputs = self.tokenizer(["sam"], return_tensors="pt").to(torch_device)
|
| 341 |
+
|
| 342 |
+
generated_utterances = self.model.generate(**model_inputs)
|
| 343 |
+
|
| 344 |
+
clean_txt = self.tokenizer.decode(
|
| 345 |
+
generated_utterances[0], skip_special_tokens=True, clean_up_tokenization_spaces=True
|
| 346 |
+
)
|
| 347 |
+
assert clean_txt in (
|
| 348 |
+
"have you ever been to a sam club? it's a great club in the south.",
|
| 349 |
+
"have you ever heard of sam harris? he's an american singer, songwriter, and actor.",
|
| 350 |
+
)
|
| 351 |
+
|
| 352 |
+
|
| 353 |
+
class BlenderbotSmallStandaloneDecoderModelTester:
|
| 354 |
+
def __init__(
|
| 355 |
+
self,
|
| 356 |
+
parent,
|
| 357 |
+
vocab_size=99,
|
| 358 |
+
batch_size=13,
|
| 359 |
+
d_model=16,
|
| 360 |
+
decoder_seq_length=7,
|
| 361 |
+
is_training=True,
|
| 362 |
+
is_decoder=True,
|
| 363 |
+
use_attention_mask=True,
|
| 364 |
+
use_cache=False,
|
| 365 |
+
use_labels=True,
|
| 366 |
+
decoder_start_token_id=2,
|
| 367 |
+
decoder_ffn_dim=32,
|
| 368 |
+
decoder_layers=2,
|
| 369 |
+
encoder_attention_heads=4,
|
| 370 |
+
decoder_attention_heads=4,
|
| 371 |
+
max_position_embeddings=30,
|
| 372 |
+
is_encoder_decoder=False,
|
| 373 |
+
pad_token_id=0,
|
| 374 |
+
bos_token_id=1,
|
| 375 |
+
eos_token_id=2,
|
| 376 |
+
scope=None,
|
| 377 |
+
):
|
| 378 |
+
self.parent = parent
|
| 379 |
+
self.batch_size = batch_size
|
| 380 |
+
self.decoder_seq_length = decoder_seq_length
|
| 381 |
+
# For common tests
|
| 382 |
+
self.seq_length = self.decoder_seq_length
|
| 383 |
+
self.is_training = is_training
|
| 384 |
+
self.use_attention_mask = use_attention_mask
|
| 385 |
+
self.use_labels = use_labels
|
| 386 |
+
|
| 387 |
+
self.vocab_size = vocab_size
|
| 388 |
+
self.d_model = d_model
|
| 389 |
+
self.hidden_size = d_model
|
| 390 |
+
self.num_hidden_layers = decoder_layers
|
| 391 |
+
self.decoder_layers = decoder_layers
|
| 392 |
+
self.decoder_ffn_dim = decoder_ffn_dim
|
| 393 |
+
self.encoder_attention_heads = encoder_attention_heads
|
| 394 |
+
self.decoder_attention_heads = decoder_attention_heads
|
| 395 |
+
self.num_attention_heads = decoder_attention_heads
|
| 396 |
+
self.eos_token_id = eos_token_id
|
| 397 |
+
self.bos_token_id = bos_token_id
|
| 398 |
+
self.pad_token_id = pad_token_id
|
| 399 |
+
self.decoder_start_token_id = decoder_start_token_id
|
| 400 |
+
self.use_cache = use_cache
|
| 401 |
+
self.max_position_embeddings = max_position_embeddings
|
| 402 |
+
self.is_encoder_decoder = is_encoder_decoder
|
| 403 |
+
|
| 404 |
+
self.scope = None
|
| 405 |
+
self.decoder_key_length = decoder_seq_length
|
| 406 |
+
self.base_model_out_len = 2
|
| 407 |
+
self.decoder_attention_idx = 1
|
| 408 |
+
|
| 409 |
+
def prepare_config_and_inputs(self):
|
| 410 |
+
input_ids = ids_tensor([self.batch_size, self.decoder_seq_length], self.vocab_size)
|
| 411 |
+
|
| 412 |
+
attention_mask = None
|
| 413 |
+
if self.use_attention_mask:
|
| 414 |
+
attention_mask = ids_tensor([self.batch_size, self.decoder_seq_length], vocab_size=2)
|
| 415 |
+
|
| 416 |
+
lm_labels = None
|
| 417 |
+
if self.use_labels:
|
| 418 |
+
lm_labels = ids_tensor([self.batch_size, self.decoder_seq_length], self.vocab_size)
|
| 419 |
+
|
| 420 |
+
config = BlenderbotSmallConfig(
|
| 421 |
+
vocab_size=self.vocab_size,
|
| 422 |
+
d_model=self.d_model,
|
| 423 |
+
decoder_layers=self.decoder_layers,
|
| 424 |
+
decoder_ffn_dim=self.decoder_ffn_dim,
|
| 425 |
+
encoder_attention_heads=self.encoder_attention_heads,
|
| 426 |
+
decoder_attention_heads=self.decoder_attention_heads,
|
| 427 |
+
eos_token_id=self.eos_token_id,
|
| 428 |
+
bos_token_id=self.bos_token_id,
|
| 429 |
+
use_cache=self.use_cache,
|
| 430 |
+
pad_token_id=self.pad_token_id,
|
| 431 |
+
decoder_start_token_id=self.decoder_start_token_id,
|
| 432 |
+
max_position_embeddings=self.max_position_embeddings,
|
| 433 |
+
is_encoder_decoder=self.is_encoder_decoder,
|
| 434 |
+
)
|
| 435 |
+
|
| 436 |
+
return (
|
| 437 |
+
config,
|
| 438 |
+
input_ids,
|
| 439 |
+
attention_mask,
|
| 440 |
+
lm_labels,
|
| 441 |
+
)
|
| 442 |
+
|
| 443 |
+
def create_and_check_decoder_model_past(
|
| 444 |
+
self,
|
| 445 |
+
config,
|
| 446 |
+
input_ids,
|
| 447 |
+
attention_mask,
|
| 448 |
+
lm_labels,
|
| 449 |
+
):
|
| 450 |
+
config.use_cache = True
|
| 451 |
+
model = BlenderbotSmallDecoder(config=config).to(torch_device).eval()
|
| 452 |
+
# first forward pass
|
| 453 |
+
outputs = model(input_ids, use_cache=True)
|
| 454 |
+
outputs_use_cache_conf = model(input_ids)
|
| 455 |
+
outputs_no_past = model(input_ids, use_cache=False)
|
| 456 |
+
|
| 457 |
+
self.parent.assertTrue(len(outputs) == len(outputs_use_cache_conf))
|
| 458 |
+
self.parent.assertTrue(len(outputs) == len(outputs_no_past) + 1)
|
| 459 |
+
|
| 460 |
+
past_key_values = outputs["past_key_values"]
|
| 461 |
+
|
| 462 |
+
# create hypothetical next token and extent to next_input_ids
|
| 463 |
+
next_tokens = ids_tensor((self.batch_size, 1), config.vocab_size)
|
| 464 |
+
|
| 465 |
+
# append to next input_ids and
|
| 466 |
+
next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
|
| 467 |
+
|
| 468 |
+
output_from_no_past = model(next_input_ids)["last_hidden_state"]
|
| 469 |
+
output_from_past = model(next_tokens, past_key_values=past_key_values)["last_hidden_state"]
|
| 470 |
+
|
| 471 |
+
# select random slice
|
| 472 |
+
random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()
|
| 473 |
+
output_from_no_past_slice = output_from_no_past[:, next_input_ids.shape[-1] - 1, random_slice_idx].detach()
|
| 474 |
+
output_from_past_slice = output_from_past[:, 0, random_slice_idx].detach()
|
| 475 |
+
|
| 476 |
+
# test that outputs are equal for slice
|
| 477 |
+
assert torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3)
|
| 478 |
+
|
| 479 |
+
def create_and_check_decoder_model_attention_mask_past(
|
| 480 |
+
self,
|
| 481 |
+
config,
|
| 482 |
+
input_ids,
|
| 483 |
+
attention_mask,
|
| 484 |
+
lm_labels,
|
| 485 |
+
):
|
| 486 |
+
model = BlenderbotSmallDecoder(config=config).to(torch_device).eval()
|
| 487 |
+
|
| 488 |
+
# create attention mask
|
| 489 |
+
attn_mask = torch.ones(input_ids.shape, dtype=torch.long, device=torch_device)
|
| 490 |
+
|
| 491 |
+
half_seq_length = input_ids.shape[-1] // 2
|
| 492 |
+
attn_mask[:, half_seq_length:] = 0
|
| 493 |
+
|
| 494 |
+
# first forward pass
|
| 495 |
+
past_key_values = model(input_ids, attention_mask=attn_mask, use_cache=True)["past_key_values"]
|
| 496 |
+
|
| 497 |
+
# create hypothetical next token and extent to next_input_ids
|
| 498 |
+
next_tokens = ids_tensor((self.batch_size, 1), config.vocab_size)
|
| 499 |
+
|
| 500 |
+
# change a random masked slice from input_ids
|
| 501 |
+
random_seq_idx_to_change = ids_tensor((1,), half_seq_length).item() + 1
|
| 502 |
+
random_other_next_tokens = ids_tensor((self.batch_size, 1), config.vocab_size).squeeze(-1)
|
| 503 |
+
input_ids[:, -random_seq_idx_to_change] = random_other_next_tokens
|
| 504 |
+
|
| 505 |
+
# append to next input_ids and attn_mask
|
| 506 |
+
next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
|
| 507 |
+
attn_mask = torch.cat(
|
| 508 |
+
[attn_mask, torch.ones((attn_mask.shape[0], 1), dtype=torch.long, device=torch_device)],
|
| 509 |
+
dim=1,
|
| 510 |
+
)
|
| 511 |
+
|
| 512 |
+
# get two different outputs
|
| 513 |
+
output_from_no_past = model(next_input_ids, attention_mask=attn_mask)["last_hidden_state"]
|
| 514 |
+
output_from_past = model(next_tokens, past_key_values=past_key_values, attention_mask=attn_mask)[
|
| 515 |
+
"last_hidden_state"
|
| 516 |
+
]
|
| 517 |
+
|
| 518 |
+
# select random slice
|
| 519 |
+
random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()
|
| 520 |
+
output_from_no_past_slice = output_from_no_past[:, next_input_ids.shape[-1] - 1, random_slice_idx].detach()
|
| 521 |
+
output_from_past_slice = output_from_past[:, 0, random_slice_idx].detach()
|
| 522 |
+
|
| 523 |
+
# test that outputs are equal for slice
|
| 524 |
+
assert torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3)
|
| 525 |
+
|
| 526 |
+
def prepare_config_and_inputs_for_common(self):
|
| 527 |
+
config_and_inputs = self.prepare_config_and_inputs()
|
| 528 |
+
(
|
| 529 |
+
config,
|
| 530 |
+
input_ids,
|
| 531 |
+
attention_mask,
|
| 532 |
+
lm_labels,
|
| 533 |
+
) = config_and_inputs
|
| 534 |
+
|
| 535 |
+
inputs_dict = {
|
| 536 |
+
"input_ids": input_ids,
|
| 537 |
+
"attention_mask": attention_mask,
|
| 538 |
+
}
|
| 539 |
+
return config, inputs_dict
|
| 540 |
+
|
| 541 |
+
|
| 542 |
+
@require_torch
|
| 543 |
+
class BlenderbotSmallStandaloneDecoderModelTest(ModelTesterMixin, GenerationTesterMixin, unittest.TestCase):
|
| 544 |
+
all_model_classes = (BlenderbotSmallDecoder, BlenderbotSmallForCausalLM) if is_torch_available() else ()
|
| 545 |
+
all_generative_model_classes = (BlenderbotSmallForCausalLM,) if is_torch_available() else ()
|
| 546 |
+
test_pruning = False
|
| 547 |
+
is_encoder_decoder = False
|
| 548 |
+
|
| 549 |
+
def setUp(
|
| 550 |
+
self,
|
| 551 |
+
):
|
| 552 |
+
self.model_tester = BlenderbotSmallStandaloneDecoderModelTester(self, is_training=False)
|
| 553 |
+
self.config_tester = ConfigTester(self, config_class=BlenderbotSmallConfig)
|
| 554 |
+
|
| 555 |
+
def test_config(self):
|
| 556 |
+
self.config_tester.run_common_tests()
|
| 557 |
+
|
| 558 |
+
def test_decoder_model_past(self):
|
| 559 |
+
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
| 560 |
+
self.model_tester.create_and_check_decoder_model_past(*config_and_inputs)
|
| 561 |
+
|
| 562 |
+
def test_decoder_model_attn_mask_past(self):
|
| 563 |
+
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
| 564 |
+
self.model_tester.create_and_check_decoder_model_attention_mask_past(*config_and_inputs)
|
| 565 |
+
|
| 566 |
+
@unittest.skip(reason="decoder cannot keep gradients")
|
| 567 |
+
def test_retain_grad_hidden_states_attentions(self):
|
| 568 |
+
return
|
tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/blenderbot_small/test_modeling_flax_blenderbot_small.py
ADDED
|
@@ -0,0 +1,411 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Copyright 2021 The HuggingFace Team. All rights reserved.
|
| 2 |
+
#
|
| 3 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
| 4 |
+
# you may not use this file except in compliance with the License.
|
| 5 |
+
# You may obtain a copy of the License at
|
| 6 |
+
#
|
| 7 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
| 8 |
+
#
|
| 9 |
+
# Unless required by applicable law or agreed to in writing, software
|
| 10 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
| 11 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
| 12 |
+
# See the License for the specific language governing permissions and
|
| 13 |
+
# limitations under the License.
|
| 14 |
+
|
| 15 |
+
import unittest
|
| 16 |
+
|
| 17 |
+
import numpy as np
|
| 18 |
+
import timeout_decorator # noqa
|
| 19 |
+
|
| 20 |
+
from transformers import BlenderbotSmallConfig, is_flax_available
|
| 21 |
+
from transformers.testing_utils import require_flax, slow
|
| 22 |
+
|
| 23 |
+
from ...generation.test_flax_utils import FlaxGenerationTesterMixin
|
| 24 |
+
from ...test_modeling_flax_common import FlaxModelTesterMixin, ids_tensor
|
| 25 |
+
|
| 26 |
+
|
| 27 |
+
if is_flax_available():
|
| 28 |
+
import os
|
| 29 |
+
|
| 30 |
+
# The slow tests are often failing with OOM error on GPU
|
| 31 |
+
# This makes JAX allocate exactly what is needed on demand, and deallocate memory that is no longer needed
|
| 32 |
+
# but will be slower as stated here https://jax.readthedocs.io/en/latest/gpu_memory_allocation.html
|
| 33 |
+
os.environ["XLA_PYTHON_CLIENT_ALLOCATOR"] = "platform"
|
| 34 |
+
|
| 35 |
+
import jax
|
| 36 |
+
import jax.numpy as jnp
|
| 37 |
+
|
| 38 |
+
from transformers.models.blenderbot_small.modeling_flax_blenderbot_small import (
|
| 39 |
+
FlaxBlenderbotSmallForConditionalGeneration,
|
| 40 |
+
FlaxBlenderbotSmallModel,
|
| 41 |
+
shift_tokens_right,
|
| 42 |
+
)
|
| 43 |
+
|
| 44 |
+
|
| 45 |
+
def prepare_blenderbot_inputs_dict(
|
| 46 |
+
config,
|
| 47 |
+
input_ids,
|
| 48 |
+
decoder_input_ids=None,
|
| 49 |
+
attention_mask=None,
|
| 50 |
+
decoder_attention_mask=None,
|
| 51 |
+
head_mask=None,
|
| 52 |
+
decoder_head_mask=None,
|
| 53 |
+
cross_attn_head_mask=None,
|
| 54 |
+
):
|
| 55 |
+
if attention_mask is None:
|
| 56 |
+
attention_mask = np.where(input_ids != config.pad_token_id, 1, 0)
|
| 57 |
+
if decoder_attention_mask is None:
|
| 58 |
+
decoder_attention_mask = np.where(decoder_input_ids != config.pad_token_id, 1, 0)
|
| 59 |
+
if head_mask is None:
|
| 60 |
+
head_mask = np.ones((config.encoder_layers, config.encoder_attention_heads))
|
| 61 |
+
if decoder_head_mask is None:
|
| 62 |
+
decoder_head_mask = np.ones((config.decoder_layers, config.decoder_attention_heads))
|
| 63 |
+
if cross_attn_head_mask is None:
|
| 64 |
+
cross_attn_head_mask = np.ones((config.decoder_layers, config.decoder_attention_heads))
|
| 65 |
+
return {
|
| 66 |
+
"input_ids": input_ids,
|
| 67 |
+
"decoder_input_ids": decoder_input_ids,
|
| 68 |
+
"attention_mask": attention_mask,
|
| 69 |
+
"decoder_attention_mask": attention_mask,
|
| 70 |
+
}
|
| 71 |
+
|
| 72 |
+
|
| 73 |
+
class FlaxBlenderbotSmallModelTester:
|
| 74 |
+
def __init__(
|
| 75 |
+
self,
|
| 76 |
+
parent,
|
| 77 |
+
batch_size=13,
|
| 78 |
+
seq_length=7,
|
| 79 |
+
is_training=True,
|
| 80 |
+
use_labels=False,
|
| 81 |
+
vocab_size=99,
|
| 82 |
+
hidden_size=16,
|
| 83 |
+
num_hidden_layers=2,
|
| 84 |
+
num_attention_heads=4,
|
| 85 |
+
intermediate_size=4,
|
| 86 |
+
hidden_act="gelu",
|
| 87 |
+
hidden_dropout_prob=0.1,
|
| 88 |
+
attention_probs_dropout_prob=0.1,
|
| 89 |
+
max_position_embeddings=50,
|
| 90 |
+
eos_token_id=2,
|
| 91 |
+
pad_token_id=1,
|
| 92 |
+
bos_token_id=0,
|
| 93 |
+
initializer_range=0.02,
|
| 94 |
+
):
|
| 95 |
+
self.parent = parent
|
| 96 |
+
self.batch_size = batch_size
|
| 97 |
+
self.seq_length = seq_length
|
| 98 |
+
self.is_training = is_training
|
| 99 |
+
self.use_labels = use_labels
|
| 100 |
+
self.vocab_size = vocab_size
|
| 101 |
+
self.hidden_size = hidden_size
|
| 102 |
+
self.num_hidden_layers = num_hidden_layers
|
| 103 |
+
self.num_attention_heads = num_attention_heads
|
| 104 |
+
self.intermediate_size = intermediate_size
|
| 105 |
+
self.hidden_act = hidden_act
|
| 106 |
+
self.hidden_dropout_prob = hidden_dropout_prob
|
| 107 |
+
self.attention_probs_dropout_prob = attention_probs_dropout_prob
|
| 108 |
+
self.max_position_embeddings = max_position_embeddings
|
| 109 |
+
self.eos_token_id = eos_token_id
|
| 110 |
+
self.pad_token_id = pad_token_id
|
| 111 |
+
self.bos_token_id = bos_token_id
|
| 112 |
+
self.initializer_range = initializer_range
|
| 113 |
+
|
| 114 |
+
def prepare_config_and_inputs(self):
|
| 115 |
+
input_ids = np.clip(ids_tensor([self.batch_size, self.seq_length - 1], self.vocab_size), 3, self.vocab_size)
|
| 116 |
+
input_ids = np.concatenate((input_ids, 2 * np.ones((self.batch_size, 1), dtype=np.int64)), -1)
|
| 117 |
+
|
| 118 |
+
decoder_input_ids = shift_tokens_right(input_ids, 1, 2)
|
| 119 |
+
|
| 120 |
+
config = BlenderbotSmallConfig(
|
| 121 |
+
vocab_size=self.vocab_size,
|
| 122 |
+
d_model=self.hidden_size,
|
| 123 |
+
encoder_layers=self.num_hidden_layers,
|
| 124 |
+
decoder_layers=self.num_hidden_layers,
|
| 125 |
+
encoder_attention_heads=self.num_attention_heads,
|
| 126 |
+
decoder_attention_heads=self.num_attention_heads,
|
| 127 |
+
encoder_ffn_dim=self.intermediate_size,
|
| 128 |
+
decoder_ffn_dim=self.intermediate_size,
|
| 129 |
+
dropout=self.hidden_dropout_prob,
|
| 130 |
+
attention_dropout=self.attention_probs_dropout_prob,
|
| 131 |
+
max_position_embeddings=self.max_position_embeddings,
|
| 132 |
+
eos_token_id=self.eos_token_id,
|
| 133 |
+
bos_token_id=self.bos_token_id,
|
| 134 |
+
pad_token_id=self.pad_token_id,
|
| 135 |
+
initializer_range=self.initializer_range,
|
| 136 |
+
use_cache=False,
|
| 137 |
+
)
|
| 138 |
+
inputs_dict = prepare_blenderbot_inputs_dict(config, input_ids, decoder_input_ids)
|
| 139 |
+
return config, inputs_dict
|
| 140 |
+
|
| 141 |
+
def prepare_config_and_inputs_for_common(self):
|
| 142 |
+
config, inputs_dict = self.prepare_config_and_inputs()
|
| 143 |
+
return config, inputs_dict
|
| 144 |
+
|
| 145 |
+
def check_use_cache_forward(self, model_class_name, config, inputs_dict):
|
| 146 |
+
max_decoder_length = 20
|
| 147 |
+
model = model_class_name(config)
|
| 148 |
+
|
| 149 |
+
encoder_outputs = model.encode(inputs_dict["input_ids"])
|
| 150 |
+
|
| 151 |
+
decoder_input_ids, decoder_attention_mask = (
|
| 152 |
+
inputs_dict["decoder_input_ids"],
|
| 153 |
+
inputs_dict["decoder_attention_mask"],
|
| 154 |
+
)
|
| 155 |
+
|
| 156 |
+
past_key_values = model.init_cache(decoder_input_ids.shape[0], max_decoder_length, encoder_outputs)
|
| 157 |
+
decoder_attention_mask = jnp.ones((decoder_input_ids.shape[0], max_decoder_length), dtype="i4")
|
| 158 |
+
|
| 159 |
+
decoder_position_ids = jnp.broadcast_to(
|
| 160 |
+
jnp.arange(decoder_input_ids.shape[-1] - 1)[None, :],
|
| 161 |
+
(decoder_input_ids.shape[0], decoder_input_ids.shape[-1] - 1),
|
| 162 |
+
)
|
| 163 |
+
outputs_cache = model.decode(
|
| 164 |
+
decoder_input_ids[:, :-1],
|
| 165 |
+
encoder_outputs,
|
| 166 |
+
decoder_attention_mask=decoder_attention_mask,
|
| 167 |
+
past_key_values=past_key_values,
|
| 168 |
+
decoder_position_ids=decoder_position_ids,
|
| 169 |
+
)
|
| 170 |
+
|
| 171 |
+
decoder_position_ids = jnp.array(decoder_input_ids.shape[0] * [[decoder_input_ids.shape[-1] - 1]], dtype="i4")
|
| 172 |
+
outputs_cache_next = model.decode(
|
| 173 |
+
decoder_input_ids[:, -1:],
|
| 174 |
+
encoder_outputs,
|
| 175 |
+
decoder_attention_mask=decoder_attention_mask,
|
| 176 |
+
past_key_values=outputs_cache.past_key_values,
|
| 177 |
+
decoder_position_ids=decoder_position_ids,
|
| 178 |
+
)
|
| 179 |
+
|
| 180 |
+
outputs = model.decode(decoder_input_ids, encoder_outputs)
|
| 181 |
+
|
| 182 |
+
diff = np.max(np.abs((outputs_cache_next[0][:, -1, :5] - outputs[0][:, -1, :5])))
|
| 183 |
+
self.parent.assertTrue(diff < 1e-3, msg=f"Max diff is {diff}")
|
| 184 |
+
|
| 185 |
+
def check_use_cache_forward_with_attn_mask(self, model_class_name, config, inputs_dict):
|
| 186 |
+
max_decoder_length = 20
|
| 187 |
+
model = model_class_name(config)
|
| 188 |
+
|
| 189 |
+
encoder_outputs = model.encode(inputs_dict["input_ids"])
|
| 190 |
+
|
| 191 |
+
decoder_input_ids, decoder_attention_mask = (
|
| 192 |
+
inputs_dict["decoder_input_ids"],
|
| 193 |
+
inputs_dict["decoder_attention_mask"],
|
| 194 |
+
)
|
| 195 |
+
|
| 196 |
+
decoder_attention_mask_cache = jnp.concatenate(
|
| 197 |
+
[
|
| 198 |
+
decoder_attention_mask,
|
| 199 |
+
jnp.zeros((decoder_attention_mask.shape[0], max_decoder_length - decoder_attention_mask.shape[1])),
|
| 200 |
+
],
|
| 201 |
+
axis=-1,
|
| 202 |
+
)
|
| 203 |
+
|
| 204 |
+
past_key_values = model.init_cache(decoder_input_ids.shape[0], max_decoder_length, encoder_outputs)
|
| 205 |
+
decoder_position_ids = jnp.broadcast_to(
|
| 206 |
+
jnp.arange(decoder_input_ids.shape[-1] - 1)[None, :],
|
| 207 |
+
(decoder_input_ids.shape[0], decoder_input_ids.shape[-1] - 1),
|
| 208 |
+
)
|
| 209 |
+
|
| 210 |
+
outputs_cache = model.decode(
|
| 211 |
+
decoder_input_ids[:, :-1],
|
| 212 |
+
encoder_outputs,
|
| 213 |
+
decoder_attention_mask=decoder_attention_mask_cache,
|
| 214 |
+
past_key_values=past_key_values,
|
| 215 |
+
decoder_position_ids=decoder_position_ids,
|
| 216 |
+
)
|
| 217 |
+
decoder_position_ids = jnp.array(decoder_input_ids.shape[0] * [[decoder_input_ids.shape[-1] - 1]], dtype="i4")
|
| 218 |
+
outputs_cache_next = model.decode(
|
| 219 |
+
decoder_input_ids[:, -1:],
|
| 220 |
+
encoder_outputs,
|
| 221 |
+
past_key_values=outputs_cache.past_key_values,
|
| 222 |
+
decoder_attention_mask=decoder_attention_mask_cache,
|
| 223 |
+
decoder_position_ids=decoder_position_ids,
|
| 224 |
+
)
|
| 225 |
+
|
| 226 |
+
outputs = model.decode(decoder_input_ids, encoder_outputs, decoder_attention_mask=decoder_attention_mask)
|
| 227 |
+
|
| 228 |
+
diff = np.max(np.abs((outputs_cache_next[0][:, -1, :5] - outputs[0][:, -1, :5])))
|
| 229 |
+
self.parent.assertTrue(diff < 1e-3, msg=f"Max diff is {diff}")
|
| 230 |
+
|
| 231 |
+
|
| 232 |
+
@require_flax
|
| 233 |
+
class BlenderbotHeadTests(unittest.TestCase):
|
| 234 |
+
vocab_size = 99
|
| 235 |
+
|
| 236 |
+
def _get_config_and_data(self):
|
| 237 |
+
input_ids = np.array(
|
| 238 |
+
[
|
| 239 |
+
[71, 82, 18, 33, 46, 91, 2],
|
| 240 |
+
[68, 34, 26, 58, 30, 82, 2],
|
| 241 |
+
[5, 97, 17, 39, 94, 40, 2],
|
| 242 |
+
[76, 83, 94, 25, 70, 78, 2],
|
| 243 |
+
[87, 59, 41, 35, 48, 66, 2],
|
| 244 |
+
[55, 13, 16, 58, 5, 2, 1], # note padding
|
| 245 |
+
[64, 27, 31, 51, 12, 75, 2],
|
| 246 |
+
[52, 64, 86, 17, 83, 39, 2],
|
| 247 |
+
[48, 61, 9, 24, 71, 82, 2],
|
| 248 |
+
[26, 1, 60, 48, 22, 13, 2],
|
| 249 |
+
[21, 5, 62, 28, 14, 76, 2],
|
| 250 |
+
[45, 98, 37, 86, 59, 48, 2],
|
| 251 |
+
[70, 70, 50, 9, 28, 0, 2],
|
| 252 |
+
],
|
| 253 |
+
dtype=np.int64,
|
| 254 |
+
)
|
| 255 |
+
|
| 256 |
+
batch_size = input_ids.shape[0]
|
| 257 |
+
config = BlenderbotSmallConfig(
|
| 258 |
+
vocab_size=self.vocab_size,
|
| 259 |
+
d_model=24,
|
| 260 |
+
encoder_layers=2,
|
| 261 |
+
decoder_layers=2,
|
| 262 |
+
encoder_attention_heads=2,
|
| 263 |
+
decoder_attention_heads=2,
|
| 264 |
+
encoder_ffn_dim=32,
|
| 265 |
+
decoder_ffn_dim=32,
|
| 266 |
+
max_position_embeddings=48,
|
| 267 |
+
eos_token_id=2,
|
| 268 |
+
pad_token_id=1,
|
| 269 |
+
bos_token_id=0,
|
| 270 |
+
)
|
| 271 |
+
return config, input_ids, batch_size
|
| 272 |
+
|
| 273 |
+
# @timeout_decorator.timeout(1) # not working with the decorator so far
|
| 274 |
+
def test_lm_forward(self):
|
| 275 |
+
config, input_ids, batch_size = self._get_config_and_data()
|
| 276 |
+
lm_model = FlaxBlenderbotSmallForConditionalGeneration(config)
|
| 277 |
+
outputs = lm_model(input_ids=input_ids)
|
| 278 |
+
expected_shape = (batch_size, input_ids.shape[1], config.vocab_size)
|
| 279 |
+
self.assertEqual(outputs["logits"].shape, expected_shape)
|
| 280 |
+
|
| 281 |
+
def test_lm_uneven_forward(self):
|
| 282 |
+
config = BlenderbotSmallConfig(
|
| 283 |
+
vocab_size=self.vocab_size,
|
| 284 |
+
d_model=14,
|
| 285 |
+
encoder_layers=2,
|
| 286 |
+
decoder_layers=2,
|
| 287 |
+
encoder_attention_heads=2,
|
| 288 |
+
decoder_attention_heads=2,
|
| 289 |
+
encoder_ffn_dim=8,
|
| 290 |
+
decoder_ffn_dim=8,
|
| 291 |
+
max_position_embeddings=48,
|
| 292 |
+
)
|
| 293 |
+
lm_model = FlaxBlenderbotSmallForConditionalGeneration(config)
|
| 294 |
+
context = np.array([[71, 82, 18, 33, 46, 91, 2], [68, 34, 26, 58, 30, 2, 1]], dtype=np.int64)
|
| 295 |
+
summary = np.array([[82, 71, 82, 18, 2], [58, 68, 2, 1, 1]], dtype=np.int64)
|
| 296 |
+
outputs = lm_model(input_ids=context, decoder_input_ids=summary)
|
| 297 |
+
expected_shape = (*summary.shape, config.vocab_size)
|
| 298 |
+
self.assertEqual(outputs["logits"].shape, expected_shape)
|
| 299 |
+
|
| 300 |
+
def test_shift_tokens_right(self):
|
| 301 |
+
input_ids = np.array([[71, 82, 18, 33, 2, 1, 1], [68, 34, 26, 58, 30, 82, 2]], dtype=np.int64)
|
| 302 |
+
shifted = shift_tokens_right(input_ids, 1, 2)
|
| 303 |
+
n_pad_before = np.equal(input_ids, 1).astype(np.float32).sum()
|
| 304 |
+
n_pad_after = np.equal(shifted, 1).astype(np.float32).sum()
|
| 305 |
+
self.assertEqual(shifted.shape, input_ids.shape)
|
| 306 |
+
self.assertEqual(n_pad_after, n_pad_before - 1)
|
| 307 |
+
self.assertTrue(np.equal(shifted[:, 0], 2).all())
|
| 308 |
+
|
| 309 |
+
|
| 310 |
+
@require_flax
|
| 311 |
+
class FlaxBlenderbotSmallModelTest(FlaxModelTesterMixin, unittest.TestCase, FlaxGenerationTesterMixin):
|
| 312 |
+
is_encoder_decoder = True
|
| 313 |
+
all_model_classes = (
|
| 314 |
+
(
|
| 315 |
+
FlaxBlenderbotSmallModel,
|
| 316 |
+
FlaxBlenderbotSmallForConditionalGeneration,
|
| 317 |
+
)
|
| 318 |
+
if is_flax_available()
|
| 319 |
+
else ()
|
| 320 |
+
)
|
| 321 |
+
all_generative_model_classes = (FlaxBlenderbotSmallForConditionalGeneration,) if is_flax_available() else ()
|
| 322 |
+
|
| 323 |
+
def is_pipeline_test_to_skip(
|
| 324 |
+
self,
|
| 325 |
+
pipeline_test_case_name,
|
| 326 |
+
config_class,
|
| 327 |
+
model_architecture,
|
| 328 |
+
tokenizer_name,
|
| 329 |
+
image_processor_name,
|
| 330 |
+
feature_extractor_name,
|
| 331 |
+
processor_name,
|
| 332 |
+
):
|
| 333 |
+
return pipeline_test_case_name == "TextGenerationPipelineTests"
|
| 334 |
+
|
| 335 |
+
def setUp(self):
|
| 336 |
+
self.model_tester = FlaxBlenderbotSmallModelTester(self)
|
| 337 |
+
|
| 338 |
+
def test_use_cache_forward(self):
|
| 339 |
+
config, inputs_dict = self.model_tester.prepare_config_and_inputs()
|
| 340 |
+
for model_class in self.all_model_classes:
|
| 341 |
+
self.model_tester.check_use_cache_forward(model_class, config, inputs_dict)
|
| 342 |
+
|
| 343 |
+
def test_use_cache_forward_with_attn_mask(self):
|
| 344 |
+
config, inputs_dict = self.model_tester.prepare_config_and_inputs()
|
| 345 |
+
for model_class in self.all_model_classes:
|
| 346 |
+
self.model_tester.check_use_cache_forward_with_attn_mask(model_class, config, inputs_dict)
|
| 347 |
+
|
| 348 |
+
def test_encode(self):
|
| 349 |
+
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
| 350 |
+
|
| 351 |
+
for model_class in self.all_model_classes:
|
| 352 |
+
with self.subTest(model_class.__name__):
|
| 353 |
+
prepared_inputs_dict = self._prepare_for_class(inputs_dict, model_class)
|
| 354 |
+
model = model_class(config)
|
| 355 |
+
|
| 356 |
+
@jax.jit
|
| 357 |
+
def encode_jitted(input_ids, attention_mask=None, **kwargs):
|
| 358 |
+
return model.encode(input_ids=input_ids, attention_mask=attention_mask)
|
| 359 |
+
|
| 360 |
+
with self.subTest("JIT Enabled"):
|
| 361 |
+
jitted_outputs = encode_jitted(**prepared_inputs_dict).to_tuple()
|
| 362 |
+
|
| 363 |
+
with self.subTest("JIT Disabled"):
|
| 364 |
+
with jax.disable_jit():
|
| 365 |
+
outputs = encode_jitted(**prepared_inputs_dict).to_tuple()
|
| 366 |
+
|
| 367 |
+
self.assertEqual(len(outputs), len(jitted_outputs))
|
| 368 |
+
for jitted_output, output in zip(jitted_outputs, outputs):
|
| 369 |
+
self.assertEqual(jitted_output.shape, output.shape)
|
| 370 |
+
|
| 371 |
+
def test_decode(self):
|
| 372 |
+
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
| 373 |
+
|
| 374 |
+
for model_class in self.all_model_classes:
|
| 375 |
+
with self.subTest(model_class.__name__):
|
| 376 |
+
model = model_class(config)
|
| 377 |
+
encoder_outputs = model.encode(inputs_dict["input_ids"], inputs_dict["attention_mask"])
|
| 378 |
+
|
| 379 |
+
prepared_inputs_dict = {
|
| 380 |
+
"decoder_input_ids": inputs_dict["decoder_input_ids"],
|
| 381 |
+
"decoder_attention_mask": inputs_dict["decoder_attention_mask"],
|
| 382 |
+
"encoder_outputs": encoder_outputs,
|
| 383 |
+
}
|
| 384 |
+
|
| 385 |
+
@jax.jit
|
| 386 |
+
def decode_jitted(decoder_input_ids, decoder_attention_mask, encoder_outputs):
|
| 387 |
+
return model.decode(
|
| 388 |
+
decoder_input_ids=decoder_input_ids,
|
| 389 |
+
decoder_attention_mask=decoder_attention_mask,
|
| 390 |
+
encoder_outputs=encoder_outputs,
|
| 391 |
+
)
|
| 392 |
+
|
| 393 |
+
with self.subTest("JIT Enabled"):
|
| 394 |
+
jitted_outputs = decode_jitted(**prepared_inputs_dict).to_tuple()
|
| 395 |
+
|
| 396 |
+
with self.subTest("JIT Disabled"):
|
| 397 |
+
with jax.disable_jit():
|
| 398 |
+
outputs = decode_jitted(**prepared_inputs_dict).to_tuple()
|
| 399 |
+
|
| 400 |
+
self.assertEqual(len(outputs), len(jitted_outputs))
|
| 401 |
+
for jitted_output, output in zip(jitted_outputs, outputs):
|
| 402 |
+
self.assertEqual(jitted_output.shape, output.shape)
|
| 403 |
+
|
| 404 |
+
@slow
|
| 405 |
+
def test_model_from_pretrained(self):
|
| 406 |
+
for model_class_name in self.all_model_classes:
|
| 407 |
+
model = model_class_name.from_pretrained("facebook/blenderbot_small-90M")
|
| 408 |
+
# FlaxBlenderbotForSequenceClassification expects eos token in input_ids
|
| 409 |
+
input_ids = np.ones((1, 1)) * model.config.eos_token_id
|
| 410 |
+
outputs = model(input_ids)
|
| 411 |
+
self.assertIsNotNone(outputs)
|
tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/blenderbot_small/test_modeling_tf_blenderbot_small.py
ADDED
|
@@ -0,0 +1,257 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# coding=utf-8
|
| 2 |
+
# Copyright 2021 The HuggingFace Inc. team. All rights reserved.
|
| 3 |
+
#
|
| 4 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
| 5 |
+
# you may not use this file except in compliance with the License.
|
| 6 |
+
# You may obtain a copy of the License at
|
| 7 |
+
#
|
| 8 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
| 9 |
+
#
|
| 10 |
+
# Unless required by applicable law or agreed to in writing, software
|
| 11 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
| 12 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
| 13 |
+
# See the License for the specific language governing permissions and
|
| 14 |
+
# limitations under the License.
|
| 15 |
+
|
| 16 |
+
|
| 17 |
+
from __future__ import annotations
|
| 18 |
+
|
| 19 |
+
import unittest
|
| 20 |
+
|
| 21 |
+
from transformers import BlenderbotSmallConfig, BlenderbotSmallTokenizer, is_tf_available
|
| 22 |
+
from transformers.testing_utils import require_tf, require_tokenizers, slow
|
| 23 |
+
from transformers.utils import cached_property
|
| 24 |
+
|
| 25 |
+
from ...test_configuration_common import ConfigTester
|
| 26 |
+
from ...test_modeling_tf_common import TFModelTesterMixin, ids_tensor
|
| 27 |
+
from ...test_pipeline_mixin import PipelineTesterMixin
|
| 28 |
+
|
| 29 |
+
|
| 30 |
+
if is_tf_available():
|
| 31 |
+
import tensorflow as tf
|
| 32 |
+
|
| 33 |
+
from transformers import TFAutoModelForSeq2SeqLM, TFBlenderbotSmallForConditionalGeneration, TFBlenderbotSmallModel
|
| 34 |
+
|
| 35 |
+
|
| 36 |
+
@require_tf
|
| 37 |
+
class TFBlenderbotSmallModelTester:
|
| 38 |
+
config_cls = BlenderbotSmallConfig
|
| 39 |
+
config_updates = {}
|
| 40 |
+
hidden_act = "gelu"
|
| 41 |
+
|
| 42 |
+
def __init__(
|
| 43 |
+
self,
|
| 44 |
+
parent,
|
| 45 |
+
batch_size=13,
|
| 46 |
+
seq_length=7,
|
| 47 |
+
is_training=True,
|
| 48 |
+
use_labels=False,
|
| 49 |
+
vocab_size=99,
|
| 50 |
+
hidden_size=32,
|
| 51 |
+
num_hidden_layers=2,
|
| 52 |
+
num_attention_heads=4,
|
| 53 |
+
intermediate_size=37,
|
| 54 |
+
hidden_dropout_prob=0.1,
|
| 55 |
+
attention_probs_dropout_prob=0.1,
|
| 56 |
+
max_position_embeddings=50,
|
| 57 |
+
eos_token_id=2,
|
| 58 |
+
pad_token_id=1,
|
| 59 |
+
bos_token_id=0,
|
| 60 |
+
):
|
| 61 |
+
self.parent = parent
|
| 62 |
+
self.batch_size = batch_size
|
| 63 |
+
self.seq_length = seq_length
|
| 64 |
+
self.is_training = is_training
|
| 65 |
+
self.use_labels = use_labels
|
| 66 |
+
self.vocab_size = vocab_size
|
| 67 |
+
self.hidden_size = hidden_size
|
| 68 |
+
self.num_hidden_layers = num_hidden_layers
|
| 69 |
+
self.num_attention_heads = num_attention_heads
|
| 70 |
+
self.intermediate_size = intermediate_size
|
| 71 |
+
|
| 72 |
+
self.hidden_dropout_prob = hidden_dropout_prob
|
| 73 |
+
self.attention_probs_dropout_prob = attention_probs_dropout_prob
|
| 74 |
+
self.max_position_embeddings = max_position_embeddings
|
| 75 |
+
self.eos_token_id = eos_token_id
|
| 76 |
+
self.pad_token_id = pad_token_id
|
| 77 |
+
self.bos_token_id = bos_token_id
|
| 78 |
+
|
| 79 |
+
def prepare_config_and_inputs_for_common(self):
|
| 80 |
+
input_ids = ids_tensor([self.batch_size, self.seq_length - 1], self.vocab_size)
|
| 81 |
+
eos_tensor = tf.expand_dims(tf.constant([self.eos_token_id] * self.batch_size), 1)
|
| 82 |
+
input_ids = tf.concat([input_ids, eos_tensor], axis=1)
|
| 83 |
+
|
| 84 |
+
decoder_input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
|
| 85 |
+
|
| 86 |
+
config = self.config_cls(
|
| 87 |
+
vocab_size=self.vocab_size,
|
| 88 |
+
d_model=self.hidden_size,
|
| 89 |
+
encoder_layers=self.num_hidden_layers,
|
| 90 |
+
decoder_layers=self.num_hidden_layers,
|
| 91 |
+
encoder_attention_heads=self.num_attention_heads,
|
| 92 |
+
decoder_attention_heads=self.num_attention_heads,
|
| 93 |
+
encoder_ffn_dim=self.intermediate_size,
|
| 94 |
+
decoder_ffn_dim=self.intermediate_size,
|
| 95 |
+
dropout=self.hidden_dropout_prob,
|
| 96 |
+
attention_dropout=self.attention_probs_dropout_prob,
|
| 97 |
+
max_position_embeddings=self.max_position_embeddings,
|
| 98 |
+
eos_token_ids=[2],
|
| 99 |
+
bos_token_id=self.bos_token_id,
|
| 100 |
+
pad_token_id=self.pad_token_id,
|
| 101 |
+
decoder_start_token_id=self.pad_token_id,
|
| 102 |
+
**self.config_updates,
|
| 103 |
+
)
|
| 104 |
+
inputs_dict = prepare_blenderbot_small_inputs_dict(config, input_ids, decoder_input_ids)
|
| 105 |
+
return config, inputs_dict
|
| 106 |
+
|
| 107 |
+
def check_decoder_model_past_large_inputs(self, config, inputs_dict):
|
| 108 |
+
model = TFBlenderbotSmallModel(config=config).get_decoder()
|
| 109 |
+
input_ids = inputs_dict["input_ids"]
|
| 110 |
+
|
| 111 |
+
input_ids = input_ids[:1, :]
|
| 112 |
+
attention_mask = inputs_dict["attention_mask"][:1, :]
|
| 113 |
+
head_mask = inputs_dict["head_mask"]
|
| 114 |
+
self.batch_size = 1
|
| 115 |
+
|
| 116 |
+
# first forward pass
|
| 117 |
+
outputs = model(input_ids, attention_mask=attention_mask, head_mask=head_mask, use_cache=True)
|
| 118 |
+
|
| 119 |
+
output, past_key_values = outputs.to_tuple()
|
| 120 |
+
|
| 121 |
+
# create hypothetical next token and extent to next_input_ids
|
| 122 |
+
next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size)
|
| 123 |
+
next_attn_mask = tf.cast(ids_tensor((self.batch_size, 3), 2), tf.int8)
|
| 124 |
+
|
| 125 |
+
# append to next input_ids and
|
| 126 |
+
next_input_ids = tf.concat([input_ids, next_tokens], axis=-1)
|
| 127 |
+
next_attention_mask = tf.concat([attention_mask, next_attn_mask], axis=-1)
|
| 128 |
+
|
| 129 |
+
output_from_no_past = model(next_input_ids, attention_mask=next_attention_mask)[0]
|
| 130 |
+
output_from_past = model(next_tokens, attention_mask=next_attention_mask, past_key_values=past_key_values)[0]
|
| 131 |
+
|
| 132 |
+
self.parent.assertEqual(next_tokens.shape[1], output_from_past.shape[1])
|
| 133 |
+
|
| 134 |
+
# select random slice
|
| 135 |
+
random_slice_idx = int(ids_tensor((1,), output_from_past.shape[-1]))
|
| 136 |
+
output_from_no_past_slice = output_from_no_past[:, -3:, random_slice_idx]
|
| 137 |
+
output_from_past_slice = output_from_past[:, :, random_slice_idx]
|
| 138 |
+
|
| 139 |
+
# test that outputs are equal for slice
|
| 140 |
+
tf.debugging.assert_near(output_from_past_slice, output_from_no_past_slice, rtol=1e-3)
|
| 141 |
+
|
| 142 |
+
|
| 143 |
+
def prepare_blenderbot_small_inputs_dict(
|
| 144 |
+
config,
|
| 145 |
+
input_ids,
|
| 146 |
+
decoder_input_ids,
|
| 147 |
+
attention_mask=None,
|
| 148 |
+
decoder_attention_mask=None,
|
| 149 |
+
head_mask=None,
|
| 150 |
+
decoder_head_mask=None,
|
| 151 |
+
cross_attn_head_mask=None,
|
| 152 |
+
):
|
| 153 |
+
if attention_mask is None:
|
| 154 |
+
attention_mask = tf.cast(tf.math.not_equal(input_ids, config.pad_token_id), tf.int8)
|
| 155 |
+
if decoder_attention_mask is None:
|
| 156 |
+
decoder_attention_mask = tf.concat(
|
| 157 |
+
[
|
| 158 |
+
tf.ones(decoder_input_ids[:, :1].shape, dtype=tf.int8),
|
| 159 |
+
tf.cast(tf.math.not_equal(decoder_input_ids[:, 1:], config.pad_token_id), tf.int8),
|
| 160 |
+
],
|
| 161 |
+
axis=-1,
|
| 162 |
+
)
|
| 163 |
+
if head_mask is None:
|
| 164 |
+
head_mask = tf.ones((config.encoder_layers, config.encoder_attention_heads))
|
| 165 |
+
if decoder_head_mask is None:
|
| 166 |
+
decoder_head_mask = tf.ones((config.decoder_layers, config.decoder_attention_heads))
|
| 167 |
+
if cross_attn_head_mask is None:
|
| 168 |
+
cross_attn_head_mask = tf.ones((config.decoder_layers, config.decoder_attention_heads))
|
| 169 |
+
return {
|
| 170 |
+
"input_ids": input_ids,
|
| 171 |
+
"decoder_input_ids": decoder_input_ids,
|
| 172 |
+
"attention_mask": attention_mask,
|
| 173 |
+
"decoder_attention_mask": decoder_attention_mask,
|
| 174 |
+
"head_mask": head_mask,
|
| 175 |
+
"decoder_head_mask": decoder_head_mask,
|
| 176 |
+
"cross_attn_head_mask": cross_attn_head_mask,
|
| 177 |
+
}
|
| 178 |
+
|
| 179 |
+
|
| 180 |
+
@require_tf
|
| 181 |
+
class TFBlenderbotSmallModelTest(TFModelTesterMixin, PipelineTesterMixin, unittest.TestCase):
|
| 182 |
+
all_model_classes = (
|
| 183 |
+
(TFBlenderbotSmallForConditionalGeneration, TFBlenderbotSmallModel) if is_tf_available() else ()
|
| 184 |
+
)
|
| 185 |
+
all_generative_model_classes = (TFBlenderbotSmallForConditionalGeneration,) if is_tf_available() else ()
|
| 186 |
+
pipeline_model_mapping = (
|
| 187 |
+
{
|
| 188 |
+
"feature-extraction": TFBlenderbotSmallModel,
|
| 189 |
+
"summarization": TFBlenderbotSmallForConditionalGeneration,
|
| 190 |
+
"text2text-generation": TFBlenderbotSmallForConditionalGeneration,
|
| 191 |
+
"translation": TFBlenderbotSmallForConditionalGeneration,
|
| 192 |
+
}
|
| 193 |
+
if is_tf_available()
|
| 194 |
+
else {}
|
| 195 |
+
)
|
| 196 |
+
is_encoder_decoder = True
|
| 197 |
+
test_pruning = False
|
| 198 |
+
test_onnx = False
|
| 199 |
+
|
| 200 |
+
def is_pipeline_test_to_skip(
|
| 201 |
+
self,
|
| 202 |
+
pipeline_test_case_name,
|
| 203 |
+
config_class,
|
| 204 |
+
model_architecture,
|
| 205 |
+
tokenizer_name,
|
| 206 |
+
image_processor_name,
|
| 207 |
+
feature_extractor_name,
|
| 208 |
+
processor_name,
|
| 209 |
+
):
|
| 210 |
+
return pipeline_test_case_name == "TextGenerationPipelineTests"
|
| 211 |
+
|
| 212 |
+
def setUp(self):
|
| 213 |
+
self.model_tester = TFBlenderbotSmallModelTester(self)
|
| 214 |
+
self.config_tester = ConfigTester(self, config_class=BlenderbotSmallConfig)
|
| 215 |
+
|
| 216 |
+
def test_config(self):
|
| 217 |
+
self.config_tester.run_common_tests()
|
| 218 |
+
|
| 219 |
+
def test_decoder_model_past_large_inputs(self):
|
| 220 |
+
config_and_inputs = self.model_tester.prepare_config_and_inputs_for_common()
|
| 221 |
+
self.model_tester.check_decoder_model_past_large_inputs(*config_and_inputs)
|
| 222 |
+
|
| 223 |
+
|
| 224 |
+
@require_tokenizers
|
| 225 |
+
@require_tf
|
| 226 |
+
class TFBlenderbot90MIntegrationTests(unittest.TestCase):
|
| 227 |
+
src_text = [
|
| 228 |
+
"Social anxiety\nWow, I am never shy. Do you have anxiety?\nYes. I end up sweating and blushing and feel like "
|
| 229 |
+
" i'm going to throw up.\nand why is that?"
|
| 230 |
+
]
|
| 231 |
+
model_name = "facebook/blenderbot_small-90M"
|
| 232 |
+
|
| 233 |
+
@cached_property
|
| 234 |
+
def tokenizer(self):
|
| 235 |
+
# use "old" tokenizer here because of bug when downloading new tokenizer
|
| 236 |
+
return BlenderbotSmallTokenizer.from_pretrained("facebook/blenderbot-90M")
|
| 237 |
+
|
| 238 |
+
@cached_property
|
| 239 |
+
def model(self):
|
| 240 |
+
model = TFAutoModelForSeq2SeqLM.from_pretrained(self.model_name)
|
| 241 |
+
return model
|
| 242 |
+
|
| 243 |
+
@slow
|
| 244 |
+
def test_90_generation_from_long_input(self):
|
| 245 |
+
model_inputs = self.tokenizer(self.src_text, return_tensors="tf")
|
| 246 |
+
generated_ids = self.model.generate(
|
| 247 |
+
model_inputs.input_ids,
|
| 248 |
+
attention_mask=model_inputs.attention_mask,
|
| 249 |
+
num_beams=2,
|
| 250 |
+
use_cache=True,
|
| 251 |
+
)
|
| 252 |
+
generated_words = self.tokenizer.batch_decode(generated_ids.numpy(), skip_special_tokens=True)[0]
|
| 253 |
+
assert generated_words in (
|
| 254 |
+
"i don't know. i just feel like i'm going to throw up. it's not fun.",
|
| 255 |
+
"i'm not sure. i just feel like i've been feeling like i have to be in a certain place",
|
| 256 |
+
"i'm not sure. i just feel like i've been in a bad situation.",
|
| 257 |
+
)
|
tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/blenderbot_small/test_tokenization_blenderbot_small.py
ADDED
|
@@ -0,0 +1,88 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/usr/bin/env python3
|
| 2 |
+
# coding=utf-8
|
| 3 |
+
# Copyright 2020 The HuggingFace Team. All rights reserved.
|
| 4 |
+
#
|
| 5 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
| 6 |
+
# you may not use this file except in compliance with the License.
|
| 7 |
+
# You may obtain a copy of the License at
|
| 8 |
+
#
|
| 9 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
| 10 |
+
#
|
| 11 |
+
# Unless required by applicable law or agreed to in writing, software
|
| 12 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
| 13 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
| 14 |
+
# See the License for the specific language governing permissions and
|
| 15 |
+
# limitations under the License.
|
| 16 |
+
"""Tests for the Blenderbot small tokenizer."""
|
| 17 |
+
|
| 18 |
+
import json
|
| 19 |
+
import os
|
| 20 |
+
import unittest
|
| 21 |
+
|
| 22 |
+
from transformers.models.blenderbot_small.tokenization_blenderbot_small import (
|
| 23 |
+
VOCAB_FILES_NAMES,
|
| 24 |
+
BlenderbotSmallTokenizer,
|
| 25 |
+
)
|
| 26 |
+
|
| 27 |
+
from ...test_tokenization_common import TokenizerTesterMixin
|
| 28 |
+
|
| 29 |
+
|
| 30 |
+
class BlenderbotSmallTokenizerTest(TokenizerTesterMixin, unittest.TestCase):
|
| 31 |
+
from_pretrained_id = "facebook/blenderbot_small-90M"
|
| 32 |
+
tokenizer_class = BlenderbotSmallTokenizer
|
| 33 |
+
test_rust_tokenizer = False
|
| 34 |
+
|
| 35 |
+
def setUp(self):
|
| 36 |
+
super().setUp()
|
| 37 |
+
|
| 38 |
+
vocab = ["__start__", "adapt", "act", "ap@@", "te", "__end__", "__unk__"]
|
| 39 |
+
vocab_tokens = dict(zip(vocab, range(len(vocab))))
|
| 40 |
+
|
| 41 |
+
merges = ["#version: 0.2", "a p", "t e</w>", "ap t</w>", "a d", "ad apt</w>", "a c", "ac t</w>", ""]
|
| 42 |
+
self.special_tokens_map = {"unk_token": "__unk__", "bos_token": "__start__", "eos_token": "__end__"}
|
| 43 |
+
|
| 44 |
+
self.vocab_file = os.path.join(self.tmpdirname, VOCAB_FILES_NAMES["vocab_file"])
|
| 45 |
+
self.merges_file = os.path.join(self.tmpdirname, VOCAB_FILES_NAMES["merges_file"])
|
| 46 |
+
with open(self.vocab_file, "w", encoding="utf-8") as fp:
|
| 47 |
+
fp.write(json.dumps(vocab_tokens) + "\n")
|
| 48 |
+
with open(self.merges_file, "w", encoding="utf-8") as fp:
|
| 49 |
+
fp.write("\n".join(merges))
|
| 50 |
+
|
| 51 |
+
def get_tokenizer(self, **kwargs):
|
| 52 |
+
kwargs.update(self.special_tokens_map)
|
| 53 |
+
return BlenderbotSmallTokenizer.from_pretrained(self.tmpdirname, **kwargs)
|
| 54 |
+
|
| 55 |
+
def get_input_output_texts(self, tokenizer):
|
| 56 |
+
input_text = "adapt act apte"
|
| 57 |
+
output_text = "adapt act apte"
|
| 58 |
+
return input_text, output_text
|
| 59 |
+
|
| 60 |
+
def test_full_blenderbot_small_tokenizer(self):
|
| 61 |
+
tokenizer = BlenderbotSmallTokenizer(self.vocab_file, self.merges_file, **self.special_tokens_map)
|
| 62 |
+
text = "adapt act apte"
|
| 63 |
+
bpe_tokens = ["adapt", "act", "ap@@", "te"]
|
| 64 |
+
tokens = tokenizer.tokenize(text)
|
| 65 |
+
self.assertListEqual(tokens, bpe_tokens)
|
| 66 |
+
|
| 67 |
+
input_tokens = [tokenizer.bos_token] + tokens + [tokenizer.eos_token]
|
| 68 |
+
|
| 69 |
+
input_bpe_tokens = [0, 1, 2, 3, 4, 5]
|
| 70 |
+
self.assertListEqual(tokenizer.convert_tokens_to_ids(input_tokens), input_bpe_tokens)
|
| 71 |
+
|
| 72 |
+
def test_special_tokens_small_tok(self):
|
| 73 |
+
tok = BlenderbotSmallTokenizer.from_pretrained("facebook/blenderbot-90M")
|
| 74 |
+
assert tok("sam").input_ids == [1384]
|
| 75 |
+
src_text = "I am a small frog."
|
| 76 |
+
encoded = tok([src_text], padding=False, truncation=False)["input_ids"]
|
| 77 |
+
decoded = tok.batch_decode(encoded, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
|
| 78 |
+
assert src_text != decoded # I wish it did!
|
| 79 |
+
assert decoded == "i am a small frog ."
|
| 80 |
+
|
| 81 |
+
def test_empty_word_small_tok(self):
|
| 82 |
+
tok = BlenderbotSmallTokenizer.from_pretrained("facebook/blenderbot-90M")
|
| 83 |
+
src_text = "I am a small frog ."
|
| 84 |
+
src_text_dot = "."
|
| 85 |
+
encoded = tok(src_text)["input_ids"]
|
| 86 |
+
encoded_dot = tok(src_text_dot)["input_ids"]
|
| 87 |
+
|
| 88 |
+
assert encoded[-1] == encoded_dot[0]
|
tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/blip/__init__.py
ADDED
|
File without changes
|
tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/blip/test_image_processing_blip.py
ADDED
|
@@ -0,0 +1,148 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# coding=utf-8
|
| 2 |
+
# Copyright 2022 HuggingFace Inc.
|
| 3 |
+
#
|
| 4 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
| 5 |
+
# you may not use this file except in compliance with the License.
|
| 6 |
+
# You may obtain a copy of the License at
|
| 7 |
+
#
|
| 8 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
| 9 |
+
#
|
| 10 |
+
# Unless required by applicable law or agreed to in writing, software
|
| 11 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
| 12 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
| 13 |
+
# See the License for the specific language governing permissions and
|
| 14 |
+
# limitations under the License.
|
| 15 |
+
|
| 16 |
+
|
| 17 |
+
import unittest
|
| 18 |
+
|
| 19 |
+
from transformers.testing_utils import require_torch, require_vision
|
| 20 |
+
from transformers.utils import is_vision_available
|
| 21 |
+
|
| 22 |
+
from ...test_image_processing_common import ImageProcessingTestMixin, prepare_image_inputs
|
| 23 |
+
|
| 24 |
+
|
| 25 |
+
if is_vision_available():
|
| 26 |
+
from transformers import BlipImageProcessor
|
| 27 |
+
|
| 28 |
+
|
| 29 |
+
class BlipImageProcessingTester(unittest.TestCase):
|
| 30 |
+
def __init__(
|
| 31 |
+
self,
|
| 32 |
+
parent,
|
| 33 |
+
batch_size=7,
|
| 34 |
+
num_channels=3,
|
| 35 |
+
image_size=18,
|
| 36 |
+
min_resolution=30,
|
| 37 |
+
max_resolution=400,
|
| 38 |
+
do_resize=True,
|
| 39 |
+
size=None,
|
| 40 |
+
do_normalize=True,
|
| 41 |
+
do_pad=False,
|
| 42 |
+
image_mean=[0.48145466, 0.4578275, 0.40821073],
|
| 43 |
+
image_std=[0.26862954, 0.26130258, 0.27577711],
|
| 44 |
+
do_convert_rgb=True,
|
| 45 |
+
):
|
| 46 |
+
super().__init__()
|
| 47 |
+
size = size if size is not None else {"height": 20, "width": 20}
|
| 48 |
+
self.parent = parent
|
| 49 |
+
self.batch_size = batch_size
|
| 50 |
+
self.num_channels = num_channels
|
| 51 |
+
self.image_size = image_size
|
| 52 |
+
self.min_resolution = min_resolution
|
| 53 |
+
self.max_resolution = max_resolution
|
| 54 |
+
self.do_resize = do_resize
|
| 55 |
+
self.size = size
|
| 56 |
+
self.do_normalize = do_normalize
|
| 57 |
+
self.image_mean = image_mean
|
| 58 |
+
self.image_std = image_std
|
| 59 |
+
self.do_pad = do_pad
|
| 60 |
+
self.do_convert_rgb = do_convert_rgb
|
| 61 |
+
|
| 62 |
+
def prepare_image_processor_dict(self):
|
| 63 |
+
return {
|
| 64 |
+
"do_resize": self.do_resize,
|
| 65 |
+
"size": self.size,
|
| 66 |
+
"do_normalize": self.do_normalize,
|
| 67 |
+
"image_mean": self.image_mean,
|
| 68 |
+
"image_std": self.image_std,
|
| 69 |
+
"do_convert_rgb": self.do_convert_rgb,
|
| 70 |
+
"do_pad": self.do_pad,
|
| 71 |
+
}
|
| 72 |
+
|
| 73 |
+
def expected_output_image_shape(self, images):
|
| 74 |
+
return self.num_channels, self.size["height"], self.size["width"]
|
| 75 |
+
|
| 76 |
+
def prepare_image_inputs(self, equal_resolution=False, numpify=False, torchify=False):
|
| 77 |
+
return prepare_image_inputs(
|
| 78 |
+
batch_size=self.batch_size,
|
| 79 |
+
num_channels=self.num_channels,
|
| 80 |
+
min_resolution=self.min_resolution,
|
| 81 |
+
max_resolution=self.max_resolution,
|
| 82 |
+
equal_resolution=equal_resolution,
|
| 83 |
+
numpify=numpify,
|
| 84 |
+
torchify=torchify,
|
| 85 |
+
)
|
| 86 |
+
|
| 87 |
+
|
| 88 |
+
@require_torch
|
| 89 |
+
@require_vision
|
| 90 |
+
class BlipImageProcessingTest(ImageProcessingTestMixin, unittest.TestCase):
|
| 91 |
+
image_processing_class = BlipImageProcessor if is_vision_available() else None
|
| 92 |
+
|
| 93 |
+
def setUp(self):
|
| 94 |
+
super().setUp()
|
| 95 |
+
self.image_processor_tester = BlipImageProcessingTester(self)
|
| 96 |
+
|
| 97 |
+
@property
|
| 98 |
+
def image_processor_dict(self):
|
| 99 |
+
return self.image_processor_tester.prepare_image_processor_dict()
|
| 100 |
+
|
| 101 |
+
def test_image_processor_properties(self):
|
| 102 |
+
image_processor = self.image_processing_class(**self.image_processor_dict)
|
| 103 |
+
self.assertTrue(hasattr(image_processor, "do_resize"))
|
| 104 |
+
self.assertTrue(hasattr(image_processor, "size"))
|
| 105 |
+
self.assertTrue(hasattr(image_processor, "do_normalize"))
|
| 106 |
+
self.assertTrue(hasattr(image_processor, "image_mean"))
|
| 107 |
+
self.assertTrue(hasattr(image_processor, "image_std"))
|
| 108 |
+
self.assertTrue(hasattr(image_processor, "do_convert_rgb"))
|
| 109 |
+
|
| 110 |
+
|
| 111 |
+
@require_torch
|
| 112 |
+
@require_vision
|
| 113 |
+
class BlipImageProcessingTestFourChannels(ImageProcessingTestMixin, unittest.TestCase):
|
| 114 |
+
image_processing_class = BlipImageProcessor if is_vision_available() else None
|
| 115 |
+
|
| 116 |
+
def setUp(self):
|
| 117 |
+
super().setUp()
|
| 118 |
+
self.image_processor_tester = BlipImageProcessingTester(self, num_channels=4)
|
| 119 |
+
self.expected_encoded_image_num_channels = 3
|
| 120 |
+
|
| 121 |
+
@property
|
| 122 |
+
def image_processor_dict(self):
|
| 123 |
+
return self.image_processor_tester.prepare_image_processor_dict()
|
| 124 |
+
|
| 125 |
+
def test_image_processor_properties(self):
|
| 126 |
+
image_processor = self.image_processing_class(**self.image_processor_dict)
|
| 127 |
+
self.assertTrue(hasattr(image_processor, "do_resize"))
|
| 128 |
+
self.assertTrue(hasattr(image_processor, "size"))
|
| 129 |
+
self.assertTrue(hasattr(image_processor, "do_normalize"))
|
| 130 |
+
self.assertTrue(hasattr(image_processor, "image_mean"))
|
| 131 |
+
self.assertTrue(hasattr(image_processor, "image_std"))
|
| 132 |
+
self.assertTrue(hasattr(image_processor, "do_convert_rgb"))
|
| 133 |
+
|
| 134 |
+
@unittest.skip(reason="BlipImageProcessor does not support 4 channels yet") # FIXME Amy
|
| 135 |
+
def test_call_numpy(self):
|
| 136 |
+
return super().test_call_numpy()
|
| 137 |
+
|
| 138 |
+
@unittest.skip(reason="BlipImageProcessor does not support 4 channels yet") # FIXME Amy
|
| 139 |
+
def test_call_pytorch(self):
|
| 140 |
+
return super().test_call_torch()
|
| 141 |
+
|
| 142 |
+
@unittest.skip(reason="BLIP doesn't treat 4 channel PIL and numpy consistently yet") # FIXME Amy
|
| 143 |
+
def test_call_pil(self):
|
| 144 |
+
pass
|
| 145 |
+
|
| 146 |
+
@unittest.skip(reason="BLIP doesn't treat 4 channel PIL and numpy consistently yet") # FIXME Amy
|
| 147 |
+
def test_call_numpy_4_channels(self):
|
| 148 |
+
pass
|
tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/blip/test_modeling_blip.py
ADDED
|
@@ -0,0 +1,1435 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# coding=utf-8
|
| 2 |
+
# Copyright 2022 The HuggingFace Inc. team. All rights reserved.
|
| 3 |
+
#
|
| 4 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
| 5 |
+
# you may not use this file except in compliance with the License.
|
| 6 |
+
# You may obtain a copy of the License at
|
| 7 |
+
#
|
| 8 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
| 9 |
+
#
|
| 10 |
+
# Unless required by applicable law or agreed to in writing, software
|
| 11 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
| 12 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
| 13 |
+
# See the License for the specific language governing permissions and
|
| 14 |
+
# limitations under the License.
|
| 15 |
+
"""Testing suite for the PyTorch Blip model."""
|
| 16 |
+
|
| 17 |
+
import inspect
|
| 18 |
+
import os
|
| 19 |
+
import tempfile
|
| 20 |
+
import unittest
|
| 21 |
+
|
| 22 |
+
import numpy as np
|
| 23 |
+
import requests
|
| 24 |
+
|
| 25 |
+
from transformers import BlipConfig, BlipTextConfig, BlipVisionConfig
|
| 26 |
+
from transformers.testing_utils import (
|
| 27 |
+
require_torch,
|
| 28 |
+
require_torch_accelerator,
|
| 29 |
+
require_torch_fp16,
|
| 30 |
+
require_vision,
|
| 31 |
+
slow,
|
| 32 |
+
torch_device,
|
| 33 |
+
)
|
| 34 |
+
from transformers.utils import is_torch_available, is_vision_available
|
| 35 |
+
|
| 36 |
+
from ...test_configuration_common import ConfigTester
|
| 37 |
+
from ...test_modeling_common import (
|
| 38 |
+
ModelTesterMixin,
|
| 39 |
+
_config_zero_init,
|
| 40 |
+
floats_tensor,
|
| 41 |
+
ids_tensor,
|
| 42 |
+
random_attention_mask,
|
| 43 |
+
)
|
| 44 |
+
from ...test_pipeline_mixin import PipelineTesterMixin
|
| 45 |
+
|
| 46 |
+
|
| 47 |
+
if is_torch_available():
|
| 48 |
+
import torch
|
| 49 |
+
from torch import nn
|
| 50 |
+
|
| 51 |
+
from transformers import (
|
| 52 |
+
BlipForConditionalGeneration,
|
| 53 |
+
BlipForImageTextRetrieval,
|
| 54 |
+
BlipForQuestionAnswering,
|
| 55 |
+
BlipModel,
|
| 56 |
+
BlipTextModel,
|
| 57 |
+
BlipVisionModel,
|
| 58 |
+
)
|
| 59 |
+
|
| 60 |
+
|
| 61 |
+
if is_vision_available():
|
| 62 |
+
from PIL import Image
|
| 63 |
+
|
| 64 |
+
from transformers import BlipProcessor
|
| 65 |
+
|
| 66 |
+
|
| 67 |
+
class BlipVisionModelTester:
|
| 68 |
+
def __init__(
|
| 69 |
+
self,
|
| 70 |
+
parent,
|
| 71 |
+
batch_size=12,
|
| 72 |
+
image_size=30,
|
| 73 |
+
patch_size=2,
|
| 74 |
+
num_channels=3,
|
| 75 |
+
is_training=True,
|
| 76 |
+
hidden_size=32,
|
| 77 |
+
projection_dim=32,
|
| 78 |
+
num_hidden_layers=2,
|
| 79 |
+
num_attention_heads=4,
|
| 80 |
+
intermediate_size=37,
|
| 81 |
+
dropout=0.1,
|
| 82 |
+
attention_dropout=0.1,
|
| 83 |
+
initializer_range=1e-10,
|
| 84 |
+
scope=None,
|
| 85 |
+
):
|
| 86 |
+
self.parent = parent
|
| 87 |
+
self.batch_size = batch_size
|
| 88 |
+
self.image_size = image_size
|
| 89 |
+
self.patch_size = patch_size
|
| 90 |
+
self.num_channels = num_channels
|
| 91 |
+
self.is_training = is_training
|
| 92 |
+
self.hidden_size = hidden_size
|
| 93 |
+
self.projection_dim = projection_dim
|
| 94 |
+
self.num_hidden_layers = num_hidden_layers
|
| 95 |
+
self.num_attention_heads = num_attention_heads
|
| 96 |
+
self.intermediate_size = intermediate_size
|
| 97 |
+
self.dropout = dropout
|
| 98 |
+
self.attention_dropout = attention_dropout
|
| 99 |
+
self.initializer_range = initializer_range
|
| 100 |
+
self.scope = scope
|
| 101 |
+
|
| 102 |
+
# in ViT, the seq length equals the number of patches + 1 (we add 1 for the [CLS] token)
|
| 103 |
+
num_patches = (image_size // patch_size) ** 2
|
| 104 |
+
self.seq_length = num_patches + 1
|
| 105 |
+
|
| 106 |
+
def prepare_config_and_inputs(self):
|
| 107 |
+
pixel_values = floats_tensor([self.batch_size, self.num_channels, self.image_size, self.image_size])
|
| 108 |
+
config = self.get_config()
|
| 109 |
+
|
| 110 |
+
return config, pixel_values
|
| 111 |
+
|
| 112 |
+
def get_config(self):
|
| 113 |
+
return BlipVisionConfig(
|
| 114 |
+
image_size=self.image_size,
|
| 115 |
+
patch_size=self.patch_size,
|
| 116 |
+
num_channels=self.num_channels,
|
| 117 |
+
hidden_size=self.hidden_size,
|
| 118 |
+
projection_dim=self.projection_dim,
|
| 119 |
+
num_hidden_layers=self.num_hidden_layers,
|
| 120 |
+
num_attention_heads=self.num_attention_heads,
|
| 121 |
+
intermediate_size=self.intermediate_size,
|
| 122 |
+
dropout=self.dropout,
|
| 123 |
+
attention_dropout=self.attention_dropout,
|
| 124 |
+
initializer_range=self.initializer_range,
|
| 125 |
+
)
|
| 126 |
+
|
| 127 |
+
def create_and_check_model(self, config, pixel_values):
|
| 128 |
+
model = BlipVisionModel(config=config)
|
| 129 |
+
model.to(torch_device)
|
| 130 |
+
model.eval()
|
| 131 |
+
with torch.no_grad():
|
| 132 |
+
result = model(pixel_values)
|
| 133 |
+
# expected sequence length = num_patches + 1 (we add 1 for the [CLS] token)
|
| 134 |
+
image_size = (self.image_size, self.image_size)
|
| 135 |
+
patch_size = (self.patch_size, self.patch_size)
|
| 136 |
+
num_patches = (image_size[1] // patch_size[1]) * (image_size[0] // patch_size[0])
|
| 137 |
+
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, num_patches + 1, self.hidden_size))
|
| 138 |
+
self.parent.assertEqual(result.pooler_output.shape, (self.batch_size, self.hidden_size))
|
| 139 |
+
|
| 140 |
+
def prepare_config_and_inputs_for_common(self):
|
| 141 |
+
config_and_inputs = self.prepare_config_and_inputs()
|
| 142 |
+
config, pixel_values = config_and_inputs
|
| 143 |
+
inputs_dict = {"pixel_values": pixel_values}
|
| 144 |
+
return config, inputs_dict
|
| 145 |
+
|
| 146 |
+
|
| 147 |
+
@require_torch
|
| 148 |
+
class BlipVisionModelTest(ModelTesterMixin, unittest.TestCase):
|
| 149 |
+
"""
|
| 150 |
+
Here we also overwrite some of the tests of test_modeling_common.py, as Blip does not use input_ids, inputs_embeds,
|
| 151 |
+
attention_mask and seq_length.
|
| 152 |
+
"""
|
| 153 |
+
|
| 154 |
+
all_model_classes = (BlipVisionModel,) if is_torch_available() else ()
|
| 155 |
+
fx_compatible = False
|
| 156 |
+
test_pruning = False
|
| 157 |
+
test_resize_embeddings = False
|
| 158 |
+
test_head_masking = False
|
| 159 |
+
|
| 160 |
+
def setUp(self):
|
| 161 |
+
self.model_tester = BlipVisionModelTester(self)
|
| 162 |
+
self.config_tester = ConfigTester(self, config_class=BlipVisionConfig, has_text_modality=False, hidden_size=37)
|
| 163 |
+
|
| 164 |
+
def test_config(self):
|
| 165 |
+
self.config_tester.run_common_tests()
|
| 166 |
+
|
| 167 |
+
@unittest.skip(reason="Blip does not use inputs_embeds")
|
| 168 |
+
def test_inputs_embeds(self):
|
| 169 |
+
pass
|
| 170 |
+
|
| 171 |
+
def test_model_get_set_embeddings(self):
|
| 172 |
+
config, _ = self.model_tester.prepare_config_and_inputs_for_common()
|
| 173 |
+
|
| 174 |
+
for model_class in self.all_model_classes:
|
| 175 |
+
model = model_class(config)
|
| 176 |
+
self.assertIsInstance(model.get_input_embeddings(), (nn.Module))
|
| 177 |
+
x = model.get_output_embeddings()
|
| 178 |
+
self.assertTrue(x is None or isinstance(x, nn.Linear))
|
| 179 |
+
|
| 180 |
+
def test_forward_signature(self):
|
| 181 |
+
config, _ = self.model_tester.prepare_config_and_inputs_for_common()
|
| 182 |
+
|
| 183 |
+
for model_class in self.all_model_classes:
|
| 184 |
+
model = model_class(config)
|
| 185 |
+
signature = inspect.signature(model.forward)
|
| 186 |
+
# signature.parameters is an OrderedDict => so arg_names order is deterministic
|
| 187 |
+
arg_names = [*signature.parameters.keys()]
|
| 188 |
+
|
| 189 |
+
expected_arg_names = ["pixel_values"]
|
| 190 |
+
self.assertListEqual(arg_names[:1], expected_arg_names)
|
| 191 |
+
|
| 192 |
+
def test_model(self):
|
| 193 |
+
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
| 194 |
+
self.model_tester.create_and_check_model(*config_and_inputs)
|
| 195 |
+
|
| 196 |
+
@unittest.skip
|
| 197 |
+
def test_training(self):
|
| 198 |
+
pass
|
| 199 |
+
|
| 200 |
+
@unittest.skip
|
| 201 |
+
def test_training_gradient_checkpointing(self):
|
| 202 |
+
pass
|
| 203 |
+
|
| 204 |
+
@unittest.skip(
|
| 205 |
+
reason="This architecure seem to not compute gradients properly when using GC, check: https://github.com/huggingface/transformers/pull/27124"
|
| 206 |
+
)
|
| 207 |
+
def test_training_gradient_checkpointing_use_reentrant(self):
|
| 208 |
+
pass
|
| 209 |
+
|
| 210 |
+
@unittest.skip(
|
| 211 |
+
reason="This architecure seem to not compute gradients properly when using GC, check: https://github.com/huggingface/transformers/pull/27124"
|
| 212 |
+
)
|
| 213 |
+
def test_training_gradient_checkpointing_use_reentrant_false(self):
|
| 214 |
+
pass
|
| 215 |
+
|
| 216 |
+
@unittest.skip(reason="BlipVisionModel has no base class and is not available in MODEL_MAPPING")
|
| 217 |
+
def test_save_load_fast_init_from_base(self):
|
| 218 |
+
pass
|
| 219 |
+
|
| 220 |
+
@unittest.skip(reason="BlipVisionModel has no base class and is not available in MODEL_MAPPING")
|
| 221 |
+
def test_save_load_fast_init_to_base(self):
|
| 222 |
+
pass
|
| 223 |
+
|
| 224 |
+
@slow
|
| 225 |
+
def test_model_from_pretrained(self):
|
| 226 |
+
model_name = "Salesforce/blip-vqa-base"
|
| 227 |
+
model = BlipVisionModel.from_pretrained(model_name)
|
| 228 |
+
self.assertIsNotNone(model)
|
| 229 |
+
|
| 230 |
+
|
| 231 |
+
class BlipTextModelTester:
|
| 232 |
+
def __init__(
|
| 233 |
+
self,
|
| 234 |
+
parent,
|
| 235 |
+
batch_size=12,
|
| 236 |
+
seq_length=7,
|
| 237 |
+
is_training=True,
|
| 238 |
+
use_input_mask=True,
|
| 239 |
+
use_labels=True,
|
| 240 |
+
vocab_size=99,
|
| 241 |
+
hidden_size=32,
|
| 242 |
+
projection_dim=32,
|
| 243 |
+
num_hidden_layers=2,
|
| 244 |
+
num_attention_heads=4,
|
| 245 |
+
intermediate_size=37,
|
| 246 |
+
dropout=0.1,
|
| 247 |
+
attention_dropout=0.1,
|
| 248 |
+
max_position_embeddings=512,
|
| 249 |
+
initializer_range=0.02,
|
| 250 |
+
bos_token_id=0,
|
| 251 |
+
scope=None,
|
| 252 |
+
):
|
| 253 |
+
self.parent = parent
|
| 254 |
+
self.batch_size = batch_size
|
| 255 |
+
self.seq_length = seq_length
|
| 256 |
+
self.is_training = is_training
|
| 257 |
+
self.use_input_mask = use_input_mask
|
| 258 |
+
self.use_labels = use_labels
|
| 259 |
+
self.vocab_size = vocab_size
|
| 260 |
+
self.hidden_size = hidden_size
|
| 261 |
+
self.projection_dim = projection_dim
|
| 262 |
+
self.num_hidden_layers = num_hidden_layers
|
| 263 |
+
self.num_attention_heads = num_attention_heads
|
| 264 |
+
self.intermediate_size = intermediate_size
|
| 265 |
+
self.dropout = dropout
|
| 266 |
+
self.attention_dropout = attention_dropout
|
| 267 |
+
self.max_position_embeddings = max_position_embeddings
|
| 268 |
+
self.initializer_range = initializer_range
|
| 269 |
+
self.scope = scope
|
| 270 |
+
self.bos_token_id = bos_token_id
|
| 271 |
+
|
| 272 |
+
def prepare_config_and_inputs(self):
|
| 273 |
+
input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
|
| 274 |
+
|
| 275 |
+
input_mask = None
|
| 276 |
+
if self.use_input_mask:
|
| 277 |
+
input_mask = random_attention_mask([self.batch_size, self.seq_length])
|
| 278 |
+
|
| 279 |
+
if input_mask is not None:
|
| 280 |
+
batch_size, seq_length = input_mask.shape
|
| 281 |
+
rnd_start_indices = np.random.randint(1, seq_length - 1, size=(batch_size,))
|
| 282 |
+
for batch_idx, start_index in enumerate(rnd_start_indices):
|
| 283 |
+
input_mask[batch_idx, :start_index] = 1
|
| 284 |
+
input_mask[batch_idx, start_index:] = 0
|
| 285 |
+
|
| 286 |
+
config = self.get_config()
|
| 287 |
+
|
| 288 |
+
return config, input_ids, input_mask
|
| 289 |
+
|
| 290 |
+
def get_config(self):
|
| 291 |
+
return BlipTextConfig(
|
| 292 |
+
vocab_size=self.vocab_size,
|
| 293 |
+
hidden_size=self.hidden_size,
|
| 294 |
+
projection_dim=self.projection_dim,
|
| 295 |
+
num_hidden_layers=self.num_hidden_layers,
|
| 296 |
+
num_attention_heads=self.num_attention_heads,
|
| 297 |
+
intermediate_size=self.intermediate_size,
|
| 298 |
+
dropout=self.dropout,
|
| 299 |
+
attention_dropout=self.attention_dropout,
|
| 300 |
+
max_position_embeddings=self.max_position_embeddings,
|
| 301 |
+
initializer_range=self.initializer_range,
|
| 302 |
+
bos_token_id=self.bos_token_id,
|
| 303 |
+
)
|
| 304 |
+
|
| 305 |
+
def create_and_check_model(self, config, input_ids, input_mask):
|
| 306 |
+
model = BlipTextModel(config=config)
|
| 307 |
+
model.to(torch_device)
|
| 308 |
+
model.eval()
|
| 309 |
+
with torch.no_grad():
|
| 310 |
+
result = model(input_ids, attention_mask=input_mask)
|
| 311 |
+
result = model(input_ids)
|
| 312 |
+
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
|
| 313 |
+
self.parent.assertEqual(result.pooler_output.shape, (self.batch_size, self.hidden_size))
|
| 314 |
+
|
| 315 |
+
def prepare_config_and_inputs_for_common(self):
|
| 316 |
+
config_and_inputs = self.prepare_config_and_inputs()
|
| 317 |
+
config, input_ids, input_mask = config_and_inputs
|
| 318 |
+
inputs_dict = {"input_ids": input_ids, "attention_mask": input_mask}
|
| 319 |
+
return config, inputs_dict
|
| 320 |
+
|
| 321 |
+
|
| 322 |
+
@require_torch
|
| 323 |
+
class BlipTextModelTest(ModelTesterMixin, unittest.TestCase):
|
| 324 |
+
all_model_classes = (BlipTextModel,) if is_torch_available() else ()
|
| 325 |
+
fx_compatible = False
|
| 326 |
+
test_pruning = False
|
| 327 |
+
test_head_masking = False
|
| 328 |
+
|
| 329 |
+
def setUp(self):
|
| 330 |
+
self.model_tester = BlipTextModelTester(self)
|
| 331 |
+
self.config_tester = ConfigTester(self, config_class=BlipTextConfig, hidden_size=37)
|
| 332 |
+
|
| 333 |
+
def test_config(self):
|
| 334 |
+
self.config_tester.run_common_tests()
|
| 335 |
+
|
| 336 |
+
def test_model(self):
|
| 337 |
+
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
| 338 |
+
self.model_tester.create_and_check_model(*config_and_inputs)
|
| 339 |
+
|
| 340 |
+
@unittest.skip
|
| 341 |
+
def test_training(self):
|
| 342 |
+
pass
|
| 343 |
+
|
| 344 |
+
@unittest.skip
|
| 345 |
+
def test_training_gradient_checkpointing(self):
|
| 346 |
+
pass
|
| 347 |
+
|
| 348 |
+
@unittest.skip(
|
| 349 |
+
reason="This architecure seem to not compute gradients properly when using GC, check: https://github.com/huggingface/transformers/pull/27124"
|
| 350 |
+
)
|
| 351 |
+
def test_training_gradient_checkpointing_use_reentrant(self):
|
| 352 |
+
pass
|
| 353 |
+
|
| 354 |
+
@unittest.skip(
|
| 355 |
+
reason="This architecure seem to not compute gradients properly when using GC, check: https://github.com/huggingface/transformers/pull/27124"
|
| 356 |
+
)
|
| 357 |
+
def test_training_gradient_checkpointing_use_reentrant_false(self):
|
| 358 |
+
pass
|
| 359 |
+
|
| 360 |
+
@unittest.skip(reason="Blip does not use inputs_embeds")
|
| 361 |
+
def test_inputs_embeds(self):
|
| 362 |
+
pass
|
| 363 |
+
|
| 364 |
+
@unittest.skip(reason="BlipTextModel has no base class and is not available in MODEL_MAPPING")
|
| 365 |
+
def test_save_load_fast_init_from_base(self):
|
| 366 |
+
pass
|
| 367 |
+
|
| 368 |
+
@unittest.skip(reason="BlipTextModel has no base class and is not available in MODEL_MAPPING")
|
| 369 |
+
def test_save_load_fast_init_to_base(self):
|
| 370 |
+
pass
|
| 371 |
+
|
| 372 |
+
@slow
|
| 373 |
+
def test_model_from_pretrained(self):
|
| 374 |
+
model_name = "Salesforce/blip-vqa-base"
|
| 375 |
+
model = BlipTextModel.from_pretrained(model_name)
|
| 376 |
+
self.assertIsNotNone(model)
|
| 377 |
+
|
| 378 |
+
def test_pt_tf_model_equivalence(self):
|
| 379 |
+
super().test_pt_tf_model_equivalence(allow_missing_keys=True)
|
| 380 |
+
|
| 381 |
+
|
| 382 |
+
class BlipModelTester:
|
| 383 |
+
def __init__(self, parent, text_kwargs=None, vision_kwargs=None, is_training=True):
|
| 384 |
+
if text_kwargs is None:
|
| 385 |
+
text_kwargs = {}
|
| 386 |
+
if vision_kwargs is None:
|
| 387 |
+
vision_kwargs = {}
|
| 388 |
+
|
| 389 |
+
self.parent = parent
|
| 390 |
+
self.text_model_tester = BlipTextModelTester(parent, **text_kwargs)
|
| 391 |
+
self.vision_model_tester = BlipVisionModelTester(parent, **vision_kwargs)
|
| 392 |
+
self.batch_size = self.text_model_tester.batch_size # need bs for batching_equivalence test
|
| 393 |
+
self.is_training = is_training
|
| 394 |
+
|
| 395 |
+
def prepare_config_and_inputs(self):
|
| 396 |
+
text_config, input_ids, attention_mask = self.text_model_tester.prepare_config_and_inputs()
|
| 397 |
+
vision_config, pixel_values = self.vision_model_tester.prepare_config_and_inputs()
|
| 398 |
+
|
| 399 |
+
config = self.get_config()
|
| 400 |
+
|
| 401 |
+
return config, input_ids, attention_mask, pixel_values
|
| 402 |
+
|
| 403 |
+
def get_config(self):
|
| 404 |
+
return BlipConfig.from_text_vision_configs(
|
| 405 |
+
self.text_model_tester.get_config(), self.vision_model_tester.get_config(), projection_dim=64
|
| 406 |
+
)
|
| 407 |
+
|
| 408 |
+
def create_and_check_model(self, config, input_ids, attention_mask, pixel_values):
|
| 409 |
+
model = BlipModel(config).to(torch_device).eval()
|
| 410 |
+
with torch.no_grad():
|
| 411 |
+
result = model(input_ids, pixel_values, attention_mask)
|
| 412 |
+
self.parent.assertEqual(
|
| 413 |
+
result.logits_per_image.shape, (self.vision_model_tester.batch_size, self.text_model_tester.batch_size)
|
| 414 |
+
)
|
| 415 |
+
self.parent.assertEqual(
|
| 416 |
+
result.logits_per_text.shape, (self.text_model_tester.batch_size, self.vision_model_tester.batch_size)
|
| 417 |
+
)
|
| 418 |
+
|
| 419 |
+
def prepare_config_and_inputs_for_common(self):
|
| 420 |
+
config_and_inputs = self.prepare_config_and_inputs()
|
| 421 |
+
config, input_ids, attention_mask, pixel_values = config_and_inputs
|
| 422 |
+
inputs_dict = {
|
| 423 |
+
"input_ids": input_ids,
|
| 424 |
+
"attention_mask": attention_mask,
|
| 425 |
+
"pixel_values": pixel_values,
|
| 426 |
+
"return_loss": True,
|
| 427 |
+
}
|
| 428 |
+
return config, inputs_dict
|
| 429 |
+
|
| 430 |
+
|
| 431 |
+
@require_torch
|
| 432 |
+
class BlipModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase):
|
| 433 |
+
all_model_classes = (BlipModel,) if is_torch_available() else ()
|
| 434 |
+
pipeline_model_mapping = (
|
| 435 |
+
{
|
| 436 |
+
"feature-extraction": BlipModel,
|
| 437 |
+
"image-to-text": BlipForConditionalGeneration,
|
| 438 |
+
"visual-question-answering": BlipForQuestionAnswering,
|
| 439 |
+
"image-text-to-text": BlipForConditionalGeneration,
|
| 440 |
+
}
|
| 441 |
+
if is_torch_available()
|
| 442 |
+
else {}
|
| 443 |
+
)
|
| 444 |
+
fx_compatible = False
|
| 445 |
+
test_head_masking = False
|
| 446 |
+
test_pruning = False
|
| 447 |
+
test_resize_embeddings = True
|
| 448 |
+
test_attention_outputs = False
|
| 449 |
+
|
| 450 |
+
def setUp(self):
|
| 451 |
+
self.model_tester = BlipModelTester(self)
|
| 452 |
+
common_properties = ["logit_scale_init_value", "image_text_hidden_size", "projection_dim", "label_smoothing"]
|
| 453 |
+
self.config_tester = ConfigTester(
|
| 454 |
+
self, config_class=BlipConfig, has_text_modality=False, common_properties=common_properties
|
| 455 |
+
)
|
| 456 |
+
|
| 457 |
+
def test_model(self):
|
| 458 |
+
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
| 459 |
+
self.model_tester.create_and_check_model(*config_and_inputs)
|
| 460 |
+
|
| 461 |
+
def test_config(self):
|
| 462 |
+
self.config_tester.run_common_tests()
|
| 463 |
+
|
| 464 |
+
@unittest.skip(reason="Hidden_states is tested in individual model tests")
|
| 465 |
+
def test_hidden_states_output(self):
|
| 466 |
+
pass
|
| 467 |
+
|
| 468 |
+
@unittest.skip(reason="Inputs_embeds is tested in individual model tests")
|
| 469 |
+
def test_inputs_embeds(self):
|
| 470 |
+
pass
|
| 471 |
+
|
| 472 |
+
@unittest.skip(reason="Retain_grad is tested in individual model tests")
|
| 473 |
+
def test_retain_grad_hidden_states_attentions(self):
|
| 474 |
+
pass
|
| 475 |
+
|
| 476 |
+
@unittest.skip(reason="BlipModel does not have input/output embeddings")
|
| 477 |
+
def test_model_get_set_embeddings(self):
|
| 478 |
+
pass
|
| 479 |
+
|
| 480 |
+
# override as the `logit_scale` parameter initilization is different for Blip
|
| 481 |
+
def test_initialization(self):
|
| 482 |
+
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
| 483 |
+
|
| 484 |
+
configs_no_init = _config_zero_init(config)
|
| 485 |
+
for model_class in self.all_model_classes:
|
| 486 |
+
model = model_class(config=configs_no_init)
|
| 487 |
+
for name, param in model.named_parameters():
|
| 488 |
+
if param.requires_grad:
|
| 489 |
+
# check if `logit_scale` is initilized as per the original implementation
|
| 490 |
+
if name == "logit_scale":
|
| 491 |
+
self.assertAlmostEqual(
|
| 492 |
+
param.data.item(),
|
| 493 |
+
np.log(1 / 0.07),
|
| 494 |
+
delta=1e-3,
|
| 495 |
+
msg=f"Parameter {name} of model {model_class} seems not properly initialized",
|
| 496 |
+
)
|
| 497 |
+
else:
|
| 498 |
+
self.assertIn(
|
| 499 |
+
((param.data.mean() * 1e9).round() / 1e9).item(),
|
| 500 |
+
[0.0, 1.0],
|
| 501 |
+
msg=f"Parameter {name} of model {model_class} seems not properly initialized",
|
| 502 |
+
)
|
| 503 |
+
|
| 504 |
+
def _create_and_check_torchscript(self, config, inputs_dict):
|
| 505 |
+
if not self.test_torchscript:
|
| 506 |
+
self.skipTest(reason="test_torchscript is set to False")
|
| 507 |
+
|
| 508 |
+
configs_no_init = _config_zero_init(config) # To be sure we have no Nan
|
| 509 |
+
configs_no_init.torchscript = True
|
| 510 |
+
configs_no_init.return_dict = False
|
| 511 |
+
for model_class in self.all_model_classes:
|
| 512 |
+
model = model_class(config=configs_no_init)
|
| 513 |
+
model.to(torch_device)
|
| 514 |
+
model.eval()
|
| 515 |
+
|
| 516 |
+
try:
|
| 517 |
+
input_ids = inputs_dict["input_ids"]
|
| 518 |
+
pixel_values = inputs_dict["pixel_values"] # Blip needs pixel_values
|
| 519 |
+
traced_model = torch.jit.trace(model, (input_ids, pixel_values))
|
| 520 |
+
except RuntimeError:
|
| 521 |
+
self.fail("Couldn't trace module.")
|
| 522 |
+
|
| 523 |
+
with tempfile.TemporaryDirectory() as tmp_dir_name:
|
| 524 |
+
pt_file_name = os.path.join(tmp_dir_name, "traced_model.pt")
|
| 525 |
+
|
| 526 |
+
try:
|
| 527 |
+
torch.jit.save(traced_model, pt_file_name)
|
| 528 |
+
except Exception:
|
| 529 |
+
self.fail("Couldn't save module.")
|
| 530 |
+
|
| 531 |
+
try:
|
| 532 |
+
loaded_model = torch.jit.load(pt_file_name)
|
| 533 |
+
except Exception:
|
| 534 |
+
self.fail("Couldn't load module.")
|
| 535 |
+
|
| 536 |
+
model.to(torch_device)
|
| 537 |
+
model.eval()
|
| 538 |
+
|
| 539 |
+
loaded_model.to(torch_device)
|
| 540 |
+
loaded_model.eval()
|
| 541 |
+
|
| 542 |
+
model_state_dict = model.state_dict()
|
| 543 |
+
loaded_model_state_dict = loaded_model.state_dict()
|
| 544 |
+
|
| 545 |
+
non_persistent_buffers = {}
|
| 546 |
+
for key in loaded_model_state_dict.keys():
|
| 547 |
+
if key not in model_state_dict.keys():
|
| 548 |
+
non_persistent_buffers[key] = loaded_model_state_dict[key]
|
| 549 |
+
|
| 550 |
+
loaded_model_state_dict = {
|
| 551 |
+
key: value for key, value in loaded_model_state_dict.items() if key not in non_persistent_buffers
|
| 552 |
+
}
|
| 553 |
+
|
| 554 |
+
self.assertEqual(set(model_state_dict.keys()), set(loaded_model_state_dict.keys()))
|
| 555 |
+
|
| 556 |
+
model_buffers = list(model.buffers())
|
| 557 |
+
for non_persistent_buffer in non_persistent_buffers.values():
|
| 558 |
+
found_buffer = False
|
| 559 |
+
for i, model_buffer in enumerate(model_buffers):
|
| 560 |
+
if torch.equal(non_persistent_buffer, model_buffer):
|
| 561 |
+
found_buffer = True
|
| 562 |
+
break
|
| 563 |
+
|
| 564 |
+
self.assertTrue(found_buffer)
|
| 565 |
+
model_buffers.pop(i)
|
| 566 |
+
|
| 567 |
+
models_equal = True
|
| 568 |
+
for layer_name, p1 in model_state_dict.items():
|
| 569 |
+
p2 = loaded_model_state_dict[layer_name]
|
| 570 |
+
if p1.data.ne(p2.data).sum() > 0:
|
| 571 |
+
models_equal = False
|
| 572 |
+
|
| 573 |
+
self.assertTrue(models_equal)
|
| 574 |
+
|
| 575 |
+
def test_load_vision_text_config(self):
|
| 576 |
+
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
| 577 |
+
|
| 578 |
+
# Save BlipConfig and check if we can load BlipVisionConfig from it
|
| 579 |
+
with tempfile.TemporaryDirectory() as tmp_dir_name:
|
| 580 |
+
config.save_pretrained(tmp_dir_name)
|
| 581 |
+
vision_config = BlipVisionConfig.from_pretrained(tmp_dir_name)
|
| 582 |
+
self.assertDictEqual(config.vision_config.to_dict(), vision_config.to_dict())
|
| 583 |
+
|
| 584 |
+
# Save BlipConfig and check if we can load BlipTextConfig from it
|
| 585 |
+
with tempfile.TemporaryDirectory() as tmp_dir_name:
|
| 586 |
+
config.save_pretrained(tmp_dir_name)
|
| 587 |
+
text_config = BlipTextConfig.from_pretrained(tmp_dir_name)
|
| 588 |
+
self.assertDictEqual(config.text_config.to_dict(), text_config.to_dict())
|
| 589 |
+
|
| 590 |
+
@slow
|
| 591 |
+
def test_model_from_pretrained(self):
|
| 592 |
+
model_name = "Salesforce/blip-vqa-base"
|
| 593 |
+
model = BlipModel.from_pretrained(model_name)
|
| 594 |
+
self.assertIsNotNone(model)
|
| 595 |
+
|
| 596 |
+
def test_get_image_features(self):
|
| 597 |
+
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
| 598 |
+
|
| 599 |
+
keys_to_pop = ["input_ids", "attention_mask", "return_loss"]
|
| 600 |
+
|
| 601 |
+
for key in keys_to_pop:
|
| 602 |
+
inputs_dict.pop(key)
|
| 603 |
+
|
| 604 |
+
model = BlipModel(config).to(torch_device)
|
| 605 |
+
model.eval()
|
| 606 |
+
image_features = model.get_image_features(**inputs_dict)
|
| 607 |
+
self.assertEqual(
|
| 608 |
+
image_features.shape,
|
| 609 |
+
(
|
| 610 |
+
self.model_tester.batch_size,
|
| 611 |
+
model.projection_dim,
|
| 612 |
+
),
|
| 613 |
+
)
|
| 614 |
+
|
| 615 |
+
def test_get_text_features(self):
|
| 616 |
+
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
| 617 |
+
|
| 618 |
+
keys_to_pop = ["pixel_values", "return_loss"]
|
| 619 |
+
|
| 620 |
+
for key in keys_to_pop:
|
| 621 |
+
inputs_dict.pop(key)
|
| 622 |
+
|
| 623 |
+
model = BlipModel(config).to(torch_device)
|
| 624 |
+
model.eval()
|
| 625 |
+
text_features = model.get_text_features(**inputs_dict)
|
| 626 |
+
self.assertEqual(
|
| 627 |
+
text_features.shape,
|
| 628 |
+
(
|
| 629 |
+
self.model_tester.batch_size,
|
| 630 |
+
model.projection_dim,
|
| 631 |
+
),
|
| 632 |
+
)
|
| 633 |
+
|
| 634 |
+
def test_get_multimodal_features(self):
|
| 635 |
+
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
| 636 |
+
|
| 637 |
+
keys_to_pop = ["return_loss"]
|
| 638 |
+
|
| 639 |
+
for key in keys_to_pop:
|
| 640 |
+
inputs_dict.pop(key)
|
| 641 |
+
|
| 642 |
+
model = BlipModel(config).to(torch_device)
|
| 643 |
+
model.eval()
|
| 644 |
+
multimodal_features = model.get_multimodal_features(**inputs_dict)
|
| 645 |
+
self.assertEqual(
|
| 646 |
+
multimodal_features.shape,
|
| 647 |
+
(
|
| 648 |
+
self.model_tester.batch_size,
|
| 649 |
+
model.projection_dim,
|
| 650 |
+
),
|
| 651 |
+
)
|
| 652 |
+
|
| 653 |
+
def test_pt_tf_model_equivalence(self):
|
| 654 |
+
super().test_pt_tf_model_equivalence(allow_missing_keys=True)
|
| 655 |
+
|
| 656 |
+
|
| 657 |
+
class BlipTextRetrievalModelTester:
|
| 658 |
+
def __init__(self, parent, text_kwargs=None, vision_kwargs=None, is_training=True):
|
| 659 |
+
if text_kwargs is None:
|
| 660 |
+
text_kwargs = {}
|
| 661 |
+
if vision_kwargs is None:
|
| 662 |
+
vision_kwargs = {}
|
| 663 |
+
|
| 664 |
+
self.parent = parent
|
| 665 |
+
self.text_model_tester = BlipTextModelTester(parent, **text_kwargs)
|
| 666 |
+
self.vision_model_tester = BlipVisionModelTester(parent, **vision_kwargs)
|
| 667 |
+
self.batch_size = self.text_model_tester.batch_size # need bs for batching_equivalence test
|
| 668 |
+
self.is_training = is_training
|
| 669 |
+
|
| 670 |
+
def prepare_config_and_inputs(self):
|
| 671 |
+
text_config, input_ids, attention_mask = self.text_model_tester.prepare_config_and_inputs()
|
| 672 |
+
vision_config, pixel_values = self.vision_model_tester.prepare_config_and_inputs()
|
| 673 |
+
|
| 674 |
+
config = self.get_config()
|
| 675 |
+
|
| 676 |
+
return config, input_ids, attention_mask, pixel_values
|
| 677 |
+
|
| 678 |
+
def get_config(self):
|
| 679 |
+
return BlipConfig.from_text_vision_configs(
|
| 680 |
+
self.text_model_tester.get_config(), self.vision_model_tester.get_config(), projection_dim=64
|
| 681 |
+
)
|
| 682 |
+
|
| 683 |
+
def create_and_check_model(self, config, input_ids, attention_mask, pixel_values):
|
| 684 |
+
model = BlipModel(config).to(torch_device).eval()
|
| 685 |
+
with torch.no_grad():
|
| 686 |
+
result = model(input_ids, pixel_values, attention_mask)
|
| 687 |
+
self.parent.assertEqual(
|
| 688 |
+
result.logits_per_image.shape, (self.vision_model_tester.batch_size, self.text_model_tester.batch_size)
|
| 689 |
+
)
|
| 690 |
+
self.parent.assertEqual(
|
| 691 |
+
result.logits_per_text.shape, (self.text_model_tester.batch_size, self.vision_model_tester.batch_size)
|
| 692 |
+
)
|
| 693 |
+
|
| 694 |
+
def prepare_config_and_inputs_for_common(self):
|
| 695 |
+
config_and_inputs = self.prepare_config_and_inputs()
|
| 696 |
+
config, input_ids, attention_mask, pixel_values = config_and_inputs
|
| 697 |
+
inputs_dict = {
|
| 698 |
+
"input_ids": input_ids,
|
| 699 |
+
"attention_mask": attention_mask,
|
| 700 |
+
"pixel_values": pixel_values,
|
| 701 |
+
}
|
| 702 |
+
return config, inputs_dict
|
| 703 |
+
|
| 704 |
+
|
| 705 |
+
class BlipTextImageModelsModelTester:
|
| 706 |
+
def __init__(self, parent, text_kwargs=None, vision_kwargs=None, is_training=True):
|
| 707 |
+
if text_kwargs is None:
|
| 708 |
+
text_kwargs = {}
|
| 709 |
+
if vision_kwargs is None:
|
| 710 |
+
vision_kwargs = {}
|
| 711 |
+
|
| 712 |
+
self.parent = parent
|
| 713 |
+
self.text_model_tester = BlipTextModelTester(parent, **text_kwargs)
|
| 714 |
+
self.vision_model_tester = BlipVisionModelTester(parent, **vision_kwargs)
|
| 715 |
+
self.batch_size = self.text_model_tester.batch_size # need bs for batching_equivalence test
|
| 716 |
+
self.seq_length = self.text_model_tester.seq_length # need seq_length for pt-tf equivalence test
|
| 717 |
+
self.is_training = is_training
|
| 718 |
+
|
| 719 |
+
def prepare_config_and_inputs(self):
|
| 720 |
+
text_config, input_ids, attention_mask = self.text_model_tester.prepare_config_and_inputs()
|
| 721 |
+
vision_config, pixel_values = self.vision_model_tester.prepare_config_and_inputs()
|
| 722 |
+
|
| 723 |
+
config = self.get_config()
|
| 724 |
+
|
| 725 |
+
return config, input_ids, attention_mask, pixel_values
|
| 726 |
+
|
| 727 |
+
def get_config(self):
|
| 728 |
+
return BlipConfig.from_text_vision_configs(
|
| 729 |
+
self.text_model_tester.get_config(), self.vision_model_tester.get_config(), projection_dim=64
|
| 730 |
+
)
|
| 731 |
+
|
| 732 |
+
def create_and_check_model(self, config, input_ids, attention_mask, pixel_values):
|
| 733 |
+
model = BlipModel(config).to(torch_device).eval()
|
| 734 |
+
with torch.no_grad():
|
| 735 |
+
result = model(input_ids, pixel_values, attention_mask)
|
| 736 |
+
self.parent.assertEqual(
|
| 737 |
+
result.logits_per_image.shape, (self.vision_model_tester.batch_size, self.text_model_tester.batch_size)
|
| 738 |
+
)
|
| 739 |
+
self.parent.assertEqual(
|
| 740 |
+
result.logits_per_text.shape, (self.text_model_tester.batch_size, self.vision_model_tester.batch_size)
|
| 741 |
+
)
|
| 742 |
+
|
| 743 |
+
def prepare_config_and_inputs_for_common(self):
|
| 744 |
+
config_and_inputs = self.prepare_config_and_inputs()
|
| 745 |
+
config, input_ids, attention_mask, pixel_values = config_and_inputs
|
| 746 |
+
inputs_dict = {
|
| 747 |
+
"input_ids": input_ids,
|
| 748 |
+
"attention_mask": attention_mask,
|
| 749 |
+
"pixel_values": pixel_values,
|
| 750 |
+
}
|
| 751 |
+
return config, inputs_dict
|
| 752 |
+
|
| 753 |
+
|
| 754 |
+
class BlipVQAModelTester:
|
| 755 |
+
def __init__(self, parent, text_kwargs=None, vision_kwargs=None, is_training=True):
|
| 756 |
+
if text_kwargs is None:
|
| 757 |
+
text_kwargs = {}
|
| 758 |
+
if vision_kwargs is None:
|
| 759 |
+
vision_kwargs = {}
|
| 760 |
+
|
| 761 |
+
self.parent = parent
|
| 762 |
+
self.text_model_tester = BlipTextModelTester(parent, **text_kwargs)
|
| 763 |
+
self.vision_model_tester = BlipVisionModelTester(parent, **vision_kwargs)
|
| 764 |
+
self.batch_size = self.text_model_tester.batch_size # need bs for batching_equivalence test
|
| 765 |
+
self.is_training = is_training
|
| 766 |
+
|
| 767 |
+
def prepare_config_and_inputs(self):
|
| 768 |
+
text_config, input_ids, attention_mask = self.text_model_tester.prepare_config_and_inputs()
|
| 769 |
+
vision_config, pixel_values = self.vision_model_tester.prepare_config_and_inputs()
|
| 770 |
+
|
| 771 |
+
config = self.get_config()
|
| 772 |
+
|
| 773 |
+
return config, input_ids, attention_mask, pixel_values
|
| 774 |
+
|
| 775 |
+
def get_config(self):
|
| 776 |
+
return BlipConfig.from_text_vision_configs(
|
| 777 |
+
self.text_model_tester.get_config(), self.vision_model_tester.get_config(), projection_dim=64
|
| 778 |
+
)
|
| 779 |
+
|
| 780 |
+
def create_and_check_model(self, config, input_ids, attention_mask, pixel_values):
|
| 781 |
+
model = BlipModel(config).to(torch_device).eval()
|
| 782 |
+
with torch.no_grad():
|
| 783 |
+
result = model(input_ids, pixel_values, attention_mask)
|
| 784 |
+
self.parent.assertEqual(
|
| 785 |
+
result.logits_per_image.shape, (self.vision_model_tester.batch_size, self.text_model_tester.batch_size)
|
| 786 |
+
)
|
| 787 |
+
self.parent.assertEqual(
|
| 788 |
+
result.logits_per_text.shape, (self.text_model_tester.batch_size, self.vision_model_tester.batch_size)
|
| 789 |
+
)
|
| 790 |
+
|
| 791 |
+
def prepare_config_and_inputs_for_common(self):
|
| 792 |
+
config_and_inputs = self.prepare_config_and_inputs()
|
| 793 |
+
config, input_ids, attention_mask, pixel_values = config_and_inputs
|
| 794 |
+
inputs_dict = {
|
| 795 |
+
"input_ids": input_ids,
|
| 796 |
+
"decoder_input_ids": input_ids,
|
| 797 |
+
"attention_mask": attention_mask,
|
| 798 |
+
"pixel_values": pixel_values,
|
| 799 |
+
"labels": input_ids,
|
| 800 |
+
}
|
| 801 |
+
return config, inputs_dict
|
| 802 |
+
|
| 803 |
+
|
| 804 |
+
@require_torch
|
| 805 |
+
@require_vision
|
| 806 |
+
class BlipVQAModelTest(ModelTesterMixin, unittest.TestCase):
|
| 807 |
+
all_model_classes = (BlipForQuestionAnswering,) if is_torch_available() else ()
|
| 808 |
+
fx_compatible = False
|
| 809 |
+
test_head_masking = False
|
| 810 |
+
test_pruning = False
|
| 811 |
+
test_resize_embeddings = True
|
| 812 |
+
test_attention_outputs = False
|
| 813 |
+
test_torchscript = False
|
| 814 |
+
|
| 815 |
+
def setUp(self):
|
| 816 |
+
self.model_tester = BlipVQAModelTester(self)
|
| 817 |
+
|
| 818 |
+
def _prepare_inputs_for_vqa(self):
|
| 819 |
+
_, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
| 820 |
+
inputs_dict["decoder_input_ids"] = inputs_dict["input_ids"]
|
| 821 |
+
inputs_dict.pop("return_loss")
|
| 822 |
+
return inputs_dict
|
| 823 |
+
|
| 824 |
+
def test_class_name_consistency(self):
|
| 825 |
+
"""
|
| 826 |
+
Tests that all VQA models have a class name that ends with "ForQuestionAnswering"
|
| 827 |
+
"""
|
| 828 |
+
for model_class in self.all_model_classes:
|
| 829 |
+
model = model_class(self.model_tester.get_config())
|
| 830 |
+
self.assertTrue(
|
| 831 |
+
model.__class__.__name__.endswith("ForQuestionAnswering"),
|
| 832 |
+
f"Class name should end with 'ForVisualQuestionAnswering' got {model.__class__.__name__}",
|
| 833 |
+
)
|
| 834 |
+
|
| 835 |
+
def test_training(self):
|
| 836 |
+
"""
|
| 837 |
+
Tests that all VQA models can be trained on a single batch
|
| 838 |
+
"""
|
| 839 |
+
for model_class in self.all_model_classes:
|
| 840 |
+
model = model_class(self.model_tester.get_config()).to(torch_device)
|
| 841 |
+
model.train()
|
| 842 |
+
loss = model(**self.model_tester.prepare_config_and_inputs_for_common()[1]).loss
|
| 843 |
+
loss.backward()
|
| 844 |
+
|
| 845 |
+
# verify the gradients are not None
|
| 846 |
+
for name, param in model.named_parameters():
|
| 847 |
+
self.assertIsNotNone(param.grad, f"Gradients should not be None - got {param.grad} for {name}")
|
| 848 |
+
|
| 849 |
+
def test_forward_signature(self):
|
| 850 |
+
"""
|
| 851 |
+
Test if the forward function has the expected arguments.
|
| 852 |
+
"""
|
| 853 |
+
for model_class in self.all_model_classes:
|
| 854 |
+
model = model_class(self.model_tester.get_config())
|
| 855 |
+
signature = inspect.signature(model.forward)
|
| 856 |
+
# signature.parameters is an OrderedDict => so args are the first n entries
|
| 857 |
+
args = list(signature.parameters.keys())
|
| 858 |
+
expected_args = [
|
| 859 |
+
"input_ids",
|
| 860 |
+
"attention_mask",
|
| 861 |
+
"labels",
|
| 862 |
+
"decoder_input_ids",
|
| 863 |
+
"decoder_attention_mask",
|
| 864 |
+
]
|
| 865 |
+
for arg in expected_args:
|
| 866 |
+
self.assertTrue(
|
| 867 |
+
arg in args,
|
| 868 |
+
f"Argument {arg} of forward function signature should include {arg}. Found {args}.",
|
| 869 |
+
)
|
| 870 |
+
|
| 871 |
+
@unittest.skip(reason="Hidden_states is tested in individual model tests")
|
| 872 |
+
def test_hidden_states_output(self):
|
| 873 |
+
pass
|
| 874 |
+
|
| 875 |
+
@unittest.skip(reason="Inputs_embeds is tested in individual model tests")
|
| 876 |
+
def test_inputs_embeds(self):
|
| 877 |
+
pass
|
| 878 |
+
|
| 879 |
+
@unittest.skip(reason="BlipModel does not have input/output embeddings")
|
| 880 |
+
def test_model_get_set_embeddings(self):
|
| 881 |
+
pass
|
| 882 |
+
|
| 883 |
+
|
| 884 |
+
@require_torch
|
| 885 |
+
class BlipTextRetrievalModelTest(ModelTesterMixin, unittest.TestCase):
|
| 886 |
+
all_model_classes = (BlipForImageTextRetrieval,) if is_torch_available() else ()
|
| 887 |
+
fx_compatible = False
|
| 888 |
+
test_head_masking = False
|
| 889 |
+
test_pruning = False
|
| 890 |
+
test_resize_embeddings = True
|
| 891 |
+
test_attention_outputs = False
|
| 892 |
+
test_torchscript = False
|
| 893 |
+
|
| 894 |
+
def setUp(self):
|
| 895 |
+
self.model_tester = BlipTextRetrievalModelTester(self)
|
| 896 |
+
|
| 897 |
+
def test_model(self):
|
| 898 |
+
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
| 899 |
+
self.model_tester.create_and_check_model(*config_and_inputs)
|
| 900 |
+
|
| 901 |
+
@unittest.skip(reason="Hidden_states is tested in individual model tests")
|
| 902 |
+
def test_hidden_states_output(self):
|
| 903 |
+
pass
|
| 904 |
+
|
| 905 |
+
@unittest.skip(reason="Inputs_embeds is tested in individual model tests")
|
| 906 |
+
def test_inputs_embeds(self):
|
| 907 |
+
pass
|
| 908 |
+
|
| 909 |
+
@unittest.skip(reason="Retain_grad is tested in individual model tests")
|
| 910 |
+
def test_retain_grad_hidden_states_attentions(self):
|
| 911 |
+
pass
|
| 912 |
+
|
| 913 |
+
@unittest.skip(reason="BlipModel does not have input/output embeddings")
|
| 914 |
+
def test_model_get_set_embeddings(self):
|
| 915 |
+
pass
|
| 916 |
+
|
| 917 |
+
def test_forward_signature(self):
|
| 918 |
+
config, _ = self.model_tester.prepare_config_and_inputs_for_common()
|
| 919 |
+
|
| 920 |
+
for model_class in self.all_model_classes:
|
| 921 |
+
model = model_class(config)
|
| 922 |
+
signature = inspect.signature(model.forward)
|
| 923 |
+
# signature.parameters is an OrderedDict => so arg_names order is deterministic
|
| 924 |
+
arg_names = [*signature.parameters.keys()]
|
| 925 |
+
|
| 926 |
+
if model.config.is_encoder_decoder:
|
| 927 |
+
expected_arg_names = [
|
| 928 |
+
"input_ids",
|
| 929 |
+
"attention_mask",
|
| 930 |
+
"decoder_input_ids",
|
| 931 |
+
"decoder_attention_mask",
|
| 932 |
+
]
|
| 933 |
+
expected_arg_names.extend(
|
| 934 |
+
["head_mask", "decoder_head_mask", "cross_attn_head_mask", "encoder_outputs"]
|
| 935 |
+
if "head_mask" and "decoder_head_mask" and "cross_attn_head_mask" in arg_names
|
| 936 |
+
else ["encoder_outputs"]
|
| 937 |
+
)
|
| 938 |
+
self.assertListEqual(arg_names[: len(expected_arg_names)], expected_arg_names)
|
| 939 |
+
else:
|
| 940 |
+
expected_arg_names = ["input_ids"] if model_class != BlipForConditionalGeneration else ["pixel_values"]
|
| 941 |
+
self.assertListEqual(arg_names[:1], expected_arg_names)
|
| 942 |
+
|
| 943 |
+
def test_training(self):
|
| 944 |
+
if not self.model_tester.is_training:
|
| 945 |
+
self.skipTest(reason="ModelTester is not setup for training")
|
| 946 |
+
|
| 947 |
+
for model_class in self.all_model_classes[:-1]:
|
| 948 |
+
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
| 949 |
+
config.return_dict = True
|
| 950 |
+
|
| 951 |
+
model = model_class(config)
|
| 952 |
+
model.to(torch_device)
|
| 953 |
+
model.train()
|
| 954 |
+
inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
|
| 955 |
+
|
| 956 |
+
# hardcode labels to be the same as input_ids
|
| 957 |
+
inputs["labels"] = inputs["input_ids"]
|
| 958 |
+
|
| 959 |
+
loss = model(**inputs).loss
|
| 960 |
+
loss.backward()
|
| 961 |
+
|
| 962 |
+
def test_training_gradient_checkpointing(self):
|
| 963 |
+
if not self.model_tester.is_training:
|
| 964 |
+
self.skipTest(reason="ModelTester is not setup for training")
|
| 965 |
+
|
| 966 |
+
for model_class in self.all_model_classes[:-1]:
|
| 967 |
+
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
| 968 |
+
config.use_cache = False
|
| 969 |
+
config.return_dict = True
|
| 970 |
+
|
| 971 |
+
model = model_class(config)
|
| 972 |
+
model.to(torch_device)
|
| 973 |
+
model.gradient_checkpointing_enable()
|
| 974 |
+
model.train()
|
| 975 |
+
inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
|
| 976 |
+
|
| 977 |
+
# hardcode labels to be the same as input_ids
|
| 978 |
+
inputs["labels"] = inputs["input_ids"]
|
| 979 |
+
|
| 980 |
+
loss = model(**inputs).loss
|
| 981 |
+
loss.backward()
|
| 982 |
+
|
| 983 |
+
@unittest.skip(
|
| 984 |
+
reason="This architecure seem to not compute gradients properly when using GC, check: https://github.com/huggingface/transformers/pull/27124"
|
| 985 |
+
)
|
| 986 |
+
def test_training_gradient_checkpointing_use_reentrant(self):
|
| 987 |
+
pass
|
| 988 |
+
|
| 989 |
+
@unittest.skip(
|
| 990 |
+
reason="This architecure seem to not compute gradients properly when using GC, check: https://github.com/huggingface/transformers/pull/27124"
|
| 991 |
+
)
|
| 992 |
+
def test_training_gradient_checkpointing_use_reentrant_false(self):
|
| 993 |
+
pass
|
| 994 |
+
|
| 995 |
+
# override as the `logit_scale` parameter initilization is different for Blip
|
| 996 |
+
def test_initialization(self):
|
| 997 |
+
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
| 998 |
+
|
| 999 |
+
configs_no_init = _config_zero_init(config)
|
| 1000 |
+
for model_class in self.all_model_classes:
|
| 1001 |
+
model = model_class(config=configs_no_init)
|
| 1002 |
+
for name, param in model.named_parameters():
|
| 1003 |
+
if param.requires_grad:
|
| 1004 |
+
# check if `logit_scale` is initilized as per the original implementation
|
| 1005 |
+
if name == "logit_scale":
|
| 1006 |
+
self.assertAlmostEqual(
|
| 1007 |
+
param.data.item(),
|
| 1008 |
+
np.log(1 / 0.07),
|
| 1009 |
+
delta=1e-3,
|
| 1010 |
+
msg=f"Parameter {name} of model {model_class} seems not properly initialized",
|
| 1011 |
+
)
|
| 1012 |
+
else:
|
| 1013 |
+
self.assertIn(
|
| 1014 |
+
((param.data.mean() * 1e9).round() / 1e9).item(),
|
| 1015 |
+
[0.0, 1.0],
|
| 1016 |
+
msg=f"Parameter {name} of model {model_class} seems not properly initialized",
|
| 1017 |
+
)
|
| 1018 |
+
|
| 1019 |
+
def _create_and_check_torchscript(self, config, inputs_dict):
|
| 1020 |
+
if not self.test_torchscript:
|
| 1021 |
+
self.skipTest(reason="test_torchscript is set to False")
|
| 1022 |
+
|
| 1023 |
+
configs_no_init = _config_zero_init(config) # To be sure we have no Nan
|
| 1024 |
+
configs_no_init.torchscript = True
|
| 1025 |
+
configs_no_init.return_dict = False
|
| 1026 |
+
for model_class in self.all_model_classes:
|
| 1027 |
+
model = model_class(config=configs_no_init)
|
| 1028 |
+
model.to(torch_device)
|
| 1029 |
+
model.eval()
|
| 1030 |
+
|
| 1031 |
+
try:
|
| 1032 |
+
input_ids = inputs_dict["input_ids"]
|
| 1033 |
+
pixel_values = inputs_dict["pixel_values"] # Blip needs pixel_values
|
| 1034 |
+
traced_model = torch.jit.trace(model, (input_ids, pixel_values))
|
| 1035 |
+
except RuntimeError:
|
| 1036 |
+
self.fail("Couldn't trace module.")
|
| 1037 |
+
|
| 1038 |
+
with tempfile.TemporaryDirectory() as tmp_dir_name:
|
| 1039 |
+
pt_file_name = os.path.join(tmp_dir_name, "traced_model.pt")
|
| 1040 |
+
|
| 1041 |
+
try:
|
| 1042 |
+
torch.jit.save(traced_model, pt_file_name)
|
| 1043 |
+
except Exception:
|
| 1044 |
+
self.fail("Couldn't save module.")
|
| 1045 |
+
|
| 1046 |
+
try:
|
| 1047 |
+
loaded_model = torch.jit.load(pt_file_name)
|
| 1048 |
+
except Exception:
|
| 1049 |
+
self.fail("Couldn't load module.")
|
| 1050 |
+
|
| 1051 |
+
model.to(torch_device)
|
| 1052 |
+
model.eval()
|
| 1053 |
+
|
| 1054 |
+
loaded_model.to(torch_device)
|
| 1055 |
+
loaded_model.eval()
|
| 1056 |
+
|
| 1057 |
+
model_state_dict = model.state_dict()
|
| 1058 |
+
loaded_model_state_dict = loaded_model.state_dict()
|
| 1059 |
+
|
| 1060 |
+
non_persistent_buffers = {}
|
| 1061 |
+
for key in loaded_model_state_dict.keys():
|
| 1062 |
+
if key not in model_state_dict.keys():
|
| 1063 |
+
non_persistent_buffers[key] = loaded_model_state_dict[key]
|
| 1064 |
+
|
| 1065 |
+
loaded_model_state_dict = {
|
| 1066 |
+
key: value for key, value in loaded_model_state_dict.items() if key not in non_persistent_buffers
|
| 1067 |
+
}
|
| 1068 |
+
|
| 1069 |
+
self.assertEqual(set(model_state_dict.keys()), set(loaded_model_state_dict.keys()))
|
| 1070 |
+
|
| 1071 |
+
model_buffers = list(model.buffers())
|
| 1072 |
+
for non_persistent_buffer in non_persistent_buffers.values():
|
| 1073 |
+
found_buffer = False
|
| 1074 |
+
for i, model_buffer in enumerate(model_buffers):
|
| 1075 |
+
if torch.equal(non_persistent_buffer, model_buffer):
|
| 1076 |
+
found_buffer = True
|
| 1077 |
+
break
|
| 1078 |
+
|
| 1079 |
+
self.assertTrue(found_buffer)
|
| 1080 |
+
model_buffers.pop(i)
|
| 1081 |
+
|
| 1082 |
+
models_equal = True
|
| 1083 |
+
for layer_name, p1 in model_state_dict.items():
|
| 1084 |
+
p2 = loaded_model_state_dict[layer_name]
|
| 1085 |
+
if p1.data.ne(p2.data).sum() > 0:
|
| 1086 |
+
models_equal = False
|
| 1087 |
+
|
| 1088 |
+
self.assertTrue(models_equal)
|
| 1089 |
+
|
| 1090 |
+
def test_load_vision_text_config(self):
|
| 1091 |
+
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
| 1092 |
+
|
| 1093 |
+
# Save BlipConfig and check if we can load BlipVisionConfig from it
|
| 1094 |
+
with tempfile.TemporaryDirectory() as tmp_dir_name:
|
| 1095 |
+
config.save_pretrained(tmp_dir_name)
|
| 1096 |
+
vision_config = BlipVisionConfig.from_pretrained(tmp_dir_name)
|
| 1097 |
+
self.assertDictEqual(config.vision_config.to_dict(), vision_config.to_dict())
|
| 1098 |
+
|
| 1099 |
+
# Save BlipConfig and check if we can load BlipTextConfig from it
|
| 1100 |
+
with tempfile.TemporaryDirectory() as tmp_dir_name:
|
| 1101 |
+
config.save_pretrained(tmp_dir_name)
|
| 1102 |
+
text_config = BlipTextConfig.from_pretrained(tmp_dir_name)
|
| 1103 |
+
self.assertDictEqual(config.text_config.to_dict(), text_config.to_dict())
|
| 1104 |
+
|
| 1105 |
+
@slow
|
| 1106 |
+
def test_model_from_pretrained(self):
|
| 1107 |
+
model_name = "Salesforce/blip-vqa-base"
|
| 1108 |
+
model = BlipModel.from_pretrained(model_name)
|
| 1109 |
+
self.assertIsNotNone(model)
|
| 1110 |
+
|
| 1111 |
+
|
| 1112 |
+
@require_torch
|
| 1113 |
+
class BlipTextImageModelTest(ModelTesterMixin, unittest.TestCase):
|
| 1114 |
+
all_model_classes = (BlipForConditionalGeneration,) if is_torch_available() else ()
|
| 1115 |
+
fx_compatible = False
|
| 1116 |
+
test_head_masking = False
|
| 1117 |
+
test_pruning = False
|
| 1118 |
+
test_resize_embeddings = True
|
| 1119 |
+
test_attention_outputs = False
|
| 1120 |
+
test_torchscript = False
|
| 1121 |
+
|
| 1122 |
+
def setUp(self):
|
| 1123 |
+
self.model_tester = BlipTextImageModelsModelTester(self)
|
| 1124 |
+
|
| 1125 |
+
def test_model(self):
|
| 1126 |
+
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
| 1127 |
+
self.model_tester.create_and_check_model(*config_and_inputs)
|
| 1128 |
+
|
| 1129 |
+
@unittest.skip(reason="Hidden_states is tested in individual model tests")
|
| 1130 |
+
def test_hidden_states_output(self):
|
| 1131 |
+
pass
|
| 1132 |
+
|
| 1133 |
+
@unittest.skip(reason="Inputs_embeds is tested in individual model tests")
|
| 1134 |
+
def test_inputs_embeds(self):
|
| 1135 |
+
pass
|
| 1136 |
+
|
| 1137 |
+
@unittest.skip(reason="Retain_grad is tested in individual model tests")
|
| 1138 |
+
def test_retain_grad_hidden_states_attentions(self):
|
| 1139 |
+
pass
|
| 1140 |
+
|
| 1141 |
+
@unittest.skip(reason="BlipModel does not have input/output embeddings")
|
| 1142 |
+
def test_model_get_set_embeddings(self):
|
| 1143 |
+
pass
|
| 1144 |
+
|
| 1145 |
+
def test_forward_signature(self):
|
| 1146 |
+
config, _ = self.model_tester.prepare_config_and_inputs_for_common()
|
| 1147 |
+
|
| 1148 |
+
for model_class in self.all_model_classes:
|
| 1149 |
+
model = model_class(config)
|
| 1150 |
+
signature = inspect.signature(model.forward)
|
| 1151 |
+
# signature.parameters is an OrderedDict => so arg_names order is deterministic
|
| 1152 |
+
arg_names = [*signature.parameters.keys()]
|
| 1153 |
+
|
| 1154 |
+
if model.config.is_encoder_decoder:
|
| 1155 |
+
expected_arg_names = [
|
| 1156 |
+
"input_ids",
|
| 1157 |
+
"attention_mask",
|
| 1158 |
+
"decoder_input_ids",
|
| 1159 |
+
"decoder_attention_mask",
|
| 1160 |
+
]
|
| 1161 |
+
expected_arg_names.extend(
|
| 1162 |
+
["head_mask", "decoder_head_mask", "cross_attn_head_mask", "encoder_outputs"]
|
| 1163 |
+
if "head_mask" and "decoder_head_mask" and "cross_attn_head_mask" in arg_names
|
| 1164 |
+
else ["encoder_outputs"]
|
| 1165 |
+
)
|
| 1166 |
+
self.assertListEqual(arg_names[: len(expected_arg_names)], expected_arg_names)
|
| 1167 |
+
else:
|
| 1168 |
+
expected_arg_names = ["input_ids"] if model_class != BlipForConditionalGeneration else ["pixel_values"]
|
| 1169 |
+
self.assertListEqual(arg_names[:1], expected_arg_names)
|
| 1170 |
+
|
| 1171 |
+
def test_training(self):
|
| 1172 |
+
if not self.model_tester.is_training:
|
| 1173 |
+
self.skipTest(reason="ModelTester is not setup for training")
|
| 1174 |
+
|
| 1175 |
+
for model_class in self.all_model_classes[:-1]:
|
| 1176 |
+
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
| 1177 |
+
config.return_dict = True
|
| 1178 |
+
|
| 1179 |
+
model = model_class(config)
|
| 1180 |
+
model.to(torch_device)
|
| 1181 |
+
model.train()
|
| 1182 |
+
inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
|
| 1183 |
+
|
| 1184 |
+
# hardcode labels to be the same as input_ids
|
| 1185 |
+
inputs["labels"] = inputs["input_ids"]
|
| 1186 |
+
|
| 1187 |
+
loss = model(**inputs).loss
|
| 1188 |
+
loss.backward()
|
| 1189 |
+
|
| 1190 |
+
def test_training_gradient_checkpointing(self):
|
| 1191 |
+
if not self.model_tester.is_training:
|
| 1192 |
+
self.skipTest(reason="ModelTester is not setup for training")
|
| 1193 |
+
|
| 1194 |
+
for model_class in self.all_model_classes[:-1]:
|
| 1195 |
+
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
| 1196 |
+
config.use_cache = False
|
| 1197 |
+
config.return_dict = True
|
| 1198 |
+
|
| 1199 |
+
model = model_class(config)
|
| 1200 |
+
model.to(torch_device)
|
| 1201 |
+
model.gradient_checkpointing_enable()
|
| 1202 |
+
model.train()
|
| 1203 |
+
inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
|
| 1204 |
+
|
| 1205 |
+
# hardcode labels to be the same as input_ids
|
| 1206 |
+
inputs["labels"] = inputs["input_ids"]
|
| 1207 |
+
|
| 1208 |
+
loss = model(**inputs).loss
|
| 1209 |
+
loss.backward()
|
| 1210 |
+
|
| 1211 |
+
# override as the `logit_scale` parameter initilization is different for Blip
|
| 1212 |
+
def test_initialization(self):
|
| 1213 |
+
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
| 1214 |
+
|
| 1215 |
+
configs_no_init = _config_zero_init(config)
|
| 1216 |
+
for model_class in self.all_model_classes:
|
| 1217 |
+
model = model_class(config=configs_no_init)
|
| 1218 |
+
for name, param in model.named_parameters():
|
| 1219 |
+
if param.requires_grad:
|
| 1220 |
+
# check if `logit_scale` is initilized as per the original implementation
|
| 1221 |
+
if name == "logit_scale":
|
| 1222 |
+
self.assertAlmostEqual(
|
| 1223 |
+
param.data.item(),
|
| 1224 |
+
np.log(1 / 0.07),
|
| 1225 |
+
delta=1e-3,
|
| 1226 |
+
msg=f"Parameter {name} of model {model_class} seems not properly initialized",
|
| 1227 |
+
)
|
| 1228 |
+
else:
|
| 1229 |
+
self.assertIn(
|
| 1230 |
+
((param.data.mean() * 1e9).round() / 1e9).item(),
|
| 1231 |
+
[0.0, 1.0],
|
| 1232 |
+
msg=f"Parameter {name} of model {model_class} seems not properly initialized",
|
| 1233 |
+
)
|
| 1234 |
+
|
| 1235 |
+
def _create_and_check_torchscript(self, config, inputs_dict):
|
| 1236 |
+
if not self.test_torchscript:
|
| 1237 |
+
self.skipTest(reason="test_torchscript is set to False")
|
| 1238 |
+
|
| 1239 |
+
configs_no_init = _config_zero_init(config) # To be sure we have no Nan
|
| 1240 |
+
configs_no_init.torchscript = True
|
| 1241 |
+
configs_no_init.return_dict = False
|
| 1242 |
+
for model_class in self.all_model_classes:
|
| 1243 |
+
model = model_class(config=configs_no_init)
|
| 1244 |
+
model.to(torch_device)
|
| 1245 |
+
model.eval()
|
| 1246 |
+
|
| 1247 |
+
try:
|
| 1248 |
+
input_ids = inputs_dict["input_ids"]
|
| 1249 |
+
pixel_values = inputs_dict["pixel_values"] # Blip needs pixel_values
|
| 1250 |
+
traced_model = torch.jit.trace(model, (input_ids, pixel_values))
|
| 1251 |
+
except RuntimeError:
|
| 1252 |
+
self.fail("Couldn't trace module.")
|
| 1253 |
+
|
| 1254 |
+
with tempfile.TemporaryDirectory() as tmp_dir_name:
|
| 1255 |
+
pt_file_name = os.path.join(tmp_dir_name, "traced_model.pt")
|
| 1256 |
+
|
| 1257 |
+
try:
|
| 1258 |
+
torch.jit.save(traced_model, pt_file_name)
|
| 1259 |
+
except Exception:
|
| 1260 |
+
self.fail("Couldn't save module.")
|
| 1261 |
+
|
| 1262 |
+
try:
|
| 1263 |
+
loaded_model = torch.jit.load(pt_file_name)
|
| 1264 |
+
except Exception:
|
| 1265 |
+
self.fail("Couldn't load module.")
|
| 1266 |
+
|
| 1267 |
+
model.to(torch_device)
|
| 1268 |
+
model.eval()
|
| 1269 |
+
|
| 1270 |
+
loaded_model.to(torch_device)
|
| 1271 |
+
loaded_model.eval()
|
| 1272 |
+
|
| 1273 |
+
model_state_dict = model.state_dict()
|
| 1274 |
+
loaded_model_state_dict = loaded_model.state_dict()
|
| 1275 |
+
|
| 1276 |
+
non_persistent_buffers = {}
|
| 1277 |
+
for key in loaded_model_state_dict.keys():
|
| 1278 |
+
if key not in model_state_dict.keys():
|
| 1279 |
+
non_persistent_buffers[key] = loaded_model_state_dict[key]
|
| 1280 |
+
|
| 1281 |
+
loaded_model_state_dict = {
|
| 1282 |
+
key: value for key, value in loaded_model_state_dict.items() if key not in non_persistent_buffers
|
| 1283 |
+
}
|
| 1284 |
+
|
| 1285 |
+
self.assertEqual(set(model_state_dict.keys()), set(loaded_model_state_dict.keys()))
|
| 1286 |
+
|
| 1287 |
+
model_buffers = list(model.buffers())
|
| 1288 |
+
for non_persistent_buffer in non_persistent_buffers.values():
|
| 1289 |
+
found_buffer = False
|
| 1290 |
+
for i, model_buffer in enumerate(model_buffers):
|
| 1291 |
+
if torch.equal(non_persistent_buffer, model_buffer):
|
| 1292 |
+
found_buffer = True
|
| 1293 |
+
break
|
| 1294 |
+
|
| 1295 |
+
self.assertTrue(found_buffer)
|
| 1296 |
+
model_buffers.pop(i)
|
| 1297 |
+
|
| 1298 |
+
models_equal = True
|
| 1299 |
+
for layer_name, p1 in model_state_dict.items():
|
| 1300 |
+
p2 = loaded_model_state_dict[layer_name]
|
| 1301 |
+
if p1.data.ne(p2.data).sum() > 0:
|
| 1302 |
+
models_equal = False
|
| 1303 |
+
|
| 1304 |
+
self.assertTrue(models_equal)
|
| 1305 |
+
|
| 1306 |
+
def test_load_vision_text_config(self):
|
| 1307 |
+
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
| 1308 |
+
|
| 1309 |
+
# Save BlipConfig and check if we can load BlipVisionConfig from it
|
| 1310 |
+
with tempfile.TemporaryDirectory() as tmp_dir_name:
|
| 1311 |
+
config.save_pretrained(tmp_dir_name)
|
| 1312 |
+
vision_config = BlipVisionConfig.from_pretrained(tmp_dir_name)
|
| 1313 |
+
self.assertDictEqual(config.vision_config.to_dict(), vision_config.to_dict())
|
| 1314 |
+
|
| 1315 |
+
# Save BlipConfig and check if we can load BlipTextConfig from it
|
| 1316 |
+
with tempfile.TemporaryDirectory() as tmp_dir_name:
|
| 1317 |
+
config.save_pretrained(tmp_dir_name)
|
| 1318 |
+
text_config = BlipTextConfig.from_pretrained(tmp_dir_name)
|
| 1319 |
+
self.assertDictEqual(config.text_config.to_dict(), text_config.to_dict())
|
| 1320 |
+
|
| 1321 |
+
@slow
|
| 1322 |
+
def test_model_from_pretrained(self):
|
| 1323 |
+
model_name = "Salesforce/blip-vqa-base"
|
| 1324 |
+
model = BlipModel.from_pretrained(model_name)
|
| 1325 |
+
self.assertIsNotNone(model)
|
| 1326 |
+
|
| 1327 |
+
|
| 1328 |
+
# We will verify our results on an image of cute cats
|
| 1329 |
+
def prepare_img():
|
| 1330 |
+
url = "https://huggingface.co/hf-internal-testing/blip-test-image/resolve/main/demo.jpg"
|
| 1331 |
+
im = Image.open(requests.get(url, stream=True).raw)
|
| 1332 |
+
return im
|
| 1333 |
+
|
| 1334 |
+
|
| 1335 |
+
@require_vision
|
| 1336 |
+
@require_torch
|
| 1337 |
+
@slow
|
| 1338 |
+
class BlipModelIntegrationTest(unittest.TestCase):
|
| 1339 |
+
def test_inference_image_captioning(self):
|
| 1340 |
+
model = BlipForConditionalGeneration.from_pretrained("Salesforce/blip-image-captioning-base").to(torch_device)
|
| 1341 |
+
processor = BlipProcessor.from_pretrained("Salesforce/blip-image-captioning-base")
|
| 1342 |
+
image = prepare_img()
|
| 1343 |
+
|
| 1344 |
+
# image only
|
| 1345 |
+
inputs = processor(images=image, return_tensors="pt").to(torch_device)
|
| 1346 |
+
|
| 1347 |
+
predictions = model.generate(**inputs)
|
| 1348 |
+
|
| 1349 |
+
# Test output
|
| 1350 |
+
self.assertEqual(predictions[0].tolist(), [30522, 1037, 2450, 3564, 2006, 1996, 3509, 2007, 2014, 3899, 102])
|
| 1351 |
+
|
| 1352 |
+
# image and context
|
| 1353 |
+
context = ["a picture of"]
|
| 1354 |
+
inputs = processor(images=image, text=context, return_tensors="pt").to(torch_device)
|
| 1355 |
+
|
| 1356 |
+
predictions = model.generate(**inputs)
|
| 1357 |
+
|
| 1358 |
+
# Test output
|
| 1359 |
+
self.assertEqual(
|
| 1360 |
+
predictions[0].tolist(),
|
| 1361 |
+
[30522, 1037, 3861, 1997, 1037, 2450, 1998, 2014, 3899, 2006, 1996, 3509, 102],
|
| 1362 |
+
)
|
| 1363 |
+
|
| 1364 |
+
@require_torch_accelerator
|
| 1365 |
+
@require_torch_fp16
|
| 1366 |
+
def test_inference_image_captioning_fp16(self):
|
| 1367 |
+
model = BlipForConditionalGeneration.from_pretrained(
|
| 1368 |
+
"Salesforce/blip-image-captioning-base", torch_dtype=torch.float16
|
| 1369 |
+
).to(torch_device)
|
| 1370 |
+
processor = BlipProcessor.from_pretrained("Salesforce/blip-image-captioning-base")
|
| 1371 |
+
image = prepare_img()
|
| 1372 |
+
|
| 1373 |
+
# image only
|
| 1374 |
+
inputs = processor(images=image, return_tensors="pt").to(torch_device, torch.float16)
|
| 1375 |
+
|
| 1376 |
+
predictions = model.generate(**inputs)
|
| 1377 |
+
|
| 1378 |
+
# Test output
|
| 1379 |
+
self.assertEqual(predictions[0].tolist(), [30522, 1037, 2450, 3564, 2006, 1996, 3509, 2007, 2014, 3899, 102])
|
| 1380 |
+
|
| 1381 |
+
# image and context
|
| 1382 |
+
context = ["a picture of"]
|
| 1383 |
+
inputs = processor(images=image, text=context, return_tensors="pt").to(torch_device, torch.float16)
|
| 1384 |
+
|
| 1385 |
+
predictions = model.generate(**inputs)
|
| 1386 |
+
|
| 1387 |
+
# Test output
|
| 1388 |
+
self.assertEqual(
|
| 1389 |
+
predictions[0].tolist(),
|
| 1390 |
+
[30522, 1037, 3861, 1997, 1037, 2450, 1998, 2014, 3899, 2006, 1996, 3509, 102],
|
| 1391 |
+
)
|
| 1392 |
+
|
| 1393 |
+
def test_inference_interpolate_pos_encoding(self):
|
| 1394 |
+
model = BlipForConditionalGeneration.from_pretrained("Salesforce/blip-image-captioning-base").to(torch_device)
|
| 1395 |
+
processor = BlipProcessor.from_pretrained("Salesforce/blip-image-captioning-base")
|
| 1396 |
+
processor.image_processor.size = {"height": 500, "width": 500}
|
| 1397 |
+
|
| 1398 |
+
image = prepare_img()
|
| 1399 |
+
inputs = processor(images=image, return_tensors="pt").to(torch_device)
|
| 1400 |
+
|
| 1401 |
+
predictions = model.generate(**inputs, interpolate_pos_encoding=True)
|
| 1402 |
+
generated_text = processor.batch_decode(predictions, skip_special_tokens=True)[0].strip()
|
| 1403 |
+
|
| 1404 |
+
self.assertEqual(predictions[0].tolist(), [30522, 1037, 2450, 3564, 2006, 1996, 3509, 2007, 1037, 3899, 102])
|
| 1405 |
+
self.assertEqual(generated_text, "a woman sitting on the beach with a dog")
|
| 1406 |
+
|
| 1407 |
+
def test_inference_vqa(self):
|
| 1408 |
+
model = BlipForQuestionAnswering.from_pretrained("Salesforce/blip-vqa-base").to(torch_device)
|
| 1409 |
+
processor = BlipProcessor.from_pretrained("Salesforce/blip-vqa-base")
|
| 1410 |
+
|
| 1411 |
+
image = prepare_img()
|
| 1412 |
+
text = "how many dogs are in the picture?"
|
| 1413 |
+
|
| 1414 |
+
inputs = processor(image, text=text, return_tensors="pt").to(torch_device)
|
| 1415 |
+
out = model.generate(**inputs)
|
| 1416 |
+
|
| 1417 |
+
# Test output
|
| 1418 |
+
self.assertEqual(out[0].tolist(), [30522, 1015, 102])
|
| 1419 |
+
|
| 1420 |
+
def test_inference_itm(self):
|
| 1421 |
+
model = BlipForImageTextRetrieval.from_pretrained("Salesforce/blip-itm-base-coco").to(torch_device)
|
| 1422 |
+
processor = BlipProcessor.from_pretrained("Salesforce/blip-itm-base-coco")
|
| 1423 |
+
|
| 1424 |
+
image = prepare_img()
|
| 1425 |
+
text = "A woman and her dog sitting in a beach"
|
| 1426 |
+
|
| 1427 |
+
inputs = processor(image, text, return_tensors="pt").to(torch_device)
|
| 1428 |
+
|
| 1429 |
+
out_itm = model(**inputs)
|
| 1430 |
+
out = model(**inputs, use_itm_head=False)
|
| 1431 |
+
|
| 1432 |
+
expected_scores = torch.Tensor([[0.0029, 0.9971]])
|
| 1433 |
+
|
| 1434 |
+
self.assertTrue(torch.allclose(torch.nn.Softmax()(out_itm[0].cpu()), expected_scores, rtol=1e-3, atol=1e-3))
|
| 1435 |
+
self.assertTrue(torch.allclose(out[0].cpu(), torch.Tensor([[0.5162]]), rtol=1e-3, atol=1e-3))
|
tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/blip/test_modeling_blip_text.py
ADDED
|
@@ -0,0 +1,183 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# coding=utf-8
|
| 2 |
+
# Copyright 2022 The HuggingFace Inc. team. All rights reserved.
|
| 3 |
+
#
|
| 4 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
| 5 |
+
# you may not use this file except in compliance with the License.
|
| 6 |
+
# You may obtain a copy of the License at
|
| 7 |
+
#
|
| 8 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
| 9 |
+
#
|
| 10 |
+
# Unless required by applicable law or agreed to in writing, software
|
| 11 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
| 12 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
| 13 |
+
# See the License for the specific language governing permissions and
|
| 14 |
+
# limitations under the License.
|
| 15 |
+
"""Testing suite for the PyTorch Blip model."""
|
| 16 |
+
|
| 17 |
+
import unittest
|
| 18 |
+
|
| 19 |
+
import numpy as np
|
| 20 |
+
|
| 21 |
+
from transformers import BlipTextConfig
|
| 22 |
+
from transformers.testing_utils import require_torch, slow, torch_device
|
| 23 |
+
from transformers.utils import is_torch_available
|
| 24 |
+
|
| 25 |
+
from ...test_configuration_common import ConfigTester
|
| 26 |
+
from ...test_modeling_common import ModelTesterMixin, ids_tensor, random_attention_mask
|
| 27 |
+
|
| 28 |
+
|
| 29 |
+
if is_torch_available():
|
| 30 |
+
import torch
|
| 31 |
+
|
| 32 |
+
from transformers import BlipTextModel
|
| 33 |
+
|
| 34 |
+
|
| 35 |
+
class BlipTextModelTester:
|
| 36 |
+
def __init__(
|
| 37 |
+
self,
|
| 38 |
+
parent,
|
| 39 |
+
batch_size=12,
|
| 40 |
+
seq_length=7,
|
| 41 |
+
is_training=True,
|
| 42 |
+
use_input_mask=True,
|
| 43 |
+
use_labels=True,
|
| 44 |
+
vocab_size=99,
|
| 45 |
+
hidden_size=32,
|
| 46 |
+
projection_dim=32,
|
| 47 |
+
num_hidden_layers=2,
|
| 48 |
+
num_attention_heads=4,
|
| 49 |
+
intermediate_size=37,
|
| 50 |
+
dropout=0.1,
|
| 51 |
+
attention_dropout=0.1,
|
| 52 |
+
max_position_embeddings=512,
|
| 53 |
+
initializer_range=0.02,
|
| 54 |
+
bos_token_id=0,
|
| 55 |
+
scope=None,
|
| 56 |
+
):
|
| 57 |
+
self.parent = parent
|
| 58 |
+
self.batch_size = batch_size
|
| 59 |
+
self.seq_length = seq_length
|
| 60 |
+
self.is_training = is_training
|
| 61 |
+
self.use_input_mask = use_input_mask
|
| 62 |
+
self.use_labels = use_labels
|
| 63 |
+
self.vocab_size = vocab_size
|
| 64 |
+
self.hidden_size = hidden_size
|
| 65 |
+
self.projection_dim = projection_dim
|
| 66 |
+
self.num_hidden_layers = num_hidden_layers
|
| 67 |
+
self.num_attention_heads = num_attention_heads
|
| 68 |
+
self.intermediate_size = intermediate_size
|
| 69 |
+
self.dropout = dropout
|
| 70 |
+
self.attention_dropout = attention_dropout
|
| 71 |
+
self.max_position_embeddings = max_position_embeddings
|
| 72 |
+
self.initializer_range = initializer_range
|
| 73 |
+
self.scope = scope
|
| 74 |
+
self.bos_token_id = bos_token_id
|
| 75 |
+
|
| 76 |
+
def prepare_config_and_inputs(self):
|
| 77 |
+
input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
|
| 78 |
+
|
| 79 |
+
input_mask = None
|
| 80 |
+
if self.use_input_mask:
|
| 81 |
+
input_mask = random_attention_mask([self.batch_size, self.seq_length])
|
| 82 |
+
|
| 83 |
+
if input_mask is not None:
|
| 84 |
+
batch_size, seq_length = input_mask.shape
|
| 85 |
+
rnd_start_indices = np.random.randint(1, seq_length - 1, size=(batch_size,))
|
| 86 |
+
for batch_idx, start_index in enumerate(rnd_start_indices):
|
| 87 |
+
input_mask[batch_idx, :start_index] = 1
|
| 88 |
+
input_mask[batch_idx, start_index:] = 0
|
| 89 |
+
|
| 90 |
+
config = self.get_config()
|
| 91 |
+
|
| 92 |
+
return config, input_ids, input_mask
|
| 93 |
+
|
| 94 |
+
def get_config(self):
|
| 95 |
+
return BlipTextConfig(
|
| 96 |
+
vocab_size=self.vocab_size,
|
| 97 |
+
hidden_size=self.hidden_size,
|
| 98 |
+
projection_dim=self.projection_dim,
|
| 99 |
+
num_hidden_layers=self.num_hidden_layers,
|
| 100 |
+
num_attention_heads=self.num_attention_heads,
|
| 101 |
+
intermediate_size=self.intermediate_size,
|
| 102 |
+
dropout=self.dropout,
|
| 103 |
+
attention_dropout=self.attention_dropout,
|
| 104 |
+
max_position_embeddings=self.max_position_embeddings,
|
| 105 |
+
initializer_range=self.initializer_range,
|
| 106 |
+
bos_token_id=self.bos_token_id,
|
| 107 |
+
)
|
| 108 |
+
|
| 109 |
+
def create_and_check_model(self, config, input_ids, input_mask):
|
| 110 |
+
model = BlipTextModel(config=config)
|
| 111 |
+
model.to(torch_device)
|
| 112 |
+
model.eval()
|
| 113 |
+
with torch.no_grad():
|
| 114 |
+
result = model(input_ids, attention_mask=input_mask)
|
| 115 |
+
result = model(input_ids)
|
| 116 |
+
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
|
| 117 |
+
self.parent.assertEqual(result.pooler_output.shape, (self.batch_size, self.hidden_size))
|
| 118 |
+
|
| 119 |
+
def prepare_config_and_inputs_for_common(self):
|
| 120 |
+
config_and_inputs = self.prepare_config_and_inputs()
|
| 121 |
+
config, input_ids, input_mask = config_and_inputs
|
| 122 |
+
inputs_dict = {"input_ids": input_ids, "attention_mask": input_mask}
|
| 123 |
+
return config, inputs_dict
|
| 124 |
+
|
| 125 |
+
|
| 126 |
+
@require_torch
|
| 127 |
+
class BlipTextModelTest(ModelTesterMixin, unittest.TestCase):
|
| 128 |
+
all_model_classes = (BlipTextModel,) if is_torch_available() else ()
|
| 129 |
+
fx_compatible = False
|
| 130 |
+
test_pruning = False
|
| 131 |
+
test_head_masking = False
|
| 132 |
+
|
| 133 |
+
def setUp(self):
|
| 134 |
+
self.model_tester = BlipTextModelTester(self)
|
| 135 |
+
self.config_tester = ConfigTester(self, config_class=BlipTextConfig, hidden_size=37)
|
| 136 |
+
|
| 137 |
+
def test_config(self):
|
| 138 |
+
self.config_tester.run_common_tests()
|
| 139 |
+
|
| 140 |
+
def test_model(self):
|
| 141 |
+
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
| 142 |
+
self.model_tester.create_and_check_model(*config_and_inputs)
|
| 143 |
+
|
| 144 |
+
@unittest.skip
|
| 145 |
+
def test_training(self):
|
| 146 |
+
pass
|
| 147 |
+
|
| 148 |
+
@unittest.skip
|
| 149 |
+
def test_training_gradient_checkpointing(self):
|
| 150 |
+
pass
|
| 151 |
+
|
| 152 |
+
@unittest.skip(
|
| 153 |
+
reason="This architecure seem to not compute gradients properly when using GC, check: https://github.com/huggingface/transformers/pull/27124"
|
| 154 |
+
)
|
| 155 |
+
def test_training_gradient_checkpointing_use_reentrant(self):
|
| 156 |
+
pass
|
| 157 |
+
|
| 158 |
+
@unittest.skip(
|
| 159 |
+
reason="This architecure seem to not compute gradients properly when using GC, check: https://github.com/huggingface/transformers/pull/27124"
|
| 160 |
+
)
|
| 161 |
+
def test_training_gradient_checkpointing_use_reentrant_false(self):
|
| 162 |
+
pass
|
| 163 |
+
|
| 164 |
+
@unittest.skip(reason="Blip does not use inputs_embeds")
|
| 165 |
+
def test_inputs_embeds(self):
|
| 166 |
+
pass
|
| 167 |
+
|
| 168 |
+
@unittest.skip(reason="BlipTextModel has no base class and is not available in MODEL_MAPPING")
|
| 169 |
+
def test_save_load_fast_init_from_base(self):
|
| 170 |
+
pass
|
| 171 |
+
|
| 172 |
+
@unittest.skip(reason="BlipTextModel has no base class and is not available in MODEL_MAPPING")
|
| 173 |
+
def test_save_load_fast_init_to_base(self):
|
| 174 |
+
pass
|
| 175 |
+
|
| 176 |
+
@slow
|
| 177 |
+
def test_model_from_pretrained(self):
|
| 178 |
+
model_name = "Salesforce/blip-vqa-base"
|
| 179 |
+
model = BlipTextModel.from_pretrained(model_name)
|
| 180 |
+
self.assertIsNotNone(model)
|
| 181 |
+
|
| 182 |
+
def test_pt_tf_model_equivalence(self):
|
| 183 |
+
super().test_pt_tf_model_equivalence(allow_missing_keys=True)
|
tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/blip/test_modeling_tf_blip.py
ADDED
|
@@ -0,0 +1,901 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# coding=utf-8
|
| 2 |
+
# Copyright 2023 The HuggingFace Inc. team. All rights reserved.
|
| 3 |
+
#
|
| 4 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
| 5 |
+
# you may not use this file except in compliance with the License.
|
| 6 |
+
# You may obtain a copy of the License at
|
| 7 |
+
#
|
| 8 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
| 9 |
+
#
|
| 10 |
+
# Unless required by applicable law or agreed to in writing, software
|
| 11 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
| 12 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
| 13 |
+
# See the License for the specific language governing permissions and
|
| 14 |
+
# limitations under the License.
|
| 15 |
+
"""Testing suite for the TensorFlow Blip model."""
|
| 16 |
+
|
| 17 |
+
from __future__ import annotations
|
| 18 |
+
|
| 19 |
+
import inspect
|
| 20 |
+
import tempfile
|
| 21 |
+
import unittest
|
| 22 |
+
|
| 23 |
+
import numpy as np
|
| 24 |
+
import requests
|
| 25 |
+
|
| 26 |
+
from transformers import BlipConfig, BlipTextConfig, BlipVisionConfig
|
| 27 |
+
from transformers.testing_utils import require_tf, require_vision, slow
|
| 28 |
+
from transformers.utils import is_tf_available, is_vision_available
|
| 29 |
+
|
| 30 |
+
from ...test_configuration_common import ConfigTester
|
| 31 |
+
from ...test_modeling_tf_common import TFModelTesterMixin, floats_tensor, ids_tensor, random_attention_mask
|
| 32 |
+
from ...test_pipeline_mixin import PipelineTesterMixin
|
| 33 |
+
|
| 34 |
+
|
| 35 |
+
if is_tf_available():
|
| 36 |
+
import tensorflow as tf
|
| 37 |
+
|
| 38 |
+
from transformers import (
|
| 39 |
+
TFBlipForConditionalGeneration,
|
| 40 |
+
TFBlipForImageTextRetrieval,
|
| 41 |
+
TFBlipForQuestionAnswering,
|
| 42 |
+
TFBlipModel,
|
| 43 |
+
TFBlipTextModel,
|
| 44 |
+
TFBlipVisionModel,
|
| 45 |
+
)
|
| 46 |
+
from transformers.modeling_tf_utils import keras
|
| 47 |
+
|
| 48 |
+
|
| 49 |
+
if is_vision_available():
|
| 50 |
+
from PIL import Image
|
| 51 |
+
|
| 52 |
+
from transformers import BlipProcessor
|
| 53 |
+
|
| 54 |
+
|
| 55 |
+
class TFBlipVisionModelTester:
|
| 56 |
+
def __init__(
|
| 57 |
+
self,
|
| 58 |
+
parent,
|
| 59 |
+
batch_size=12,
|
| 60 |
+
image_size=30,
|
| 61 |
+
patch_size=2,
|
| 62 |
+
num_channels=3,
|
| 63 |
+
is_training=True,
|
| 64 |
+
hidden_size=32,
|
| 65 |
+
projection_dim=32,
|
| 66 |
+
num_hidden_layers=2,
|
| 67 |
+
num_attention_heads=4,
|
| 68 |
+
intermediate_size=37,
|
| 69 |
+
dropout=0.1,
|
| 70 |
+
attention_dropout=0.1,
|
| 71 |
+
initializer_range=1e-10,
|
| 72 |
+
scope=None,
|
| 73 |
+
):
|
| 74 |
+
self.parent = parent
|
| 75 |
+
self.batch_size = batch_size
|
| 76 |
+
self.image_size = image_size
|
| 77 |
+
self.patch_size = patch_size
|
| 78 |
+
self.num_channels = num_channels
|
| 79 |
+
self.is_training = is_training
|
| 80 |
+
self.hidden_size = hidden_size
|
| 81 |
+
self.projection_dim = projection_dim
|
| 82 |
+
self.num_hidden_layers = num_hidden_layers
|
| 83 |
+
self.num_attention_heads = num_attention_heads
|
| 84 |
+
self.intermediate_size = intermediate_size
|
| 85 |
+
self.dropout = dropout
|
| 86 |
+
self.attention_dropout = attention_dropout
|
| 87 |
+
self.initializer_range = initializer_range
|
| 88 |
+
self.scope = scope
|
| 89 |
+
|
| 90 |
+
# in ViT, the seq length equals the number of patches + 1 (we add 1 for the [CLS] token)
|
| 91 |
+
num_patches = (image_size // patch_size) ** 2
|
| 92 |
+
self.seq_length = num_patches + 1
|
| 93 |
+
|
| 94 |
+
def prepare_config_and_inputs(self):
|
| 95 |
+
pixel_values = floats_tensor([self.batch_size, self.num_channels, self.image_size, self.image_size])
|
| 96 |
+
config = self.get_config()
|
| 97 |
+
|
| 98 |
+
return config, pixel_values
|
| 99 |
+
|
| 100 |
+
def get_config(self):
|
| 101 |
+
return BlipVisionConfig(
|
| 102 |
+
image_size=self.image_size,
|
| 103 |
+
patch_size=self.patch_size,
|
| 104 |
+
num_channels=self.num_channels,
|
| 105 |
+
hidden_size=self.hidden_size,
|
| 106 |
+
projection_dim=self.projection_dim,
|
| 107 |
+
num_hidden_layers=self.num_hidden_layers,
|
| 108 |
+
num_attention_heads=self.num_attention_heads,
|
| 109 |
+
intermediate_size=self.intermediate_size,
|
| 110 |
+
dropout=self.dropout,
|
| 111 |
+
attention_dropout=self.attention_dropout,
|
| 112 |
+
initializer_range=self.initializer_range,
|
| 113 |
+
)
|
| 114 |
+
|
| 115 |
+
def create_and_check_model(self, config, pixel_values):
|
| 116 |
+
model = TFBlipVisionModel(config=config)
|
| 117 |
+
result = model(pixel_values)
|
| 118 |
+
# expected sequence length = num_patches + 1 (we add 1 for the [CLS] token)
|
| 119 |
+
image_size = (self.image_size, self.image_size)
|
| 120 |
+
patch_size = (self.patch_size, self.patch_size)
|
| 121 |
+
num_patches = (image_size[1] // patch_size[1]) * (image_size[0] // patch_size[0])
|
| 122 |
+
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, num_patches + 1, self.hidden_size))
|
| 123 |
+
self.parent.assertEqual(result.pooler_output.shape, (self.batch_size, self.hidden_size))
|
| 124 |
+
|
| 125 |
+
def prepare_config_and_inputs_for_common(self):
|
| 126 |
+
config_and_inputs = self.prepare_config_and_inputs()
|
| 127 |
+
config, pixel_values = config_and_inputs
|
| 128 |
+
inputs_dict = {"pixel_values": pixel_values}
|
| 129 |
+
return config, inputs_dict
|
| 130 |
+
|
| 131 |
+
|
| 132 |
+
@require_tf
|
| 133 |
+
class TFBlipVisionModelTest(TFModelTesterMixin, unittest.TestCase):
|
| 134 |
+
"""
|
| 135 |
+
Here we also overwrite some of the tests of test_modeling_common.py, as Blip does not use input_ids, inputs_embeds,
|
| 136 |
+
attention_mask and seq_length.
|
| 137 |
+
"""
|
| 138 |
+
|
| 139 |
+
all_model_classes = (TFBlipVisionModel,) if is_tf_available() else ()
|
| 140 |
+
fx_compatible = False
|
| 141 |
+
test_pruning = False
|
| 142 |
+
test_resize_embeddings = False
|
| 143 |
+
test_head_masking = False
|
| 144 |
+
test_onnx = False
|
| 145 |
+
|
| 146 |
+
def setUp(self):
|
| 147 |
+
self.model_tester = TFBlipVisionModelTester(self)
|
| 148 |
+
self.config_tester = ConfigTester(self, config_class=BlipVisionConfig, has_text_modality=False, hidden_size=37)
|
| 149 |
+
|
| 150 |
+
def test_config(self):
|
| 151 |
+
self.config_tester.run_common_tests()
|
| 152 |
+
|
| 153 |
+
@unittest.skip(reason="Blip does not use inputs_embeds")
|
| 154 |
+
def test_inputs_embeds(self):
|
| 155 |
+
pass
|
| 156 |
+
|
| 157 |
+
def test_forward_signature(self):
|
| 158 |
+
config, _ = self.model_tester.prepare_config_and_inputs_for_common()
|
| 159 |
+
|
| 160 |
+
for model_class in self.all_model_classes:
|
| 161 |
+
model = model_class(config)
|
| 162 |
+
signature = inspect.signature(model.call)
|
| 163 |
+
# signature.parameters is an OrderedDict => so arg_names order is deterministic
|
| 164 |
+
arg_names = [*signature.parameters.keys()]
|
| 165 |
+
|
| 166 |
+
expected_arg_names = ["pixel_values"]
|
| 167 |
+
self.assertListEqual(arg_names[:1], expected_arg_names)
|
| 168 |
+
|
| 169 |
+
def test_model_common_attributes(self):
|
| 170 |
+
config, _ = self.model_tester.prepare_config_and_inputs_for_common()
|
| 171 |
+
|
| 172 |
+
for model_class in self.all_model_classes:
|
| 173 |
+
model = model_class(config)
|
| 174 |
+
self.assertIsInstance(model.get_input_embeddings(), (keras.layers.Layer))
|
| 175 |
+
x = model.get_output_embeddings()
|
| 176 |
+
self.assertTrue(x is None or isinstance(x, keras.layers.Layer))
|
| 177 |
+
|
| 178 |
+
def test_model(self):
|
| 179 |
+
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
| 180 |
+
self.model_tester.create_and_check_model(*config_and_inputs)
|
| 181 |
+
|
| 182 |
+
@unittest.skip(reason="BlipVisionModel has no base class and is not available in MODEL_MAPPING")
|
| 183 |
+
def test_save_load_fast_init_from_base(self):
|
| 184 |
+
pass
|
| 185 |
+
|
| 186 |
+
@unittest.skip(reason="BlipVisionModel has no base class and is not available in MODEL_MAPPING")
|
| 187 |
+
def test_save_load_fast_init_to_base(self):
|
| 188 |
+
pass
|
| 189 |
+
|
| 190 |
+
@slow
|
| 191 |
+
def test_model_from_pretrained(self):
|
| 192 |
+
model_name = "Salesforce/blip-vqa-base"
|
| 193 |
+
model = TFBlipVisionModel.from_pretrained(model_name)
|
| 194 |
+
self.assertIsNotNone(model)
|
| 195 |
+
|
| 196 |
+
|
| 197 |
+
class TFBlipTextModelTester:
|
| 198 |
+
def __init__(
|
| 199 |
+
self,
|
| 200 |
+
parent,
|
| 201 |
+
batch_size=12,
|
| 202 |
+
seq_length=7,
|
| 203 |
+
is_training=True,
|
| 204 |
+
use_input_mask=True,
|
| 205 |
+
use_labels=True,
|
| 206 |
+
vocab_size=99,
|
| 207 |
+
hidden_size=32,
|
| 208 |
+
projection_dim=32,
|
| 209 |
+
num_hidden_layers=2,
|
| 210 |
+
num_attention_heads=4,
|
| 211 |
+
intermediate_size=37,
|
| 212 |
+
dropout=0.1,
|
| 213 |
+
attention_dropout=0.1,
|
| 214 |
+
max_position_embeddings=512,
|
| 215 |
+
initializer_range=0.02,
|
| 216 |
+
bos_token_id=0,
|
| 217 |
+
scope=None,
|
| 218 |
+
):
|
| 219 |
+
self.parent = parent
|
| 220 |
+
self.batch_size = batch_size
|
| 221 |
+
self.seq_length = seq_length
|
| 222 |
+
self.is_training = is_training
|
| 223 |
+
self.use_input_mask = use_input_mask
|
| 224 |
+
self.use_labels = use_labels
|
| 225 |
+
self.vocab_size = vocab_size
|
| 226 |
+
self.hidden_size = hidden_size
|
| 227 |
+
self.projection_dim = projection_dim
|
| 228 |
+
self.num_hidden_layers = num_hidden_layers
|
| 229 |
+
self.num_attention_heads = num_attention_heads
|
| 230 |
+
self.intermediate_size = intermediate_size
|
| 231 |
+
self.dropout = dropout
|
| 232 |
+
self.attention_dropout = attention_dropout
|
| 233 |
+
self.max_position_embeddings = max_position_embeddings
|
| 234 |
+
self.initializer_range = initializer_range
|
| 235 |
+
self.scope = scope
|
| 236 |
+
self.bos_token_id = bos_token_id
|
| 237 |
+
|
| 238 |
+
def prepare_config_and_inputs(self):
|
| 239 |
+
input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
|
| 240 |
+
|
| 241 |
+
input_mask = None
|
| 242 |
+
if self.use_input_mask:
|
| 243 |
+
input_mask = random_attention_mask([self.batch_size, self.seq_length])
|
| 244 |
+
|
| 245 |
+
if input_mask is not None:
|
| 246 |
+
input_mask = input_mask.numpy()
|
| 247 |
+
batch_size, seq_length = input_mask.shape
|
| 248 |
+
rnd_start_indices = np.random.randint(1, seq_length - 1, size=(batch_size,))
|
| 249 |
+
for batch_idx, start_index in enumerate(rnd_start_indices):
|
| 250 |
+
input_mask[batch_idx, :start_index] = 1
|
| 251 |
+
input_mask[batch_idx, start_index:] = 0
|
| 252 |
+
input_mask = tf.convert_to_tensor(input_mask)
|
| 253 |
+
|
| 254 |
+
config = self.get_config()
|
| 255 |
+
|
| 256 |
+
return config, input_ids, input_mask
|
| 257 |
+
|
| 258 |
+
def get_config(self):
|
| 259 |
+
return BlipTextConfig(
|
| 260 |
+
vocab_size=self.vocab_size,
|
| 261 |
+
hidden_size=self.hidden_size,
|
| 262 |
+
projection_dim=self.projection_dim,
|
| 263 |
+
num_hidden_layers=self.num_hidden_layers,
|
| 264 |
+
num_attention_heads=self.num_attention_heads,
|
| 265 |
+
intermediate_size=self.intermediate_size,
|
| 266 |
+
dropout=self.dropout,
|
| 267 |
+
attention_dropout=self.attention_dropout,
|
| 268 |
+
max_position_embeddings=self.max_position_embeddings,
|
| 269 |
+
initializer_range=self.initializer_range,
|
| 270 |
+
bos_token_id=self.bos_token_id,
|
| 271 |
+
)
|
| 272 |
+
|
| 273 |
+
def create_and_check_model(self, config, input_ids, input_mask):
|
| 274 |
+
model = TFBlipTextModel(config=config)
|
| 275 |
+
result = model(input_ids, attention_mask=input_mask, training=False)
|
| 276 |
+
result = model(input_ids, training=False)
|
| 277 |
+
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
|
| 278 |
+
self.parent.assertEqual(result.pooler_output.shape, (self.batch_size, self.hidden_size))
|
| 279 |
+
|
| 280 |
+
def prepare_config_and_inputs_for_common(self):
|
| 281 |
+
config_and_inputs = self.prepare_config_and_inputs()
|
| 282 |
+
config, input_ids, input_mask = config_and_inputs
|
| 283 |
+
inputs_dict = {"input_ids": input_ids, "attention_mask": input_mask}
|
| 284 |
+
return config, inputs_dict
|
| 285 |
+
|
| 286 |
+
|
| 287 |
+
@require_tf
|
| 288 |
+
class TFBlipTextModelTest(TFModelTesterMixin, unittest.TestCase):
|
| 289 |
+
all_model_classes = (TFBlipTextModel,) if is_tf_available() else ()
|
| 290 |
+
fx_compatible = False
|
| 291 |
+
test_pruning = False
|
| 292 |
+
test_head_masking = False
|
| 293 |
+
test_onnx = False
|
| 294 |
+
|
| 295 |
+
def setUp(self):
|
| 296 |
+
self.model_tester = TFBlipTextModelTester(self)
|
| 297 |
+
self.config_tester = ConfigTester(self, config_class=BlipTextConfig, hidden_size=37)
|
| 298 |
+
|
| 299 |
+
def test_config(self):
|
| 300 |
+
self.config_tester.run_common_tests()
|
| 301 |
+
|
| 302 |
+
def test_model(self):
|
| 303 |
+
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
| 304 |
+
self.model_tester.create_and_check_model(*config_and_inputs)
|
| 305 |
+
|
| 306 |
+
@unittest.skip(reason="Blip does not use inputs_embeds")
|
| 307 |
+
def test_inputs_embeds(self):
|
| 308 |
+
pass
|
| 309 |
+
|
| 310 |
+
@unittest.skip(reason="BlipTextModel has no base class and is not available in MODEL_MAPPING")
|
| 311 |
+
def test_save_load_fast_init_from_base(self):
|
| 312 |
+
pass
|
| 313 |
+
|
| 314 |
+
@unittest.skip(reason="BlipTextModel has no base class and is not available in MODEL_MAPPING")
|
| 315 |
+
def test_save_load_fast_init_to_base(self):
|
| 316 |
+
pass
|
| 317 |
+
|
| 318 |
+
@slow
|
| 319 |
+
def test_model_from_pretrained(self):
|
| 320 |
+
model_name = "Salesforce/blip-vqa-base"
|
| 321 |
+
model = TFBlipTextModel.from_pretrained(model_name)
|
| 322 |
+
self.assertIsNotNone(model)
|
| 323 |
+
|
| 324 |
+
def test_pt_tf_model_equivalence(self, allow_missing_keys=True):
|
| 325 |
+
super().test_pt_tf_model_equivalence(allow_missing_keys=allow_missing_keys)
|
| 326 |
+
|
| 327 |
+
|
| 328 |
+
class TFBlipModelTester:
|
| 329 |
+
def __init__(self, parent, text_kwargs=None, vision_kwargs=None, is_training=True):
|
| 330 |
+
if text_kwargs is None:
|
| 331 |
+
text_kwargs = {}
|
| 332 |
+
if vision_kwargs is None:
|
| 333 |
+
vision_kwargs = {}
|
| 334 |
+
|
| 335 |
+
self.parent = parent
|
| 336 |
+
self.text_model_tester = TFBlipTextModelTester(parent, **text_kwargs)
|
| 337 |
+
self.vision_model_tester = TFBlipVisionModelTester(parent, **vision_kwargs)
|
| 338 |
+
self.is_training = is_training
|
| 339 |
+
|
| 340 |
+
def prepare_config_and_inputs(self):
|
| 341 |
+
text_config, input_ids, attention_mask = self.text_model_tester.prepare_config_and_inputs()
|
| 342 |
+
vision_config, pixel_values = self.vision_model_tester.prepare_config_and_inputs()
|
| 343 |
+
|
| 344 |
+
config = self.get_config()
|
| 345 |
+
|
| 346 |
+
return config, input_ids, attention_mask, pixel_values
|
| 347 |
+
|
| 348 |
+
def get_config(self):
|
| 349 |
+
return BlipConfig.from_text_vision_configs(
|
| 350 |
+
self.text_model_tester.get_config(), self.vision_model_tester.get_config(), projection_dim=64
|
| 351 |
+
)
|
| 352 |
+
|
| 353 |
+
def create_and_check_model(self, config, input_ids, attention_mask, pixel_values):
|
| 354 |
+
model = TFBlipModel(config)
|
| 355 |
+
result = model(input_ids, pixel_values, attention_mask, training=False)
|
| 356 |
+
self.parent.assertEqual(
|
| 357 |
+
result.logits_per_image.shape, (self.vision_model_tester.batch_size, self.text_model_tester.batch_size)
|
| 358 |
+
)
|
| 359 |
+
self.parent.assertEqual(
|
| 360 |
+
result.logits_per_text.shape, (self.text_model_tester.batch_size, self.vision_model_tester.batch_size)
|
| 361 |
+
)
|
| 362 |
+
|
| 363 |
+
def prepare_config_and_inputs_for_common(self):
|
| 364 |
+
config_and_inputs = self.prepare_config_and_inputs()
|
| 365 |
+
config, input_ids, attention_mask, pixel_values = config_and_inputs
|
| 366 |
+
inputs_dict = {
|
| 367 |
+
"input_ids": input_ids,
|
| 368 |
+
"attention_mask": attention_mask,
|
| 369 |
+
"pixel_values": pixel_values,
|
| 370 |
+
"return_loss": True,
|
| 371 |
+
}
|
| 372 |
+
return config, inputs_dict
|
| 373 |
+
|
| 374 |
+
|
| 375 |
+
@require_tf
|
| 376 |
+
class TFBlipModelTest(TFModelTesterMixin, PipelineTesterMixin, unittest.TestCase):
|
| 377 |
+
all_model_classes = (TFBlipModel,) if is_tf_available() else ()
|
| 378 |
+
pipeline_model_mapping = (
|
| 379 |
+
{"feature-extraction": TFBlipModel, "image-to-text": TFBlipForConditionalGeneration}
|
| 380 |
+
if is_tf_available()
|
| 381 |
+
else {}
|
| 382 |
+
)
|
| 383 |
+
test_head_masking = False
|
| 384 |
+
test_pruning = False
|
| 385 |
+
test_resize_embeddings = False
|
| 386 |
+
test_attention_outputs = False
|
| 387 |
+
test_onnx = False
|
| 388 |
+
|
| 389 |
+
def setUp(self):
|
| 390 |
+
self.model_tester = TFBlipModelTester(self)
|
| 391 |
+
|
| 392 |
+
def test_model(self):
|
| 393 |
+
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
| 394 |
+
self.model_tester.create_and_check_model(*config_and_inputs)
|
| 395 |
+
|
| 396 |
+
@unittest.skip(reason="Hidden_states is tested in individual model tests")
|
| 397 |
+
def test_hidden_states_output(self):
|
| 398 |
+
pass
|
| 399 |
+
|
| 400 |
+
@unittest.skip(reason="Inputs_embeds is tested in individual model tests")
|
| 401 |
+
def test_inputs_embeds(self):
|
| 402 |
+
pass
|
| 403 |
+
|
| 404 |
+
@unittest.skip(reason="Retain_grad is tested in individual model tests")
|
| 405 |
+
def test_retain_grad_hidden_states_attentions(self):
|
| 406 |
+
pass
|
| 407 |
+
|
| 408 |
+
@unittest.skip(reason="BlipModel does not have input/output embeddings")
|
| 409 |
+
def test_model_common_attributes(self):
|
| 410 |
+
pass
|
| 411 |
+
|
| 412 |
+
def test_load_vision_text_config(self):
|
| 413 |
+
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
| 414 |
+
|
| 415 |
+
# Save BlipConfig and check if we can load BlipVisionConfig from it
|
| 416 |
+
with tempfile.TemporaryDirectory() as tmp_dir_name:
|
| 417 |
+
config.save_pretrained(tmp_dir_name)
|
| 418 |
+
vision_config = BlipVisionConfig.from_pretrained(tmp_dir_name)
|
| 419 |
+
self.assertDictEqual(config.vision_config.to_dict(), vision_config.to_dict())
|
| 420 |
+
|
| 421 |
+
# Save BlipConfig and check if we can load BlipTextConfig from it
|
| 422 |
+
with tempfile.TemporaryDirectory() as tmp_dir_name:
|
| 423 |
+
config.save_pretrained(tmp_dir_name)
|
| 424 |
+
text_config = BlipTextConfig.from_pretrained(tmp_dir_name)
|
| 425 |
+
self.assertDictEqual(config.text_config.to_dict(), text_config.to_dict())
|
| 426 |
+
|
| 427 |
+
@slow
|
| 428 |
+
def test_model_from_pretrained(self):
|
| 429 |
+
model_name = "Salesforce/blip-vqa-base"
|
| 430 |
+
model = TFBlipModel.from_pretrained(model_name)
|
| 431 |
+
self.assertIsNotNone(model)
|
| 432 |
+
|
| 433 |
+
def test_pt_tf_model_equivalence(self, allow_missing_keys=True):
|
| 434 |
+
super().test_pt_tf_model_equivalence(allow_missing_keys=allow_missing_keys)
|
| 435 |
+
|
| 436 |
+
@unittest.skip("Matt: Re-enable this test when we have a proper export function for TF models.")
|
| 437 |
+
def test_saved_model_creation(self):
|
| 438 |
+
# This fails because the if return_loss: conditional can return None or a Tensor and TF hates that.
|
| 439 |
+
# We could fix that by setting the bool to a constant when exporting, but that requires a dedicated export
|
| 440 |
+
# function that we don't have yet.
|
| 441 |
+
pass
|
| 442 |
+
|
| 443 |
+
|
| 444 |
+
class BlipTextRetrievalModelTester:
|
| 445 |
+
def __init__(self, parent, text_kwargs=None, vision_kwargs=None, is_training=True):
|
| 446 |
+
if text_kwargs is None:
|
| 447 |
+
text_kwargs = {}
|
| 448 |
+
if vision_kwargs is None:
|
| 449 |
+
vision_kwargs = {}
|
| 450 |
+
|
| 451 |
+
self.parent = parent
|
| 452 |
+
self.text_model_tester = TFBlipTextModelTester(parent, **text_kwargs)
|
| 453 |
+
self.vision_model_tester = TFBlipVisionModelTester(parent, **vision_kwargs)
|
| 454 |
+
self.is_training = is_training
|
| 455 |
+
|
| 456 |
+
def prepare_config_and_inputs(self):
|
| 457 |
+
text_config, input_ids, attention_mask = self.text_model_tester.prepare_config_and_inputs()
|
| 458 |
+
vision_config, pixel_values = self.vision_model_tester.prepare_config_and_inputs()
|
| 459 |
+
|
| 460 |
+
config = self.get_config()
|
| 461 |
+
|
| 462 |
+
return config, input_ids, attention_mask, pixel_values
|
| 463 |
+
|
| 464 |
+
def get_config(self):
|
| 465 |
+
return BlipConfig.from_text_vision_configs(
|
| 466 |
+
self.text_model_tester.get_config(), self.vision_model_tester.get_config(), projection_dim=64
|
| 467 |
+
)
|
| 468 |
+
|
| 469 |
+
def create_and_check_model(self, config, input_ids, attention_mask, pixel_values):
|
| 470 |
+
model = TFBlipModel(config)
|
| 471 |
+
result = model(input_ids, pixel_values, attention_mask, training=False)
|
| 472 |
+
self.parent.assertEqual(
|
| 473 |
+
result.logits_per_image.shape, (self.vision_model_tester.batch_size, self.text_model_tester.batch_size)
|
| 474 |
+
)
|
| 475 |
+
self.parent.assertEqual(
|
| 476 |
+
result.logits_per_text.shape, (self.text_model_tester.batch_size, self.vision_model_tester.batch_size)
|
| 477 |
+
)
|
| 478 |
+
|
| 479 |
+
def prepare_config_and_inputs_for_common(self):
|
| 480 |
+
config_and_inputs = self.prepare_config_and_inputs()
|
| 481 |
+
config, input_ids, attention_mask, pixel_values = config_and_inputs
|
| 482 |
+
inputs_dict = {
|
| 483 |
+
"input_ids": input_ids,
|
| 484 |
+
"attention_mask": attention_mask,
|
| 485 |
+
"pixel_values": pixel_values,
|
| 486 |
+
}
|
| 487 |
+
return config, inputs_dict
|
| 488 |
+
|
| 489 |
+
|
| 490 |
+
class BlipTextImageModelsModelTester:
|
| 491 |
+
def __init__(self, parent, text_kwargs=None, vision_kwargs=None, is_training=True):
|
| 492 |
+
if text_kwargs is None:
|
| 493 |
+
text_kwargs = {}
|
| 494 |
+
if vision_kwargs is None:
|
| 495 |
+
vision_kwargs = {}
|
| 496 |
+
|
| 497 |
+
self.parent = parent
|
| 498 |
+
self.text_model_tester = TFBlipTextModelTester(parent, **text_kwargs)
|
| 499 |
+
self.vision_model_tester = TFBlipVisionModelTester(parent, **vision_kwargs)
|
| 500 |
+
self.is_training = is_training
|
| 501 |
+
|
| 502 |
+
def prepare_config_and_inputs(self):
|
| 503 |
+
text_config, input_ids, attention_mask = self.text_model_tester.prepare_config_and_inputs()
|
| 504 |
+
vision_config, pixel_values = self.vision_model_tester.prepare_config_and_inputs()
|
| 505 |
+
|
| 506 |
+
config = self.get_config()
|
| 507 |
+
|
| 508 |
+
return config, input_ids, attention_mask, pixel_values
|
| 509 |
+
|
| 510 |
+
def get_config(self):
|
| 511 |
+
return BlipConfig.from_text_vision_configs(
|
| 512 |
+
self.text_model_tester.get_config(), self.vision_model_tester.get_config(), projection_dim=64
|
| 513 |
+
)
|
| 514 |
+
|
| 515 |
+
def create_and_check_model(self, config, input_ids, attention_mask, pixel_values):
|
| 516 |
+
model = TFBlipModel(config)
|
| 517 |
+
result = model(input_ids, pixel_values, attention_mask, training=False)
|
| 518 |
+
self.parent.assertEqual(
|
| 519 |
+
result.logits_per_image.shape, (self.vision_model_tester.batch_size, self.text_model_tester.batch_size)
|
| 520 |
+
)
|
| 521 |
+
self.parent.assertEqual(
|
| 522 |
+
result.logits_per_text.shape, (self.text_model_tester.batch_size, self.vision_model_tester.batch_size)
|
| 523 |
+
)
|
| 524 |
+
|
| 525 |
+
def prepare_config_and_inputs_for_common(self):
|
| 526 |
+
config_and_inputs = self.prepare_config_and_inputs()
|
| 527 |
+
config, input_ids, attention_mask, pixel_values = config_and_inputs
|
| 528 |
+
inputs_dict = {
|
| 529 |
+
"input_ids": input_ids,
|
| 530 |
+
"labels": input_ids,
|
| 531 |
+
"attention_mask": attention_mask,
|
| 532 |
+
"pixel_values": pixel_values,
|
| 533 |
+
}
|
| 534 |
+
return config, inputs_dict
|
| 535 |
+
|
| 536 |
+
|
| 537 |
+
class BlipVQAModelsModelTester:
|
| 538 |
+
def __init__(self, parent, text_kwargs=None, vision_kwargs=None, is_training=True):
|
| 539 |
+
if text_kwargs is None:
|
| 540 |
+
text_kwargs = {}
|
| 541 |
+
if vision_kwargs is None:
|
| 542 |
+
vision_kwargs = {}
|
| 543 |
+
|
| 544 |
+
self.parent = parent
|
| 545 |
+
self.text_model_tester = TFBlipTextModelTester(parent, **text_kwargs)
|
| 546 |
+
self.vision_model_tester = TFBlipVisionModelTester(parent, **vision_kwargs)
|
| 547 |
+
self.is_training = is_training
|
| 548 |
+
|
| 549 |
+
def prepare_config_and_inputs(self):
|
| 550 |
+
text_config, input_ids, attention_mask = self.text_model_tester.prepare_config_and_inputs()
|
| 551 |
+
vision_config, pixel_values = self.vision_model_tester.prepare_config_and_inputs()
|
| 552 |
+
|
| 553 |
+
config = self.get_config()
|
| 554 |
+
|
| 555 |
+
return config, input_ids, attention_mask, pixel_values
|
| 556 |
+
|
| 557 |
+
def get_config(self):
|
| 558 |
+
return BlipConfig.from_text_vision_configs(
|
| 559 |
+
self.text_model_tester.get_config(), self.vision_model_tester.get_config(), projection_dim=64
|
| 560 |
+
)
|
| 561 |
+
|
| 562 |
+
def create_and_check_model(self, config, input_ids, attention_mask, pixel_values):
|
| 563 |
+
model = TFBlipModel(config)
|
| 564 |
+
result = model(input_ids, pixel_values, attention_mask, training=False)
|
| 565 |
+
self.parent.assertEqual(
|
| 566 |
+
result.logits_per_image.shape, (self.vision_model_tester.batch_size, self.text_model_tester.batch_size)
|
| 567 |
+
)
|
| 568 |
+
self.parent.assertEqual(
|
| 569 |
+
result.logits_per_text.shape, (self.text_model_tester.batch_size, self.vision_model_tester.batch_size)
|
| 570 |
+
)
|
| 571 |
+
|
| 572 |
+
def prepare_config_and_inputs_for_common(self):
|
| 573 |
+
config_and_inputs = self.prepare_config_and_inputs()
|
| 574 |
+
config, input_ids, attention_mask, pixel_values = config_and_inputs
|
| 575 |
+
inputs_dict = {
|
| 576 |
+
"input_ids": input_ids,
|
| 577 |
+
"decoder_input_ids": input_ids,
|
| 578 |
+
"labels": input_ids,
|
| 579 |
+
"attention_mask": attention_mask,
|
| 580 |
+
"pixel_values": pixel_values,
|
| 581 |
+
}
|
| 582 |
+
return config, inputs_dict
|
| 583 |
+
|
| 584 |
+
|
| 585 |
+
@require_tf
|
| 586 |
+
@require_vision
|
| 587 |
+
class TFBlipVQAModelTest(TFModelTesterMixin, unittest.TestCase):
|
| 588 |
+
all_model_classes = (TFBlipForQuestionAnswering,) if is_tf_available() else ()
|
| 589 |
+
test_head_masking = False
|
| 590 |
+
test_pruning = False
|
| 591 |
+
test_resize_embeddings = False
|
| 592 |
+
test_attention_outputs = False
|
| 593 |
+
test_onnx = False
|
| 594 |
+
|
| 595 |
+
def setUp(self):
|
| 596 |
+
self.model_tester = BlipVQAModelsModelTester(self)
|
| 597 |
+
|
| 598 |
+
def _prepare_inputs_for_vqa(self):
|
| 599 |
+
_, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
| 600 |
+
inputs_dict["labels"] = inputs_dict["input_ids"]
|
| 601 |
+
inputs_dict["decoder_input_ids"] = inputs_dict["input_ids"]
|
| 602 |
+
inputs_dict.pop("return_loss")
|
| 603 |
+
return inputs_dict
|
| 604 |
+
|
| 605 |
+
def test_class_name_consistency(self):
|
| 606 |
+
"""
|
| 607 |
+
Tests that all VQA models have a class name that ends with "ForQuestionAnswering"
|
| 608 |
+
"""
|
| 609 |
+
for model_class in self.all_model_classes:
|
| 610 |
+
model = model_class(self.model_tester.get_config())
|
| 611 |
+
self.assertTrue(
|
| 612 |
+
model.__class__.__name__.endswith("ForQuestionAnswering"),
|
| 613 |
+
f"Class name should end with 'ForVisualQuestionAnswering' got {model.__class__.__name__}",
|
| 614 |
+
)
|
| 615 |
+
|
| 616 |
+
def test_training(self):
|
| 617 |
+
"""
|
| 618 |
+
Tests that all VQA models can be trained on a single batch
|
| 619 |
+
"""
|
| 620 |
+
for model_class in self.all_model_classes:
|
| 621 |
+
model = model_class(self.model_tester.get_config())
|
| 622 |
+
loss = model(**self.model_tester.prepare_config_and_inputs_for_common()[1], training=True).loss
|
| 623 |
+
|
| 624 |
+
self.assertIsNotNone(loss, "Loss should not be None")
|
| 625 |
+
|
| 626 |
+
@unittest.skip(reason="Hidden_states is tested in individual model tests")
|
| 627 |
+
def test_hidden_states_output(self):
|
| 628 |
+
pass
|
| 629 |
+
|
| 630 |
+
@unittest.skip(reason="Inputs_embeds is tested in individual model tests")
|
| 631 |
+
def test_inputs_embeds(self):
|
| 632 |
+
pass
|
| 633 |
+
|
| 634 |
+
@unittest.skip(reason="Retain_grad is tested in individual model tests")
|
| 635 |
+
def test_retain_grad_hidden_states_attentions(self):
|
| 636 |
+
pass
|
| 637 |
+
|
| 638 |
+
@unittest.skip(reason="BlipModel does not have input/output embeddings")
|
| 639 |
+
def test_model_common_attributes(self):
|
| 640 |
+
pass
|
| 641 |
+
|
| 642 |
+
@unittest.skip(reason="Tested in individual model tests")
|
| 643 |
+
def test_compile_tf_model(self):
|
| 644 |
+
pass
|
| 645 |
+
|
| 646 |
+
@unittest.skip("Model doesn't have a clean loss output.")
|
| 647 |
+
def test_keras_fit(self):
|
| 648 |
+
pass
|
| 649 |
+
|
| 650 |
+
|
| 651 |
+
@require_tf
|
| 652 |
+
class TFBlipTextRetrievalModelTest(TFModelTesterMixin, unittest.TestCase):
|
| 653 |
+
all_model_classes = (TFBlipForImageTextRetrieval,) if is_tf_available() else ()
|
| 654 |
+
test_head_masking = False
|
| 655 |
+
test_pruning = False
|
| 656 |
+
test_resize_embeddings = False
|
| 657 |
+
test_attention_outputs = False
|
| 658 |
+
test_onnx = False
|
| 659 |
+
|
| 660 |
+
def setUp(self):
|
| 661 |
+
self.model_tester = BlipTextRetrievalModelTester(self)
|
| 662 |
+
|
| 663 |
+
def test_model(self):
|
| 664 |
+
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
| 665 |
+
self.model_tester.create_and_check_model(*config_and_inputs)
|
| 666 |
+
|
| 667 |
+
@unittest.skip(reason="Hidden_states is tested in individual model tests")
|
| 668 |
+
def test_hidden_states_output(self):
|
| 669 |
+
pass
|
| 670 |
+
|
| 671 |
+
@unittest.skip(reason="Inputs_embeds is tested in individual model tests")
|
| 672 |
+
def test_inputs_embeds(self):
|
| 673 |
+
pass
|
| 674 |
+
|
| 675 |
+
@unittest.skip(reason="Retain_grad is tested in individual model tests")
|
| 676 |
+
def test_retain_grad_hidden_states_attentions(self):
|
| 677 |
+
pass
|
| 678 |
+
|
| 679 |
+
@unittest.skip(reason="BlipModel does not have input/output embeddings")
|
| 680 |
+
def test_model_common_attributes(self):
|
| 681 |
+
pass
|
| 682 |
+
|
| 683 |
+
def test_training(self):
|
| 684 |
+
if not self.model_tester.is_training:
|
| 685 |
+
return
|
| 686 |
+
|
| 687 |
+
for model_class in self.all_model_classes[:-1]:
|
| 688 |
+
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
| 689 |
+
config.return_dict = True
|
| 690 |
+
|
| 691 |
+
model = model_class(config)
|
| 692 |
+
inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
|
| 693 |
+
|
| 694 |
+
# hardcode labels to be the same as input_ids
|
| 695 |
+
inputs["labels"] = inputs["input_ids"]
|
| 696 |
+
|
| 697 |
+
loss = model(**inputs, training=True).loss
|
| 698 |
+
self.assertTrue(loss is not None)
|
| 699 |
+
|
| 700 |
+
def test_load_vision_text_config(self):
|
| 701 |
+
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
| 702 |
+
|
| 703 |
+
# Save BlipConfig and check if we can load BlipVisionConfig from it
|
| 704 |
+
with tempfile.TemporaryDirectory() as tmp_dir_name:
|
| 705 |
+
config.save_pretrained(tmp_dir_name)
|
| 706 |
+
vision_config = BlipVisionConfig.from_pretrained(tmp_dir_name)
|
| 707 |
+
self.assertDictEqual(config.vision_config.to_dict(), vision_config.to_dict())
|
| 708 |
+
|
| 709 |
+
# Save BlipConfig and check if we can load BlipTextConfig from it
|
| 710 |
+
with tempfile.TemporaryDirectory() as tmp_dir_name:
|
| 711 |
+
config.save_pretrained(tmp_dir_name)
|
| 712 |
+
text_config = BlipTextConfig.from_pretrained(tmp_dir_name)
|
| 713 |
+
self.assertDictEqual(config.text_config.to_dict(), text_config.to_dict())
|
| 714 |
+
|
| 715 |
+
@slow
|
| 716 |
+
def test_model_from_pretrained(self):
|
| 717 |
+
model_name = "Salesforce/blip-vqa-base"
|
| 718 |
+
model = TFBlipModel.from_pretrained(model_name)
|
| 719 |
+
self.assertIsNotNone(model)
|
| 720 |
+
|
| 721 |
+
@unittest.skip(reason="Tested in individual model tests")
|
| 722 |
+
def test_compile_tf_model(self):
|
| 723 |
+
pass
|
| 724 |
+
|
| 725 |
+
@unittest.skip("Model doesn't have a clean loss output.")
|
| 726 |
+
def test_keras_fit(self):
|
| 727 |
+
pass
|
| 728 |
+
|
| 729 |
+
|
| 730 |
+
@require_tf
|
| 731 |
+
class TFBlipTextImageModelTest(TFModelTesterMixin, unittest.TestCase):
|
| 732 |
+
all_model_classes = (TFBlipForConditionalGeneration,) if is_tf_available() else ()
|
| 733 |
+
test_head_masking = False
|
| 734 |
+
test_pruning = False
|
| 735 |
+
test_resize_embeddings = False
|
| 736 |
+
test_attention_outputs = False
|
| 737 |
+
test_onnx = False
|
| 738 |
+
|
| 739 |
+
def setUp(self):
|
| 740 |
+
self.model_tester = BlipTextImageModelsModelTester(self)
|
| 741 |
+
|
| 742 |
+
def test_model(self):
|
| 743 |
+
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
| 744 |
+
self.model_tester.create_and_check_model(*config_and_inputs)
|
| 745 |
+
|
| 746 |
+
@unittest.skip(reason="Hidden_states is tested in individual model tests")
|
| 747 |
+
def test_hidden_states_output(self):
|
| 748 |
+
pass
|
| 749 |
+
|
| 750 |
+
@unittest.skip(reason="Inputs_embeds is tested in individual model tests")
|
| 751 |
+
def test_inputs_embeds(self):
|
| 752 |
+
pass
|
| 753 |
+
|
| 754 |
+
def test_forward_signature(self):
|
| 755 |
+
config, _ = self.model_tester.prepare_config_and_inputs_for_common()
|
| 756 |
+
|
| 757 |
+
for model_class in self.all_model_classes:
|
| 758 |
+
model = model_class(config)
|
| 759 |
+
signature = inspect.signature(model.call)
|
| 760 |
+
# signature.parameters is an OrderedDict => so arg_names order is deterministic
|
| 761 |
+
arg_names = [*signature.parameters.keys()]
|
| 762 |
+
|
| 763 |
+
if model.config.is_encoder_decoder:
|
| 764 |
+
expected_arg_names = [
|
| 765 |
+
"input_ids",
|
| 766 |
+
"attention_mask",
|
| 767 |
+
"decoder_input_ids",
|
| 768 |
+
"decoder_attention_mask",
|
| 769 |
+
]
|
| 770 |
+
expected_arg_names.extend(
|
| 771 |
+
["head_mask", "decoder_head_mask", "cross_attn_head_mask", "encoder_outputs"]
|
| 772 |
+
if "head_mask" and "decoder_head_mask" and "cross_attn_head_mask" in arg_names
|
| 773 |
+
else ["encoder_outputs"]
|
| 774 |
+
)
|
| 775 |
+
self.assertListEqual(arg_names[: len(expected_arg_names)], expected_arg_names)
|
| 776 |
+
else:
|
| 777 |
+
expected_arg_names = (
|
| 778 |
+
["input_ids"] if model_class != TFBlipForConditionalGeneration else ["pixel_values"]
|
| 779 |
+
)
|
| 780 |
+
self.assertListEqual(arg_names[:1], expected_arg_names)
|
| 781 |
+
|
| 782 |
+
@unittest.skip(reason="Tested in individual model tests")
|
| 783 |
+
def test_compile_tf_model(self):
|
| 784 |
+
pass
|
| 785 |
+
|
| 786 |
+
@unittest.skip("Has some odd input names!")
|
| 787 |
+
def test_keras_fit(self):
|
| 788 |
+
pass
|
| 789 |
+
|
| 790 |
+
@unittest.skip(reason="Retain_grad is tested in individual model tests")
|
| 791 |
+
def test_retain_grad_hidden_states_attentions(self):
|
| 792 |
+
pass
|
| 793 |
+
|
| 794 |
+
@unittest.skip(reason="BlipModel does not have input/output embeddings")
|
| 795 |
+
def test_model_common_attributes(self):
|
| 796 |
+
pass
|
| 797 |
+
|
| 798 |
+
def test_training(self):
|
| 799 |
+
if not self.model_tester.is_training:
|
| 800 |
+
return
|
| 801 |
+
|
| 802 |
+
for model_class in self.all_model_classes[:-1]:
|
| 803 |
+
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
| 804 |
+
config.return_dict = True
|
| 805 |
+
|
| 806 |
+
model = model_class(config)
|
| 807 |
+
inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
|
| 808 |
+
|
| 809 |
+
# hardcode labels to be the same as input_ids
|
| 810 |
+
inputs["labels"] = inputs["input_ids"]
|
| 811 |
+
|
| 812 |
+
loss = model(**inputs, training=True).loss
|
| 813 |
+
self.assertIsNotNone(loss)
|
| 814 |
+
|
| 815 |
+
def test_load_vision_text_config(self):
|
| 816 |
+
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
| 817 |
+
|
| 818 |
+
# Save BlipConfig and check if we can load BlipVisionConfig from it
|
| 819 |
+
with tempfile.TemporaryDirectory() as tmp_dir_name:
|
| 820 |
+
config.save_pretrained(tmp_dir_name)
|
| 821 |
+
vision_config = BlipVisionConfig.from_pretrained(tmp_dir_name)
|
| 822 |
+
self.assertDictEqual(config.vision_config.to_dict(), vision_config.to_dict())
|
| 823 |
+
|
| 824 |
+
# Save BlipConfig and check if we can load BlipTextConfig from it
|
| 825 |
+
with tempfile.TemporaryDirectory() as tmp_dir_name:
|
| 826 |
+
config.save_pretrained(tmp_dir_name)
|
| 827 |
+
text_config = BlipTextConfig.from_pretrained(tmp_dir_name)
|
| 828 |
+
self.assertDictEqual(config.text_config.to_dict(), text_config.to_dict())
|
| 829 |
+
|
| 830 |
+
@slow
|
| 831 |
+
def test_model_from_pretrained(self):
|
| 832 |
+
model_name = "Salesforce/blip-vqa-base"
|
| 833 |
+
model = TFBlipModel.from_pretrained(model_name)
|
| 834 |
+
self.assertIsNotNone(model)
|
| 835 |
+
|
| 836 |
+
|
| 837 |
+
# We will verify our results on an image of cute cats
|
| 838 |
+
def prepare_img():
|
| 839 |
+
url = "https://huggingface.co/hf-internal-testing/blip-test-image/resolve/main/demo.jpg"
|
| 840 |
+
im = Image.open(requests.get(url, stream=True).raw)
|
| 841 |
+
return im
|
| 842 |
+
|
| 843 |
+
|
| 844 |
+
@require_vision
|
| 845 |
+
@require_tf
|
| 846 |
+
@slow
|
| 847 |
+
class TFBlipModelIntegrationTest(unittest.TestCase):
|
| 848 |
+
def test_inference_image_captioning(self):
|
| 849 |
+
model = TFBlipForConditionalGeneration.from_pretrained("Salesforce/blip-image-captioning-base")
|
| 850 |
+
processor = BlipProcessor.from_pretrained("Salesforce/blip-image-captioning-base")
|
| 851 |
+
image = prepare_img()
|
| 852 |
+
|
| 853 |
+
# image only
|
| 854 |
+
inputs = processor(images=image, return_tensors="tf")
|
| 855 |
+
|
| 856 |
+
predictions = model.generate(**inputs)
|
| 857 |
+
|
| 858 |
+
# Test output
|
| 859 |
+
self.assertEqual(
|
| 860 |
+
predictions[0].numpy().tolist(), [30522, 1037, 2450, 3564, 2006, 1996, 3509, 2007, 2014, 3899, 102]
|
| 861 |
+
)
|
| 862 |
+
|
| 863 |
+
# image and context
|
| 864 |
+
context = ["a picture of"]
|
| 865 |
+
inputs = processor(images=image, text=context, return_tensors="tf")
|
| 866 |
+
|
| 867 |
+
predictions = model.generate(**inputs)
|
| 868 |
+
|
| 869 |
+
# Test output
|
| 870 |
+
self.assertEqual(
|
| 871 |
+
predictions[0].numpy().tolist(),
|
| 872 |
+
[30522, 1037, 3861, 1997, 1037, 2450, 1998, 2014, 3899, 2006, 1996, 3509, 102],
|
| 873 |
+
)
|
| 874 |
+
|
| 875 |
+
def test_inference_vqa(self):
|
| 876 |
+
model = TFBlipForQuestionAnswering.from_pretrained("Salesforce/blip-vqa-base")
|
| 877 |
+
processor = BlipProcessor.from_pretrained("Salesforce/blip-vqa-base")
|
| 878 |
+
|
| 879 |
+
image = prepare_img()
|
| 880 |
+
text = "how many dogs are in the picture?"
|
| 881 |
+
inputs = processor(image, text=text, return_tensors="tf")
|
| 882 |
+
out = model.generate(**inputs)
|
| 883 |
+
|
| 884 |
+
# Test output
|
| 885 |
+
self.assertEqual(out[0].numpy().tolist(), [30522, 1015, 102])
|
| 886 |
+
|
| 887 |
+
def test_inference_itm(self):
|
| 888 |
+
model = TFBlipForImageTextRetrieval.from_pretrained("Salesforce/blip-itm-base-coco")
|
| 889 |
+
processor = BlipProcessor.from_pretrained("Salesforce/blip-itm-base-coco")
|
| 890 |
+
|
| 891 |
+
image = prepare_img()
|
| 892 |
+
text = "A woman and her dog sitting in a beach"
|
| 893 |
+
|
| 894 |
+
inputs = processor(image, text, return_tensors="tf")
|
| 895 |
+
|
| 896 |
+
out_itm = model(**inputs)
|
| 897 |
+
out = model(**inputs, use_itm_head=False, training=False)
|
| 898 |
+
|
| 899 |
+
expected_scores = tf.convert_to_tensor([[0.0029, 0.9971]])
|
| 900 |
+
self.assertTrue(np.allclose(tf.nn.softmax(out_itm[0]).numpy(), expected_scores, rtol=1e-3, atol=1e-3))
|
| 901 |
+
self.assertTrue(np.allclose(out[0], tf.convert_to_tensor([[0.5162]]), rtol=1e-3, atol=1e-3))
|
tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/blip/test_modeling_tf_blip_text.py
ADDED
|
@@ -0,0 +1,181 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# coding=utf-8
|
| 2 |
+
# Copyright 2023 The HuggingFace Inc. team. All rights reserved.
|
| 3 |
+
#
|
| 4 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
| 5 |
+
# you may not use this file except in compliance with the License.
|
| 6 |
+
# You may obtain a copy of the License at
|
| 7 |
+
#
|
| 8 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
| 9 |
+
#
|
| 10 |
+
# Unless required by applicable law or agreed to in writing, software
|
| 11 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
| 12 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
| 13 |
+
# See the License for the specific language governing permissions and
|
| 14 |
+
# limitations under the License.
|
| 15 |
+
"""Testing suite for the TensorFlow Blip model."""
|
| 16 |
+
|
| 17 |
+
from __future__ import annotations
|
| 18 |
+
|
| 19 |
+
import unittest
|
| 20 |
+
|
| 21 |
+
import numpy as np
|
| 22 |
+
|
| 23 |
+
from transformers import BlipTextConfig
|
| 24 |
+
from transformers.testing_utils import require_tf, slow
|
| 25 |
+
from transformers.utils import is_tf_available
|
| 26 |
+
|
| 27 |
+
from ...test_configuration_common import ConfigTester
|
| 28 |
+
from ...test_modeling_tf_common import TFModelTesterMixin, ids_tensor, random_attention_mask
|
| 29 |
+
|
| 30 |
+
|
| 31 |
+
if is_tf_available():
|
| 32 |
+
import tensorflow as tf
|
| 33 |
+
|
| 34 |
+
from transformers import TFBlipTextModel
|
| 35 |
+
|
| 36 |
+
|
| 37 |
+
class BlipTextModelTester:
|
| 38 |
+
def __init__(
|
| 39 |
+
self,
|
| 40 |
+
parent,
|
| 41 |
+
batch_size=12,
|
| 42 |
+
seq_length=7,
|
| 43 |
+
is_training=True,
|
| 44 |
+
use_input_mask=True,
|
| 45 |
+
use_labels=True,
|
| 46 |
+
vocab_size=99,
|
| 47 |
+
hidden_size=32,
|
| 48 |
+
projection_dim=32,
|
| 49 |
+
num_hidden_layers=2,
|
| 50 |
+
num_attention_heads=4,
|
| 51 |
+
intermediate_size=37,
|
| 52 |
+
dropout=0.1,
|
| 53 |
+
attention_dropout=0.1,
|
| 54 |
+
max_position_embeddings=512,
|
| 55 |
+
initializer_range=0.02,
|
| 56 |
+
bos_token_id=0,
|
| 57 |
+
scope=None,
|
| 58 |
+
):
|
| 59 |
+
self.parent = parent
|
| 60 |
+
self.batch_size = batch_size
|
| 61 |
+
self.seq_length = seq_length
|
| 62 |
+
self.is_training = is_training
|
| 63 |
+
self.use_input_mask = use_input_mask
|
| 64 |
+
self.use_labels = use_labels
|
| 65 |
+
self.vocab_size = vocab_size
|
| 66 |
+
self.hidden_size = hidden_size
|
| 67 |
+
self.projection_dim = projection_dim
|
| 68 |
+
self.num_hidden_layers = num_hidden_layers
|
| 69 |
+
self.num_attention_heads = num_attention_heads
|
| 70 |
+
self.intermediate_size = intermediate_size
|
| 71 |
+
self.dropout = dropout
|
| 72 |
+
self.attention_dropout = attention_dropout
|
| 73 |
+
self.max_position_embeddings = max_position_embeddings
|
| 74 |
+
self.initializer_range = initializer_range
|
| 75 |
+
self.scope = scope
|
| 76 |
+
self.bos_token_id = bos_token_id
|
| 77 |
+
|
| 78 |
+
def prepare_config_and_inputs(self):
|
| 79 |
+
input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
|
| 80 |
+
|
| 81 |
+
input_mask = None
|
| 82 |
+
if self.use_input_mask:
|
| 83 |
+
input_mask = random_attention_mask([self.batch_size, self.seq_length])
|
| 84 |
+
|
| 85 |
+
if input_mask is not None:
|
| 86 |
+
input_mask = input_mask.numpy()
|
| 87 |
+
batch_size, seq_length = input_mask.shape
|
| 88 |
+
rnd_start_indices = np.random.randint(1, seq_length - 1, size=(batch_size,))
|
| 89 |
+
for batch_idx, start_index in enumerate(rnd_start_indices):
|
| 90 |
+
input_mask[batch_idx, :start_index] = 1
|
| 91 |
+
input_mask[batch_idx, start_index:] = 0
|
| 92 |
+
|
| 93 |
+
config = self.get_config()
|
| 94 |
+
|
| 95 |
+
return config, input_ids, tf.convert_to_tensor(input_mask)
|
| 96 |
+
|
| 97 |
+
def get_config(self):
|
| 98 |
+
return BlipTextConfig(
|
| 99 |
+
vocab_size=self.vocab_size,
|
| 100 |
+
hidden_size=self.hidden_size,
|
| 101 |
+
projection_dim=self.projection_dim,
|
| 102 |
+
num_hidden_layers=self.num_hidden_layers,
|
| 103 |
+
num_attention_heads=self.num_attention_heads,
|
| 104 |
+
intermediate_size=self.intermediate_size,
|
| 105 |
+
dropout=self.dropout,
|
| 106 |
+
attention_dropout=self.attention_dropout,
|
| 107 |
+
max_position_embeddings=self.max_position_embeddings,
|
| 108 |
+
initializer_range=self.initializer_range,
|
| 109 |
+
bos_token_id=self.bos_token_id,
|
| 110 |
+
)
|
| 111 |
+
|
| 112 |
+
def create_and_check_model(self, config, input_ids, input_mask):
|
| 113 |
+
model = TFBlipTextModel(config=config)
|
| 114 |
+
result = model(input_ids, attention_mask=input_mask, training=False)
|
| 115 |
+
result = model(input_ids, training=False)
|
| 116 |
+
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
|
| 117 |
+
self.parent.assertEqual(result.pooler_output.shape, (self.batch_size, self.hidden_size))
|
| 118 |
+
|
| 119 |
+
def prepare_config_and_inputs_for_common(self):
|
| 120 |
+
config_and_inputs = self.prepare_config_and_inputs()
|
| 121 |
+
config, input_ids, input_mask = config_and_inputs
|
| 122 |
+
inputs_dict = {"input_ids": input_ids, "attention_mask": input_mask}
|
| 123 |
+
return config, inputs_dict
|
| 124 |
+
|
| 125 |
+
|
| 126 |
+
@require_tf
|
| 127 |
+
class BlipTextModelTest(TFModelTesterMixin, unittest.TestCase):
|
| 128 |
+
all_model_classes = (TFBlipTextModel,) if is_tf_available() else ()
|
| 129 |
+
test_onnx = False
|
| 130 |
+
test_pruning = False
|
| 131 |
+
test_head_masking = False
|
| 132 |
+
|
| 133 |
+
def setUp(self):
|
| 134 |
+
self.model_tester = BlipTextModelTester(self)
|
| 135 |
+
self.config_tester = ConfigTester(self, config_class=BlipTextConfig, hidden_size=37)
|
| 136 |
+
|
| 137 |
+
def test_config(self):
|
| 138 |
+
self.config_tester.run_common_tests()
|
| 139 |
+
|
| 140 |
+
def test_model(self):
|
| 141 |
+
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
| 142 |
+
self.model_tester.create_and_check_model(*config_and_inputs)
|
| 143 |
+
|
| 144 |
+
def test_training(self):
|
| 145 |
+
pass
|
| 146 |
+
|
| 147 |
+
def test_training_gradient_checkpointing(self):
|
| 148 |
+
pass
|
| 149 |
+
|
| 150 |
+
@unittest.skip(
|
| 151 |
+
reason="This architecure seem to not compute gradients properly when using GC, check: https://github.com/huggingface/transformers/pull/27124"
|
| 152 |
+
)
|
| 153 |
+
def test_training_gradient_checkpointing_use_reentrant(self):
|
| 154 |
+
pass
|
| 155 |
+
|
| 156 |
+
@unittest.skip(
|
| 157 |
+
reason="This architecure seem to not compute gradients properly when using GC, check: https://github.com/huggingface/transformers/pull/27124"
|
| 158 |
+
)
|
| 159 |
+
def test_training_gradient_checkpointing_use_reentrant_false(self):
|
| 160 |
+
pass
|
| 161 |
+
|
| 162 |
+
@unittest.skip(reason="Blip does not use inputs_embeds")
|
| 163 |
+
def test_inputs_embeds(self):
|
| 164 |
+
pass
|
| 165 |
+
|
| 166 |
+
@unittest.skip(reason="BlipTextModel has no base class and is not available in MODEL_MAPPING")
|
| 167 |
+
def test_save_load_fast_init_from_base(self):
|
| 168 |
+
pass
|
| 169 |
+
|
| 170 |
+
@unittest.skip(reason="BlipTextModel has no base class and is not available in MODEL_MAPPING")
|
| 171 |
+
def test_save_load_fast_init_to_base(self):
|
| 172 |
+
pass
|
| 173 |
+
|
| 174 |
+
@slow
|
| 175 |
+
def test_model_from_pretrained(self):
|
| 176 |
+
model_name = "Salesforce/blip-vqa-base"
|
| 177 |
+
model = TFBlipTextModel.from_pretrained(model_name)
|
| 178 |
+
self.assertIsNotNone(model)
|
| 179 |
+
|
| 180 |
+
def test_pt_tf_model_equivalence(self, allow_missing_keys=True):
|
| 181 |
+
super().test_pt_tf_model_equivalence(allow_missing_keys=allow_missing_keys)
|
tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/blip/test_processor_blip.py
ADDED
|
@@ -0,0 +1,167 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Copyright 2022 The HuggingFace Team. All rights reserved.
|
| 2 |
+
#
|
| 3 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
| 4 |
+
# you may not use this file except in compliance with the License.
|
| 5 |
+
# You may obtain a copy of the License at
|
| 6 |
+
#
|
| 7 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
| 8 |
+
#
|
| 9 |
+
# Unless required by applicable law or agreed to in writing, software
|
| 10 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
| 11 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
| 12 |
+
# See the License for the specific language governing permissions and
|
| 13 |
+
# limitations under the License.
|
| 14 |
+
import shutil
|
| 15 |
+
import tempfile
|
| 16 |
+
import unittest
|
| 17 |
+
|
| 18 |
+
import pytest
|
| 19 |
+
|
| 20 |
+
from transformers.testing_utils import require_torch, require_vision
|
| 21 |
+
from transformers.utils import is_vision_available
|
| 22 |
+
|
| 23 |
+
from ...test_processing_common import ProcessorTesterMixin
|
| 24 |
+
|
| 25 |
+
|
| 26 |
+
if is_vision_available():
|
| 27 |
+
from transformers import AutoProcessor, BertTokenizer, BlipImageProcessor, BlipProcessor, PreTrainedTokenizerFast
|
| 28 |
+
|
| 29 |
+
|
| 30 |
+
@require_vision
|
| 31 |
+
class BlipProcessorTest(ProcessorTesterMixin, unittest.TestCase):
|
| 32 |
+
processor_class = BlipProcessor
|
| 33 |
+
|
| 34 |
+
def setUp(self):
|
| 35 |
+
self.tmpdirname = tempfile.mkdtemp()
|
| 36 |
+
|
| 37 |
+
image_processor = BlipImageProcessor()
|
| 38 |
+
tokenizer = BertTokenizer.from_pretrained("hf-internal-testing/tiny-random-BertModel")
|
| 39 |
+
|
| 40 |
+
processor = BlipProcessor(image_processor, tokenizer)
|
| 41 |
+
|
| 42 |
+
processor.save_pretrained(self.tmpdirname)
|
| 43 |
+
|
| 44 |
+
def get_tokenizer(self, **kwargs):
|
| 45 |
+
return AutoProcessor.from_pretrained(self.tmpdirname, **kwargs).tokenizer
|
| 46 |
+
|
| 47 |
+
def get_image_processor(self, **kwargs):
|
| 48 |
+
return AutoProcessor.from_pretrained(self.tmpdirname, **kwargs).image_processor
|
| 49 |
+
|
| 50 |
+
def tearDown(self):
|
| 51 |
+
shutil.rmtree(self.tmpdirname)
|
| 52 |
+
|
| 53 |
+
def test_save_load_pretrained_additional_features(self):
|
| 54 |
+
processor = BlipProcessor(tokenizer=self.get_tokenizer(), image_processor=self.get_image_processor())
|
| 55 |
+
processor.save_pretrained(self.tmpdirname)
|
| 56 |
+
|
| 57 |
+
tokenizer_add_kwargs = self.get_tokenizer(bos_token="(BOS)", eos_token="(EOS)")
|
| 58 |
+
image_processor_add_kwargs = self.get_image_processor(do_normalize=False, padding_value=1.0)
|
| 59 |
+
|
| 60 |
+
processor = BlipProcessor.from_pretrained(
|
| 61 |
+
self.tmpdirname, bos_token="(BOS)", eos_token="(EOS)", do_normalize=False, padding_value=1.0
|
| 62 |
+
)
|
| 63 |
+
|
| 64 |
+
self.assertEqual(processor.tokenizer.get_vocab(), tokenizer_add_kwargs.get_vocab())
|
| 65 |
+
self.assertIsInstance(processor.tokenizer, PreTrainedTokenizerFast)
|
| 66 |
+
|
| 67 |
+
self.assertEqual(processor.image_processor.to_json_string(), image_processor_add_kwargs.to_json_string())
|
| 68 |
+
self.assertIsInstance(processor.image_processor, BlipImageProcessor)
|
| 69 |
+
|
| 70 |
+
def test_image_processor(self):
|
| 71 |
+
image_processor = self.get_image_processor()
|
| 72 |
+
tokenizer = self.get_tokenizer()
|
| 73 |
+
|
| 74 |
+
processor = BlipProcessor(tokenizer=tokenizer, image_processor=image_processor)
|
| 75 |
+
|
| 76 |
+
image_input = self.prepare_image_inputs()
|
| 77 |
+
|
| 78 |
+
input_feat_extract = image_processor(image_input, return_tensors="np")
|
| 79 |
+
input_processor = processor(images=image_input, return_tensors="np")
|
| 80 |
+
|
| 81 |
+
for key in input_feat_extract.keys():
|
| 82 |
+
self.assertAlmostEqual(input_feat_extract[key].sum(), input_processor[key].sum(), delta=1e-2)
|
| 83 |
+
|
| 84 |
+
def test_tokenizer(self):
|
| 85 |
+
image_processor = self.get_image_processor()
|
| 86 |
+
tokenizer = self.get_tokenizer()
|
| 87 |
+
|
| 88 |
+
processor = BlipProcessor(tokenizer=tokenizer, image_processor=image_processor)
|
| 89 |
+
|
| 90 |
+
input_str = "lower newer"
|
| 91 |
+
|
| 92 |
+
encoded_processor = processor(text=input_str)
|
| 93 |
+
|
| 94 |
+
encoded_tok = tokenizer(input_str, return_token_type_ids=False)
|
| 95 |
+
|
| 96 |
+
for key in encoded_tok.keys():
|
| 97 |
+
self.assertListEqual(encoded_tok[key], encoded_processor[key])
|
| 98 |
+
|
| 99 |
+
def test_processor(self):
|
| 100 |
+
image_processor = self.get_image_processor()
|
| 101 |
+
tokenizer = self.get_tokenizer()
|
| 102 |
+
|
| 103 |
+
processor = BlipProcessor(tokenizer=tokenizer, image_processor=image_processor)
|
| 104 |
+
|
| 105 |
+
input_str = "lower newer"
|
| 106 |
+
image_input = self.prepare_image_inputs()
|
| 107 |
+
|
| 108 |
+
inputs = processor(text=input_str, images=image_input)
|
| 109 |
+
|
| 110 |
+
self.assertListEqual(list(inputs.keys()), ["pixel_values", "input_ids", "attention_mask"])
|
| 111 |
+
|
| 112 |
+
# test if it raises when no input is passed
|
| 113 |
+
with pytest.raises(ValueError):
|
| 114 |
+
processor()
|
| 115 |
+
|
| 116 |
+
def test_tokenizer_decode(self):
|
| 117 |
+
image_processor = self.get_image_processor()
|
| 118 |
+
tokenizer = self.get_tokenizer()
|
| 119 |
+
|
| 120 |
+
processor = BlipProcessor(tokenizer=tokenizer, image_processor=image_processor)
|
| 121 |
+
|
| 122 |
+
predicted_ids = [[1, 4, 5, 8, 1, 0, 8], [3, 4, 3, 1, 1, 8, 9]]
|
| 123 |
+
|
| 124 |
+
decoded_processor = processor.batch_decode(predicted_ids)
|
| 125 |
+
decoded_tok = tokenizer.batch_decode(predicted_ids)
|
| 126 |
+
|
| 127 |
+
self.assertListEqual(decoded_tok, decoded_processor)
|
| 128 |
+
|
| 129 |
+
def test_model_input_names(self):
|
| 130 |
+
image_processor = self.get_image_processor()
|
| 131 |
+
tokenizer = self.get_tokenizer()
|
| 132 |
+
|
| 133 |
+
processor = BlipProcessor(tokenizer=tokenizer, image_processor=image_processor)
|
| 134 |
+
|
| 135 |
+
input_str = "lower newer"
|
| 136 |
+
image_input = self.prepare_image_inputs()
|
| 137 |
+
|
| 138 |
+
inputs = processor(text=input_str, images=image_input)
|
| 139 |
+
|
| 140 |
+
# For now the processor supports only ['pixel_values', 'input_ids', 'attention_mask']
|
| 141 |
+
self.assertListEqual(list(inputs.keys()), ["pixel_values", "input_ids", "attention_mask"])
|
| 142 |
+
|
| 143 |
+
@require_torch
|
| 144 |
+
@require_vision
|
| 145 |
+
def test_unstructured_kwargs_batched(self):
|
| 146 |
+
if "image_processor" not in self.processor_class.attributes:
|
| 147 |
+
self.skipTest(f"image_processor attribute not present in {self.processor_class}")
|
| 148 |
+
image_processor = self.get_component("image_processor")
|
| 149 |
+
tokenizer = self.get_component("tokenizer")
|
| 150 |
+
|
| 151 |
+
processor = self.processor_class(tokenizer=tokenizer, image_processor=image_processor)
|
| 152 |
+
self.skip_processor_without_typed_kwargs(processor)
|
| 153 |
+
|
| 154 |
+
input_str = ["lower newer", "upper older longer string"]
|
| 155 |
+
image_input = self.prepare_image_inputs(batch_size=2)
|
| 156 |
+
inputs = processor(
|
| 157 |
+
text=input_str,
|
| 158 |
+
images=image_input,
|
| 159 |
+
return_tensors="pt",
|
| 160 |
+
crop_size={"height": 214, "width": 214},
|
| 161 |
+
size={"height": 214, "width": 214},
|
| 162 |
+
padding="longest",
|
| 163 |
+
max_length=76,
|
| 164 |
+
)
|
| 165 |
+
self.assertEqual(inputs["pixel_values"].shape[2], 214)
|
| 166 |
+
|
| 167 |
+
self.assertEqual(len(inputs["input_ids"][0]), 24)
|
tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/blip_2/__init__.py
ADDED
|
File without changes
|
tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/blip_2/test_modeling_blip_2.py
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/blip_2/test_processor_blip_2.py
ADDED
|
@@ -0,0 +1,141 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Copyright 2023 The HuggingFace Team. All rights reserved.
|
| 2 |
+
#
|
| 3 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
| 4 |
+
# you may not use this file except in compliance with the License.
|
| 5 |
+
# You may obtain a copy of the License at
|
| 6 |
+
#
|
| 7 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
| 8 |
+
#
|
| 9 |
+
# Unless required by applicable law or agreed to in writing, software
|
| 10 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
| 11 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
| 12 |
+
# See the License for the specific language governing permissions and
|
| 13 |
+
# limitations under the License.
|
| 14 |
+
import shutil
|
| 15 |
+
import tempfile
|
| 16 |
+
import unittest
|
| 17 |
+
|
| 18 |
+
import pytest
|
| 19 |
+
|
| 20 |
+
from transformers.testing_utils import require_vision
|
| 21 |
+
from transformers.utils import is_vision_available
|
| 22 |
+
|
| 23 |
+
from ...test_processing_common import ProcessorTesterMixin
|
| 24 |
+
|
| 25 |
+
|
| 26 |
+
if is_vision_available():
|
| 27 |
+
from transformers import AutoProcessor, Blip2Processor, BlipImageProcessor, GPT2Tokenizer, PreTrainedTokenizerFast
|
| 28 |
+
|
| 29 |
+
|
| 30 |
+
@require_vision
|
| 31 |
+
class Blip2ProcessorTest(ProcessorTesterMixin, unittest.TestCase):
|
| 32 |
+
processor_class = Blip2Processor
|
| 33 |
+
|
| 34 |
+
def setUp(self):
|
| 35 |
+
self.tmpdirname = tempfile.mkdtemp()
|
| 36 |
+
|
| 37 |
+
image_processor = BlipImageProcessor()
|
| 38 |
+
tokenizer = GPT2Tokenizer.from_pretrained("hf-internal-testing/tiny-random-GPT2Model")
|
| 39 |
+
|
| 40 |
+
processor = Blip2Processor(image_processor, tokenizer)
|
| 41 |
+
|
| 42 |
+
processor.save_pretrained(self.tmpdirname)
|
| 43 |
+
|
| 44 |
+
def get_tokenizer(self, **kwargs):
|
| 45 |
+
return AutoProcessor.from_pretrained(self.tmpdirname, **kwargs).tokenizer
|
| 46 |
+
|
| 47 |
+
def get_image_processor(self, **kwargs):
|
| 48 |
+
return AutoProcessor.from_pretrained(self.tmpdirname, **kwargs).image_processor
|
| 49 |
+
|
| 50 |
+
def tearDown(self):
|
| 51 |
+
shutil.rmtree(self.tmpdirname)
|
| 52 |
+
|
| 53 |
+
def test_save_load_pretrained_additional_features(self):
|
| 54 |
+
processor = Blip2Processor(tokenizer=self.get_tokenizer(), image_processor=self.get_image_processor())
|
| 55 |
+
processor.save_pretrained(self.tmpdirname)
|
| 56 |
+
|
| 57 |
+
tokenizer_add_kwargs = self.get_tokenizer(bos_token="(BOS)", eos_token="(EOS)")
|
| 58 |
+
image_processor_add_kwargs = self.get_image_processor(do_normalize=False, padding_value=1.0)
|
| 59 |
+
|
| 60 |
+
processor = Blip2Processor.from_pretrained(
|
| 61 |
+
self.tmpdirname, bos_token="(BOS)", eos_token="(EOS)", do_normalize=False, padding_value=1.0
|
| 62 |
+
)
|
| 63 |
+
|
| 64 |
+
self.assertEqual(processor.tokenizer.get_vocab(), tokenizer_add_kwargs.get_vocab())
|
| 65 |
+
self.assertIsInstance(processor.tokenizer, PreTrainedTokenizerFast)
|
| 66 |
+
|
| 67 |
+
self.assertEqual(processor.image_processor.to_json_string(), image_processor_add_kwargs.to_json_string())
|
| 68 |
+
self.assertIsInstance(processor.image_processor, BlipImageProcessor)
|
| 69 |
+
|
| 70 |
+
def test_image_processor(self):
|
| 71 |
+
image_processor = self.get_image_processor()
|
| 72 |
+
tokenizer = self.get_tokenizer()
|
| 73 |
+
|
| 74 |
+
processor = Blip2Processor(tokenizer=tokenizer, image_processor=image_processor)
|
| 75 |
+
|
| 76 |
+
image_input = self.prepare_image_inputs()
|
| 77 |
+
|
| 78 |
+
input_feat_extract = image_processor(image_input, return_tensors="np")
|
| 79 |
+
input_processor = processor(images=image_input, return_tensors="np")
|
| 80 |
+
|
| 81 |
+
for key in input_feat_extract.keys():
|
| 82 |
+
self.assertAlmostEqual(input_feat_extract[key].sum(), input_processor[key].sum(), delta=1e-2)
|
| 83 |
+
|
| 84 |
+
def test_tokenizer(self):
|
| 85 |
+
image_processor = self.get_image_processor()
|
| 86 |
+
tokenizer = self.get_tokenizer()
|
| 87 |
+
|
| 88 |
+
processor = Blip2Processor(tokenizer=tokenizer, image_processor=image_processor)
|
| 89 |
+
|
| 90 |
+
input_str = "lower newer"
|
| 91 |
+
|
| 92 |
+
encoded_processor = processor(text=input_str)
|
| 93 |
+
|
| 94 |
+
encoded_tok = tokenizer(input_str, return_token_type_ids=False)
|
| 95 |
+
|
| 96 |
+
for key in encoded_tok.keys():
|
| 97 |
+
self.assertListEqual(encoded_tok[key], encoded_processor[key][0])
|
| 98 |
+
|
| 99 |
+
def test_processor(self):
|
| 100 |
+
image_processor = self.get_image_processor()
|
| 101 |
+
tokenizer = self.get_tokenizer()
|
| 102 |
+
|
| 103 |
+
processor = Blip2Processor(tokenizer=tokenizer, image_processor=image_processor)
|
| 104 |
+
|
| 105 |
+
input_str = "lower newer"
|
| 106 |
+
image_input = self.prepare_image_inputs()
|
| 107 |
+
|
| 108 |
+
inputs = processor(text=input_str, images=image_input)
|
| 109 |
+
|
| 110 |
+
self.assertCountEqual(list(inputs.keys()), ["input_ids", "pixel_values", "attention_mask"])
|
| 111 |
+
|
| 112 |
+
# test if it raises when no input is passed
|
| 113 |
+
with pytest.raises(ValueError):
|
| 114 |
+
processor()
|
| 115 |
+
|
| 116 |
+
def test_tokenizer_decode(self):
|
| 117 |
+
image_processor = self.get_image_processor()
|
| 118 |
+
tokenizer = self.get_tokenizer()
|
| 119 |
+
|
| 120 |
+
processor = Blip2Processor(tokenizer=tokenizer, image_processor=image_processor)
|
| 121 |
+
|
| 122 |
+
predicted_ids = [[1, 4, 5, 8, 1, 0, 8], [3, 4, 3, 1, 1, 8, 9]]
|
| 123 |
+
|
| 124 |
+
decoded_processor = processor.batch_decode(predicted_ids)
|
| 125 |
+
decoded_tok = tokenizer.batch_decode(predicted_ids)
|
| 126 |
+
|
| 127 |
+
self.assertListEqual(decoded_tok, decoded_processor)
|
| 128 |
+
|
| 129 |
+
def test_model_input_names(self):
|
| 130 |
+
image_processor = self.get_image_processor()
|
| 131 |
+
tokenizer = self.get_tokenizer()
|
| 132 |
+
|
| 133 |
+
processor = Blip2Processor(tokenizer=tokenizer, image_processor=image_processor)
|
| 134 |
+
|
| 135 |
+
input_str = "lower newer"
|
| 136 |
+
image_input = self.prepare_image_inputs()
|
| 137 |
+
|
| 138 |
+
inputs = processor(text=input_str, images=image_input)
|
| 139 |
+
|
| 140 |
+
# For now the processor supports only ['pixel_values', 'input_ids', 'attention_mask']
|
| 141 |
+
self.assertCountEqual(list(inputs.keys()), ["input_ids", "pixel_values", "attention_mask"])
|
tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/bloom/__init__.py
ADDED
|
File without changes
|
tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/bloom/test_modeling_bloom.py
ADDED
|
@@ -0,0 +1,825 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# coding=utf-8
|
| 2 |
+
# Copyright 2022 The HuggingFace Team. All rights reserved.
|
| 3 |
+
#
|
| 4 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
| 5 |
+
# you may not use this file except in compliance with the License.
|
| 6 |
+
# You may obtain a copy of the License at
|
| 7 |
+
#
|
| 8 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
| 9 |
+
#
|
| 10 |
+
# Unless required by applicable law or agreed to in writing, software
|
| 11 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
| 12 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
| 13 |
+
# See the License for the specific language governing permissions and
|
| 14 |
+
# limitations under the License.
|
| 15 |
+
#
|
| 16 |
+
|
| 17 |
+
import math
|
| 18 |
+
import unittest
|
| 19 |
+
|
| 20 |
+
from transformers import BloomConfig, is_torch_available
|
| 21 |
+
from transformers.testing_utils import require_torch, require_torch_accelerator, slow, torch_device
|
| 22 |
+
|
| 23 |
+
from ...generation.test_utils import GenerationTesterMixin
|
| 24 |
+
from ...test_configuration_common import ConfigTester
|
| 25 |
+
from ...test_modeling_common import ModelTesterMixin, ids_tensor, random_attention_mask
|
| 26 |
+
from ...test_pipeline_mixin import PipelineTesterMixin
|
| 27 |
+
|
| 28 |
+
|
| 29 |
+
if is_torch_available():
|
| 30 |
+
import torch
|
| 31 |
+
|
| 32 |
+
from transformers import (
|
| 33 |
+
BloomForCausalLM,
|
| 34 |
+
BloomForQuestionAnswering,
|
| 35 |
+
BloomForSequenceClassification,
|
| 36 |
+
BloomForTokenClassification,
|
| 37 |
+
BloomModel,
|
| 38 |
+
BloomTokenizerFast,
|
| 39 |
+
)
|
| 40 |
+
|
| 41 |
+
|
| 42 |
+
@require_torch
|
| 43 |
+
class BloomModelTester:
|
| 44 |
+
def __init__(
|
| 45 |
+
self,
|
| 46 |
+
parent,
|
| 47 |
+
batch_size=14,
|
| 48 |
+
seq_length=7,
|
| 49 |
+
is_training=True,
|
| 50 |
+
use_token_type_ids=False,
|
| 51 |
+
use_input_mask=True,
|
| 52 |
+
use_labels=True,
|
| 53 |
+
use_mc_token_ids=True,
|
| 54 |
+
vocab_size=99,
|
| 55 |
+
hidden_size=32,
|
| 56 |
+
num_hidden_layers=2,
|
| 57 |
+
num_attention_heads=4,
|
| 58 |
+
intermediate_size=37,
|
| 59 |
+
hidden_act="gelu",
|
| 60 |
+
hidden_dropout_prob=0.1,
|
| 61 |
+
attention_dropout_prob=0.1,
|
| 62 |
+
max_position_embeddings=512,
|
| 63 |
+
type_vocab_size=16,
|
| 64 |
+
type_sequence_label_size=2,
|
| 65 |
+
initializer_range=0.02,
|
| 66 |
+
num_labels=3,
|
| 67 |
+
num_choices=4,
|
| 68 |
+
scope=None,
|
| 69 |
+
):
|
| 70 |
+
self.parent = parent
|
| 71 |
+
self.batch_size = batch_size
|
| 72 |
+
self.seq_length = seq_length
|
| 73 |
+
self.is_training = is_training
|
| 74 |
+
self.use_token_type_ids = use_token_type_ids
|
| 75 |
+
self.use_input_mask = use_input_mask
|
| 76 |
+
self.use_labels = use_labels
|
| 77 |
+
self.use_mc_token_ids = use_mc_token_ids
|
| 78 |
+
self.vocab_size = vocab_size
|
| 79 |
+
self.hidden_size = hidden_size
|
| 80 |
+
self.num_hidden_layers = num_hidden_layers
|
| 81 |
+
self.num_attention_heads = num_attention_heads
|
| 82 |
+
self.intermediate_size = intermediate_size
|
| 83 |
+
self.hidden_act = hidden_act
|
| 84 |
+
self.hidden_dropout_prob = hidden_dropout_prob
|
| 85 |
+
self.attention_dropout_prob = attention_dropout_prob
|
| 86 |
+
self.max_position_embeddings = max_position_embeddings
|
| 87 |
+
self.type_vocab_size = type_vocab_size
|
| 88 |
+
self.type_sequence_label_size = type_sequence_label_size
|
| 89 |
+
self.initializer_range = initializer_range
|
| 90 |
+
self.num_labels = num_labels
|
| 91 |
+
self.num_choices = num_choices
|
| 92 |
+
self.scope = None
|
| 93 |
+
self.bos_token_id = vocab_size - 1
|
| 94 |
+
self.eos_token_id = vocab_size - 1
|
| 95 |
+
self.pad_token_id = vocab_size - 1
|
| 96 |
+
|
| 97 |
+
def get_large_model_config(self):
|
| 98 |
+
return BloomConfig.from_pretrained("bigscience/bloom")
|
| 99 |
+
|
| 100 |
+
def prepare_config_and_inputs(self, gradient_checkpointing=False):
|
| 101 |
+
input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
|
| 102 |
+
|
| 103 |
+
input_mask = None
|
| 104 |
+
if self.use_input_mask:
|
| 105 |
+
input_mask = random_attention_mask([self.batch_size, self.seq_length])
|
| 106 |
+
|
| 107 |
+
sequence_labels = None
|
| 108 |
+
if self.use_labels:
|
| 109 |
+
sequence_labels = ids_tensor([self.batch_size], self.type_sequence_label_size)
|
| 110 |
+
|
| 111 |
+
config = self.get_config(gradient_checkpointing=gradient_checkpointing)
|
| 112 |
+
|
| 113 |
+
return (config, input_ids, input_mask, sequence_labels)
|
| 114 |
+
|
| 115 |
+
def get_config(self, gradient_checkpointing=False, slow_but_exact=True):
|
| 116 |
+
return BloomConfig(
|
| 117 |
+
vocab_size=self.vocab_size,
|
| 118 |
+
seq_length=self.seq_length,
|
| 119 |
+
hidden_size=self.hidden_size,
|
| 120 |
+
n_layer=self.num_hidden_layers,
|
| 121 |
+
n_head=self.num_attention_heads,
|
| 122 |
+
hidden_dropout=self.hidden_dropout_prob,
|
| 123 |
+
attention_dropout=self.attention_dropout_prob,
|
| 124 |
+
n_positions=self.max_position_embeddings,
|
| 125 |
+
type_vocab_size=self.type_vocab_size,
|
| 126 |
+
initializer_range=self.initializer_range,
|
| 127 |
+
use_cache=True,
|
| 128 |
+
bos_token_id=self.bos_token_id,
|
| 129 |
+
eos_token_id=self.eos_token_id,
|
| 130 |
+
pad_token_id=self.pad_token_id,
|
| 131 |
+
num_labels=self.num_labels,
|
| 132 |
+
gradient_checkpointing=gradient_checkpointing,
|
| 133 |
+
slow_but_exact=slow_but_exact,
|
| 134 |
+
dtype="float32",
|
| 135 |
+
)
|
| 136 |
+
|
| 137 |
+
def create_and_check_bloom_model(self, config, input_ids, input_mask, *args):
|
| 138 |
+
model = BloomModel(config=config)
|
| 139 |
+
model.to(torch_device)
|
| 140 |
+
model.eval()
|
| 141 |
+
|
| 142 |
+
result = model(input_ids)
|
| 143 |
+
|
| 144 |
+
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
|
| 145 |
+
self.parent.assertEqual(len(result.past_key_values), config.n_layer)
|
| 146 |
+
|
| 147 |
+
def create_and_check_bloom_model_past(self, config, input_ids, input_mask, *args):
|
| 148 |
+
model = BloomModel(config=config)
|
| 149 |
+
|
| 150 |
+
model.to(torch_device)
|
| 151 |
+
model.eval()
|
| 152 |
+
|
| 153 |
+
# first forward pass
|
| 154 |
+
outputs = model(input_ids, attention_mask=torch.ones_like(input_ids), use_cache=True)
|
| 155 |
+
outputs_use_cache_conf = model(input_ids, attention_mask=torch.ones_like(input_ids))
|
| 156 |
+
outputs_no_past = model(input_ids, use_cache=False, attention_mask=torch.ones_like(input_ids))
|
| 157 |
+
|
| 158 |
+
self.parent.assertTrue(len(outputs) == len(outputs_use_cache_conf))
|
| 159 |
+
self.parent.assertTrue(len(outputs) == len(outputs_no_past) + 1)
|
| 160 |
+
|
| 161 |
+
past = outputs["past_key_values"]
|
| 162 |
+
|
| 163 |
+
# create hypothetical next token and extent to next_input_ids
|
| 164 |
+
next_tokens = ids_tensor((self.batch_size, 1), config.vocab_size)
|
| 165 |
+
|
| 166 |
+
# append to next input_ids and token_type_ids
|
| 167 |
+
next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
|
| 168 |
+
|
| 169 |
+
output_from_no_past = model(next_input_ids)["last_hidden_state"]
|
| 170 |
+
output_from_past = model(next_tokens, past_key_values=past)["last_hidden_state"]
|
| 171 |
+
|
| 172 |
+
# select random slice
|
| 173 |
+
random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()
|
| 174 |
+
output_from_no_past_slice = output_from_no_past[:, -1, random_slice_idx].detach()
|
| 175 |
+
output_from_past_slice = output_from_past[:, 0, random_slice_idx].detach()
|
| 176 |
+
|
| 177 |
+
# test that outputs are equal for slice
|
| 178 |
+
self.parent.assertTrue(torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3))
|
| 179 |
+
|
| 180 |
+
def create_and_check_bloom_model_attention_mask_past(self, config, input_ids, input_mask, *args):
|
| 181 |
+
model = BloomModel(config=config)
|
| 182 |
+
model.to(torch_device)
|
| 183 |
+
model.eval()
|
| 184 |
+
|
| 185 |
+
# create attention mask
|
| 186 |
+
attn_mask = torch.ones(input_ids.shape, dtype=torch.long, device=torch_device)
|
| 187 |
+
half_seq_length = self.seq_length // 2
|
| 188 |
+
attn_mask[:, half_seq_length:] = 0
|
| 189 |
+
|
| 190 |
+
# first forward pass
|
| 191 |
+
output, past = model(input_ids, attention_mask=attn_mask).to_tuple()
|
| 192 |
+
|
| 193 |
+
# create hypothetical next token and extent to next_input_ids
|
| 194 |
+
next_tokens = ids_tensor((self.batch_size, 1), config.vocab_size)
|
| 195 |
+
|
| 196 |
+
# change a random masked slice from input_ids
|
| 197 |
+
random_seq_idx_to_change = ids_tensor((1,), half_seq_length).item() + 1
|
| 198 |
+
random_other_next_tokens = ids_tensor((self.batch_size, 1), config.vocab_size).squeeze(-1)
|
| 199 |
+
input_ids[:, -random_seq_idx_to_change] = random_other_next_tokens
|
| 200 |
+
|
| 201 |
+
# append to next input_ids and attn_mask
|
| 202 |
+
next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
|
| 203 |
+
attn_mask = torch.cat(
|
| 204 |
+
[attn_mask, torch.ones((attn_mask.shape[0], 1), dtype=torch.long, device=torch_device)],
|
| 205 |
+
dim=1,
|
| 206 |
+
)
|
| 207 |
+
|
| 208 |
+
# get two different outputs
|
| 209 |
+
output_from_no_past = model(next_input_ids, attention_mask=attn_mask)["last_hidden_state"]
|
| 210 |
+
output_from_past = model(next_tokens, past_key_values=past, attention_mask=attn_mask)["last_hidden_state"]
|
| 211 |
+
|
| 212 |
+
# select random slice
|
| 213 |
+
random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()
|
| 214 |
+
output_from_no_past_slice = output_from_no_past[:, -1, random_slice_idx].detach()
|
| 215 |
+
output_from_past_slice = output_from_past[:, 0, random_slice_idx].detach()
|
| 216 |
+
|
| 217 |
+
# test that outputs are equal for slice
|
| 218 |
+
self.parent.assertTrue(torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3))
|
| 219 |
+
|
| 220 |
+
def create_and_check_bloom_model_past_large_inputs(self, config, input_ids, input_mask, *args):
|
| 221 |
+
model = BloomModel(config=config)
|
| 222 |
+
model.to(torch_device)
|
| 223 |
+
model.eval()
|
| 224 |
+
|
| 225 |
+
# first forward pass
|
| 226 |
+
outputs = model(input_ids, attention_mask=input_mask, use_cache=True)
|
| 227 |
+
|
| 228 |
+
output, past = outputs.to_tuple()
|
| 229 |
+
|
| 230 |
+
# create hypothetical next token and extent to next_input_ids
|
| 231 |
+
next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size)
|
| 232 |
+
next_mask = ids_tensor((self.batch_size, 3), vocab_size=2)
|
| 233 |
+
|
| 234 |
+
# append to next input_ids and token_type_ids
|
| 235 |
+
next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
|
| 236 |
+
next_attention_mask = torch.cat([input_mask, next_mask], dim=-1)
|
| 237 |
+
|
| 238 |
+
output_from_no_past = model(next_input_ids, attention_mask=next_attention_mask)["last_hidden_state"]
|
| 239 |
+
output_from_past = model(next_tokens, attention_mask=next_attention_mask, past_key_values=past)[
|
| 240 |
+
"last_hidden_state"
|
| 241 |
+
]
|
| 242 |
+
self.parent.assertTrue(output_from_past.shape[1] == next_tokens.shape[1])
|
| 243 |
+
|
| 244 |
+
# select random slice
|
| 245 |
+
random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()
|
| 246 |
+
output_from_no_past_slice = output_from_no_past[:, -3:, random_slice_idx].detach()
|
| 247 |
+
output_from_past_slice = output_from_past[:, :, random_slice_idx].detach()
|
| 248 |
+
|
| 249 |
+
# test that outputs are equal for slice
|
| 250 |
+
self.parent.assertTrue(torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3))
|
| 251 |
+
|
| 252 |
+
def create_and_check_lm_head_model(self, config, input_ids, input_mask, *args):
|
| 253 |
+
model = BloomForCausalLM(config)
|
| 254 |
+
model.to(torch_device)
|
| 255 |
+
model.eval()
|
| 256 |
+
|
| 257 |
+
result = model(input_ids, labels=input_ids)
|
| 258 |
+
self.parent.assertEqual(result.loss.shape, ())
|
| 259 |
+
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
|
| 260 |
+
|
| 261 |
+
def create_and_check_sequence_classification_model(self, config, input_ids, input_mask, *args):
|
| 262 |
+
config.num_labels = self.num_labels
|
| 263 |
+
model = BloomForSequenceClassification(config)
|
| 264 |
+
model.to(torch_device)
|
| 265 |
+
model.eval()
|
| 266 |
+
|
| 267 |
+
result = model(input_ids, attention_mask=input_mask)
|
| 268 |
+
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.num_labels))
|
| 269 |
+
|
| 270 |
+
def create_and_check_token_classification_model(self, config, input_ids, input_mask, *args):
|
| 271 |
+
model = BloomForTokenClassification(config)
|
| 272 |
+
model.to(torch_device)
|
| 273 |
+
model.eval()
|
| 274 |
+
|
| 275 |
+
result = model(input_ids, attention_mask=input_mask)
|
| 276 |
+
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.num_labels))
|
| 277 |
+
|
| 278 |
+
def create_and_check_question_answering_model(self, config, input_ids, input_mask, *args):
|
| 279 |
+
model = BloomForQuestionAnswering(config)
|
| 280 |
+
model.to(torch_device)
|
| 281 |
+
model.eval()
|
| 282 |
+
|
| 283 |
+
result = model(input_ids, attention_mask=input_mask)
|
| 284 |
+
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.num_labels))
|
| 285 |
+
|
| 286 |
+
def create_and_check_forward_and_backwards(
|
| 287 |
+
self, config, input_ids, input_mask, *args, gradient_checkpointing=False
|
| 288 |
+
):
|
| 289 |
+
model = BloomForCausalLM(config)
|
| 290 |
+
model.to(torch_device)
|
| 291 |
+
if gradient_checkpointing:
|
| 292 |
+
model.gradient_checkpointing_enable()
|
| 293 |
+
|
| 294 |
+
result = model(input_ids, labels=input_ids)
|
| 295 |
+
self.parent.assertEqual(result.loss.shape, ())
|
| 296 |
+
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
|
| 297 |
+
result.loss.backward()
|
| 298 |
+
|
| 299 |
+
def create_and_check_bloom_weight_initialization(self, config, *args):
|
| 300 |
+
model = BloomModel(config)
|
| 301 |
+
model_std = model.config.initializer_range / math.sqrt(2 * model.config.n_layer)
|
| 302 |
+
for key in model.state_dict().keys():
|
| 303 |
+
if "c_proj" in key and "weight" in key:
|
| 304 |
+
self.parent.assertLessEqual(abs(torch.std(model.state_dict()[key]) - model_std), 0.001)
|
| 305 |
+
self.parent.assertLessEqual(abs(torch.mean(model.state_dict()[key]) - 0.0), 0.01)
|
| 306 |
+
|
| 307 |
+
def prepare_config_and_inputs_for_common(self):
|
| 308 |
+
config_and_inputs = self.prepare_config_and_inputs()
|
| 309 |
+
|
| 310 |
+
config, input_ids, input_mask, sequence_labels = config_and_inputs
|
| 311 |
+
|
| 312 |
+
inputs_dict = {"input_ids": input_ids}
|
| 313 |
+
|
| 314 |
+
return config, inputs_dict
|
| 315 |
+
|
| 316 |
+
|
| 317 |
+
@require_torch
|
| 318 |
+
class BloomModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixin, unittest.TestCase):
|
| 319 |
+
all_model_classes = (
|
| 320 |
+
(
|
| 321 |
+
BloomModel,
|
| 322 |
+
BloomForCausalLM,
|
| 323 |
+
BloomForSequenceClassification,
|
| 324 |
+
BloomForTokenClassification,
|
| 325 |
+
BloomForQuestionAnswering,
|
| 326 |
+
)
|
| 327 |
+
if is_torch_available()
|
| 328 |
+
else ()
|
| 329 |
+
)
|
| 330 |
+
|
| 331 |
+
all_generative_model_classes = (BloomForCausalLM,) if is_torch_available() else ()
|
| 332 |
+
pipeline_model_mapping = (
|
| 333 |
+
{
|
| 334 |
+
"feature-extraction": BloomModel,
|
| 335 |
+
"question-answering": BloomForQuestionAnswering,
|
| 336 |
+
"text-classification": BloomForSequenceClassification,
|
| 337 |
+
"text-generation": BloomForCausalLM,
|
| 338 |
+
"token-classification": BloomForTokenClassification,
|
| 339 |
+
"zero-shot": BloomForSequenceClassification,
|
| 340 |
+
}
|
| 341 |
+
if is_torch_available()
|
| 342 |
+
else {}
|
| 343 |
+
)
|
| 344 |
+
fx_compatible = True
|
| 345 |
+
test_missing_keys = False
|
| 346 |
+
test_pruning = False
|
| 347 |
+
test_torchscript = True # torch.autograd functions seems not to be supported
|
| 348 |
+
|
| 349 |
+
def setUp(self):
|
| 350 |
+
self.model_tester = BloomModelTester(self)
|
| 351 |
+
self.config_tester = ConfigTester(self, config_class=BloomConfig, n_embd=37)
|
| 352 |
+
|
| 353 |
+
def test_config(self):
|
| 354 |
+
self.config_tester.run_common_tests()
|
| 355 |
+
|
| 356 |
+
def test_bloom_model(self):
|
| 357 |
+
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
| 358 |
+
self.model_tester.create_and_check_bloom_model(*config_and_inputs)
|
| 359 |
+
|
| 360 |
+
def test_bloom_model_past(self):
|
| 361 |
+
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
| 362 |
+
self.model_tester.create_and_check_bloom_model_past(*config_and_inputs)
|
| 363 |
+
|
| 364 |
+
def test_bloom_model_att_mask_past(self):
|
| 365 |
+
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
| 366 |
+
self.model_tester.create_and_check_bloom_model_attention_mask_past(*config_and_inputs)
|
| 367 |
+
|
| 368 |
+
def test_bloom_model_past_large_inputs(self):
|
| 369 |
+
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
| 370 |
+
self.model_tester.create_and_check_bloom_model_past_large_inputs(*config_and_inputs)
|
| 371 |
+
|
| 372 |
+
def test_bloom_lm_head_model(self):
|
| 373 |
+
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
| 374 |
+
self.model_tester.create_and_check_lm_head_model(*config_and_inputs)
|
| 375 |
+
|
| 376 |
+
def test_bloom_sequence_classification_model(self):
|
| 377 |
+
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
| 378 |
+
self.model_tester.create_and_check_sequence_classification_model(*config_and_inputs)
|
| 379 |
+
|
| 380 |
+
def test_bloom_token_classification_model(self):
|
| 381 |
+
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
| 382 |
+
self.model_tester.create_and_check_token_classification_model(*config_and_inputs)
|
| 383 |
+
|
| 384 |
+
def test_bloom_gradient_checkpointing(self):
|
| 385 |
+
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
| 386 |
+
self.model_tester.create_and_check_forward_and_backwards(*config_and_inputs, gradient_checkpointing=True)
|
| 387 |
+
|
| 388 |
+
def test_bloom_weight_initialization(self):
|
| 389 |
+
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
| 390 |
+
self.model_tester.create_and_check_bloom_weight_initialization(*config_and_inputs)
|
| 391 |
+
|
| 392 |
+
@slow
|
| 393 |
+
def test_model_from_pretrained(self):
|
| 394 |
+
model_name = "bigscience/bigscience-small-testing"
|
| 395 |
+
model = BloomModel.from_pretrained(model_name)
|
| 396 |
+
self.assertIsNotNone(model)
|
| 397 |
+
|
| 398 |
+
@slow
|
| 399 |
+
@require_torch_accelerator
|
| 400 |
+
def test_simple_generation(self):
|
| 401 |
+
# This test is a bit flaky. For some GPU architectures, pytorch sets by default allow_fp16_reduced_precision_reduction = True and some operations
|
| 402 |
+
# do not give the same results under this configuration, especially torch.baddmm and torch.bmm. https://pytorch.org/docs/stable/notes/numerical_accuracy.html#fp16-on-mi200
|
| 403 |
+
# As we leave the default value (True) for allow_fp16_reduced_precision_reduction , the tests failed when running in half-precision with smaller models (560m)
|
| 404 |
+
# Please see: https://pytorch.org/docs/stable/notes/cuda.html#reduced-precision-reduction-in-fp16-gemms
|
| 405 |
+
# This discrepancy is observed only when using small models and seems to be stable for larger models.
|
| 406 |
+
# Our conclusion is that these operations are flaky for small inputs but seems to be stable for larger inputs (for the functions `baddmm` and `bmm`), and therefore for larger models.
|
| 407 |
+
|
| 408 |
+
# Here is a summary of an ablation study of our observations
|
| 409 |
+
# EXPECTED_OUTPUT = "I enjoy walking with my cute dog, and I love to watch the kids play. I am a very active person, and I am a very good listener. I am a very good person, and I am a very good person. I am a"
|
| 410 |
+
# 560m + allow_fp16_reduced_precision_reduction = False + torch.bmm ==> PASS
|
| 411 |
+
# 560m + allow_fp16_reduced_precision_reduction = False + torch.baddm ==> PASS
|
| 412 |
+
# 560m + allow_fp16_reduced_precision_reduction = True + torch.baddm ==> PASS
|
| 413 |
+
# 560m + allow_fp16_reduced_precision_reduction = True + torch.bmm ==> FAIL
|
| 414 |
+
|
| 415 |
+
# EXPECTED_OUTPUT = "I enjoy walking with my cute dog, but I also enjoy hiking, biking, and swimming. I love to cook and bake. I love to cook and bake. I love to cook and bake. I love to cook and bake. I love"
|
| 416 |
+
# >=1b1 + allow_fp16_reduced_precision_reduction = True + torch.baddm ==> PASS (for use_cache=True and use_cache=False)
|
| 417 |
+
# >=1b1 + allow_fp16_reduced_precision_reduction = True + torch.bmm ==> PASS
|
| 418 |
+
# >=1b1 + allow_fp16_reduced_precision_reduction = False + torch.bmm ==> PASS
|
| 419 |
+
|
| 420 |
+
path_560m = "bigscience/bloom-560m"
|
| 421 |
+
model = BloomForCausalLM.from_pretrained(path_560m, use_cache=True, revision="gs555750").to(torch_device)
|
| 422 |
+
model = model.eval()
|
| 423 |
+
tokenizer = BloomTokenizerFast.from_pretrained(path_560m)
|
| 424 |
+
|
| 425 |
+
input_sentence = "I enjoy walking with my cute dog"
|
| 426 |
+
# This output has been obtained using fp32 model on the huggingface DGX workstation - NVIDIA A100 GPU
|
| 427 |
+
EXPECTED_OUTPUT = (
|
| 428 |
+
"I enjoy walking with my cute dog, and I love to watch the kids play with the kids. I am a very "
|
| 429 |
+
"active person, and I enjoy working out, and I am a very active person. I am a very active person, and I"
|
| 430 |
+
)
|
| 431 |
+
|
| 432 |
+
input_ids = tokenizer.encode(input_sentence, return_tensors="pt")
|
| 433 |
+
greedy_output = model.generate(input_ids.to(torch_device), max_length=50)
|
| 434 |
+
|
| 435 |
+
self.assertEqual(tokenizer.decode(greedy_output[0], skip_special_tokens=True), EXPECTED_OUTPUT)
|
| 436 |
+
|
| 437 |
+
@slow
|
| 438 |
+
@require_torch_accelerator
|
| 439 |
+
def test_batch_generation(self):
|
| 440 |
+
path_560m = "bigscience/bloom-560m"
|
| 441 |
+
model = BloomForCausalLM.from_pretrained(path_560m, use_cache=True, revision="gs555750").to(torch_device)
|
| 442 |
+
model = model.eval()
|
| 443 |
+
tokenizer = BloomTokenizerFast.from_pretrained(path_560m, padding_side="left")
|
| 444 |
+
|
| 445 |
+
input_sentence = ["I enjoy walking with my cute dog", "I enjoy walking with my cute dog"]
|
| 446 |
+
|
| 447 |
+
inputs = tokenizer.batch_encode_plus(input_sentence, return_tensors="pt", padding=True)
|
| 448 |
+
input_ids = inputs["input_ids"].to(torch_device)
|
| 449 |
+
attention_mask = inputs["attention_mask"]
|
| 450 |
+
greedy_output = model.generate(input_ids, attention_mask=attention_mask, max_length=50, do_sample=False)
|
| 451 |
+
|
| 452 |
+
self.assertEqual(
|
| 453 |
+
tokenizer.decode(greedy_output[0], skip_special_tokens=True),
|
| 454 |
+
tokenizer.decode(greedy_output[1], skip_special_tokens=True),
|
| 455 |
+
)
|
| 456 |
+
|
| 457 |
+
@slow
|
| 458 |
+
@require_torch_accelerator
|
| 459 |
+
def test_batch_generation_padd(self):
|
| 460 |
+
path_560m = "bigscience/bloom-560m"
|
| 461 |
+
model = BloomForCausalLM.from_pretrained(path_560m, use_cache=True, revision="gs555750").to(torch_device)
|
| 462 |
+
model = model.eval()
|
| 463 |
+
tokenizer = BloomTokenizerFast.from_pretrained(path_560m, padding_side="left")
|
| 464 |
+
|
| 465 |
+
input_sentence = ["I enjoy walking with my cute dog", "Hello my name is"]
|
| 466 |
+
input_sentence_without_pad = "Hello my name is"
|
| 467 |
+
|
| 468 |
+
input_ids = tokenizer.batch_encode_plus(input_sentence, return_tensors="pt", padding=True)
|
| 469 |
+
input_ids_without_pad = tokenizer.encode(input_sentence_without_pad, return_tensors="pt")
|
| 470 |
+
|
| 471 |
+
input_ids, attention_mask = input_ids["input_ids"].to(torch_device), input_ids["attention_mask"]
|
| 472 |
+
greedy_output = model.generate(input_ids, attention_mask=attention_mask, max_length=50, do_sample=False)
|
| 473 |
+
greedy_output_without_pad = model.generate(
|
| 474 |
+
input_ids_without_pad.to(torch_device), max_length=50, do_sample=False
|
| 475 |
+
)
|
| 476 |
+
|
| 477 |
+
# test token values
|
| 478 |
+
self.assertEqual(greedy_output[-1, 3:].tolist(), greedy_output_without_pad[0, :-3].tolist())
|
| 479 |
+
|
| 480 |
+
# test reconstructions
|
| 481 |
+
self.assertEqual(
|
| 482 |
+
tokenizer.decode(greedy_output[-1, 3:], skip_special_tokens=True),
|
| 483 |
+
tokenizer.decode(greedy_output_without_pad[0, :-3], skip_special_tokens=True),
|
| 484 |
+
)
|
| 485 |
+
|
| 486 |
+
@slow
|
| 487 |
+
@require_torch_accelerator
|
| 488 |
+
def test_batch_generated_text(self):
|
| 489 |
+
path_560m = "bigscience/bloom-560m"
|
| 490 |
+
|
| 491 |
+
model = BloomForCausalLM.from_pretrained(path_560m, use_cache=True, revision="gs555750").to(torch_device)
|
| 492 |
+
model = model.eval()
|
| 493 |
+
tokenizer = BloomTokenizerFast.from_pretrained(path_560m, padding_side="left")
|
| 494 |
+
|
| 495 |
+
input_sentences = [
|
| 496 |
+
"Hello what is",
|
| 497 |
+
"Running a quick test with the",
|
| 498 |
+
]
|
| 499 |
+
inputs = tokenizer(input_sentences, return_tensors="pt", padding=True, truncation=True)
|
| 500 |
+
generated_ids = model.generate(
|
| 501 |
+
inputs["input_ids"].to(torch_device), attention_mask=inputs["attention_mask"], max_length=20
|
| 502 |
+
)
|
| 503 |
+
generated_text = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)
|
| 504 |
+
|
| 505 |
+
# these generations match those of the PyTorch model
|
| 506 |
+
EXPECTED_GENERATIONS = [
|
| 507 |
+
"Hello what is the best way to get the data from the server? I have tried",
|
| 508 |
+
"Running a quick test with the following command:\nsudo apt-get install python3\nsudo apt-get install python2",
|
| 509 |
+
]
|
| 510 |
+
|
| 511 |
+
self.assertListEqual(generated_text, EXPECTED_GENERATIONS)
|
| 512 |
+
|
| 513 |
+
@unittest.skip("Bloom needs a 2D attention for alibi")
|
| 514 |
+
def test_custom_4d_attention_mask(self):
|
| 515 |
+
pass
|
| 516 |
+
|
| 517 |
+
|
| 518 |
+
@require_torch
|
| 519 |
+
class BloomEmbeddingTest(unittest.TestCase):
|
| 520 |
+
"""
|
| 521 |
+
The goal here is to compare the embeddings generated by the model trained
|
| 522 |
+
using Megatron-LM with the one from the transformers library, with a small GPT2-like model
|
| 523 |
+
to ensure that the conversion from Megatron-LM to transformers has been done successfully.
|
| 524 |
+
The script compares the logits of the embedding layer and the transformer layers.
|
| 525 |
+
|
| 526 |
+
WARNING: It is expected that these logits will not have exactly the same statistics when running
|
| 527 |
+
the code on CPU or GPU. For more info, please visit:
|
| 528 |
+
- https://github.com/pytorch/pytorch/issues/76052#issuecomment-1103193548
|
| 529 |
+
- https://discuss.pytorch.org/t/reproducibility-issue-between-intel-and-amd-cpus/144779/9
|
| 530 |
+
|
| 531 |
+
|
| 532 |
+
You need to install tokenizers following this readme:
|
| 533 |
+
- https://huggingface.co/bigscience-catalogue-data-dev/byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles
|
| 534 |
+
|
| 535 |
+
Tokenizer used during training:
|
| 536 |
+
- https://huggingface.co/bigscience-catalogue-data-dev/byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles
|
| 537 |
+
|
| 538 |
+
# TODO change the script (or just add skip) when building the env with tokenizers 0.12.0
|
| 539 |
+
"""
|
| 540 |
+
|
| 541 |
+
def setUp(self):
|
| 542 |
+
super().setUp()
|
| 543 |
+
self.path_bigscience_model = "bigscience/bigscience-small-testing"
|
| 544 |
+
|
| 545 |
+
@require_torch
|
| 546 |
+
def test_embeddings(self):
|
| 547 |
+
# The config in this checkpoint has `bfloat16` as `torch_dtype` -> model in `bfloat16`
|
| 548 |
+
model = BloomForCausalLM.from_pretrained(self.path_bigscience_model, torch_dtype="auto")
|
| 549 |
+
model.eval()
|
| 550 |
+
|
| 551 |
+
EMBEDDINGS_DS_BEFORE_LN_BF_16_MEAN = {
|
| 552 |
+
3478: 0.0002307891845703125,
|
| 553 |
+
368: -0.000568389892578125,
|
| 554 |
+
109586: -0.0003910064697265625,
|
| 555 |
+
35433: -0.000194549560546875,
|
| 556 |
+
2: 0.0004138946533203125,
|
| 557 |
+
77: 0.000659942626953125,
|
| 558 |
+
132619: -0.00031280517578125,
|
| 559 |
+
2175: 0.000457763671875,
|
| 560 |
+
23714: 0.000263214111328125,
|
| 561 |
+
73173: -0.000286102294921875,
|
| 562 |
+
144252: 0.00052642822265625,
|
| 563 |
+
}
|
| 564 |
+
EMBEDDINGS_DS_BEFORE_LN_BF_16_MIN = {
|
| 565 |
+
3478: -0.00921630859375,
|
| 566 |
+
368: -0.010009765625,
|
| 567 |
+
109586: -0.01031494140625,
|
| 568 |
+
35433: -0.01177978515625,
|
| 569 |
+
2: -0.0074462890625,
|
| 570 |
+
77: -0.00848388671875,
|
| 571 |
+
132619: -0.009521484375,
|
| 572 |
+
2175: -0.0074462890625,
|
| 573 |
+
23714: -0.0145263671875,
|
| 574 |
+
73173: -0.007415771484375,
|
| 575 |
+
144252: -0.01007080078125,
|
| 576 |
+
}
|
| 577 |
+
EMBEDDINGS_DS_BEFORE_LN_BF_16_MAX = {
|
| 578 |
+
3478: 0.0128173828125,
|
| 579 |
+
368: 0.01214599609375,
|
| 580 |
+
109586: 0.0111083984375,
|
| 581 |
+
35433: 0.01019287109375,
|
| 582 |
+
2: 0.0157470703125,
|
| 583 |
+
77: 0.0174560546875,
|
| 584 |
+
132619: 0.0078125,
|
| 585 |
+
2175: 0.0113525390625,
|
| 586 |
+
23714: 0.0146484375,
|
| 587 |
+
73173: 0.01116943359375,
|
| 588 |
+
144252: 0.01141357421875,
|
| 589 |
+
}
|
| 590 |
+
EMBEDDINGS_DS_BEFORE_LN_BF_16_SUM = {"value": 0.08203125}
|
| 591 |
+
|
| 592 |
+
EMBEDDINGS_DS_BEFORE_LN_F_16_MEAN = {
|
| 593 |
+
132619: -0.00031256675720214844,
|
| 594 |
+
3478: 0.00023090839385986328,
|
| 595 |
+
368: -0.0005702972412109375,
|
| 596 |
+
109586: -0.00039124488830566406,
|
| 597 |
+
35433: -0.000194549560546875,
|
| 598 |
+
2: 0.0004146099090576172,
|
| 599 |
+
2175: 0.0004572868347167969,
|
| 600 |
+
23714: 0.00026416778564453125,
|
| 601 |
+
73173: -0.0002865791320800781,
|
| 602 |
+
144252: 0.0005254745483398438,
|
| 603 |
+
77: 0.0006618499755859375,
|
| 604 |
+
}
|
| 605 |
+
EMBEDDINGS_DS_BEFORE_LN_F_16_MIN = {
|
| 606 |
+
3478: -0.00921630859375,
|
| 607 |
+
368: -0.010009765625,
|
| 608 |
+
109586: -0.01031494140625,
|
| 609 |
+
35433: -0.01177978515625,
|
| 610 |
+
2: -0.0074462890625,
|
| 611 |
+
77: -0.00848388671875,
|
| 612 |
+
132619: -0.009521484375,
|
| 613 |
+
2175: -0.0074462890625,
|
| 614 |
+
23714: -0.0145263671875,
|
| 615 |
+
73173: -0.007415771484375,
|
| 616 |
+
144252: -0.01007080078125,
|
| 617 |
+
}
|
| 618 |
+
EMBEDDINGS_DS_BEFORE_LN_F_16_MAX = {
|
| 619 |
+
3478: 0.0128173828125,
|
| 620 |
+
368: 0.01214599609375,
|
| 621 |
+
109586: 0.0111083984375,
|
| 622 |
+
35433: 0.01019287109375,
|
| 623 |
+
2: 0.0157470703125,
|
| 624 |
+
77: 0.0174560546875,
|
| 625 |
+
132619: 0.0078125,
|
| 626 |
+
2175: 0.0113525390625,
|
| 627 |
+
23714: 0.0146484375,
|
| 628 |
+
73173: 0.01116943359375,
|
| 629 |
+
144252: 0.01141357421875,
|
| 630 |
+
}
|
| 631 |
+
EMBEDDINGS_DS_BEFORE_LN_F_16_SUM = {"value": 0.0821533203125}
|
| 632 |
+
|
| 633 |
+
EMBEDDINGS_DS_BEFORE_LN_F_32_MEAN = {
|
| 634 |
+
132619: -0.00031267106533050537,
|
| 635 |
+
3478: 0.00023087859153747559,
|
| 636 |
+
368: -0.0005701072514057159,
|
| 637 |
+
109586: -0.0003911703824996948,
|
| 638 |
+
35433: -0.0001944899559020996,
|
| 639 |
+
2: 0.0004146844148635864,
|
| 640 |
+
2175: 0.00045740045607089996,
|
| 641 |
+
23714: 0.0002641640603542328,
|
| 642 |
+
73173: -0.0002864748239517212,
|
| 643 |
+
144252: 0.0005256589502096176,
|
| 644 |
+
77: 0.0006617321632802486,
|
| 645 |
+
}
|
| 646 |
+
EMBEDDINGS_DS_BEFORE_LN_F_32_MIN = {
|
| 647 |
+
3478: -0.00921630859375,
|
| 648 |
+
368: -0.010009765625,
|
| 649 |
+
109586: -0.01031494140625,
|
| 650 |
+
35433: -0.01177978515625,
|
| 651 |
+
2: -0.0074462890625,
|
| 652 |
+
77: -0.00848388671875,
|
| 653 |
+
132619: -0.009521484375,
|
| 654 |
+
2175: -0.0074462890625,
|
| 655 |
+
23714: -0.0145263671875,
|
| 656 |
+
73173: -0.007415771484375,
|
| 657 |
+
144252: -0.01007080078125,
|
| 658 |
+
}
|
| 659 |
+
EMBEDDINGS_DS_BEFORE_LN_F_32_MAX = {
|
| 660 |
+
3478: 0.0128173828125,
|
| 661 |
+
368: 0.01214599609375,
|
| 662 |
+
109586: 0.0111083984375,
|
| 663 |
+
35433: 0.01019287109375,
|
| 664 |
+
2: 0.0157470703125,
|
| 665 |
+
77: 0.0174560546875,
|
| 666 |
+
132619: 0.0078125,
|
| 667 |
+
2175: 0.0113525390625,
|
| 668 |
+
23714: 0.0146484375,
|
| 669 |
+
73173: 0.01116943359375,
|
| 670 |
+
144252: 0.01141357421875,
|
| 671 |
+
}
|
| 672 |
+
EMBEDDINGS_DS_BEFORE_LN_F_32_SUM = {"value": 0.08217757940292358}
|
| 673 |
+
|
| 674 |
+
TEST_EMBEDDINGS = {
|
| 675 |
+
"torch.bfloat16": {
|
| 676 |
+
"mean": EMBEDDINGS_DS_BEFORE_LN_BF_16_MEAN,
|
| 677 |
+
"max": EMBEDDINGS_DS_BEFORE_LN_BF_16_MAX,
|
| 678 |
+
"min": EMBEDDINGS_DS_BEFORE_LN_BF_16_MIN,
|
| 679 |
+
"sum": EMBEDDINGS_DS_BEFORE_LN_BF_16_SUM,
|
| 680 |
+
},
|
| 681 |
+
"torch.float32": {
|
| 682 |
+
"mean": EMBEDDINGS_DS_BEFORE_LN_F_32_MEAN,
|
| 683 |
+
"max": EMBEDDINGS_DS_BEFORE_LN_F_32_MAX,
|
| 684 |
+
"min": EMBEDDINGS_DS_BEFORE_LN_F_32_MIN,
|
| 685 |
+
"sum": EMBEDDINGS_DS_BEFORE_LN_F_32_SUM,
|
| 686 |
+
},
|
| 687 |
+
"torch.float": {
|
| 688 |
+
"mean": EMBEDDINGS_DS_BEFORE_LN_F_32_MEAN,
|
| 689 |
+
"max": EMBEDDINGS_DS_BEFORE_LN_F_32_MAX,
|
| 690 |
+
"min": EMBEDDINGS_DS_BEFORE_LN_F_32_MIN,
|
| 691 |
+
"sum": EMBEDDINGS_DS_BEFORE_LN_F_32_SUM,
|
| 692 |
+
},
|
| 693 |
+
"torch.float16": {
|
| 694 |
+
"mean": EMBEDDINGS_DS_BEFORE_LN_F_16_MEAN,
|
| 695 |
+
"max": EMBEDDINGS_DS_BEFORE_LN_F_16_MAX,
|
| 696 |
+
"min": EMBEDDINGS_DS_BEFORE_LN_F_16_MIN,
|
| 697 |
+
"sum": EMBEDDINGS_DS_BEFORE_LN_F_16_SUM,
|
| 698 |
+
},
|
| 699 |
+
}
|
| 700 |
+
|
| 701 |
+
EXAMPLE_IDS = [3478, 368, 109586, 35433, 2, 77, 132619, 3478, 368, 109586, 35433, 2, 2175, 23714, 73173, 144252, 2, 77, 132619, 3478] # fmt: skip
|
| 702 |
+
|
| 703 |
+
EMBEDDINGS_DS_AFTER_LN_MEAN = {
|
| 704 |
+
3478: -6.580352783203125e-05,
|
| 705 |
+
368: 0.0001316070556640625,
|
| 706 |
+
109586: -0.00030517578125,
|
| 707 |
+
35433: 4.00543212890625e-05,
|
| 708 |
+
2: -7.2479248046875e-05,
|
| 709 |
+
77: -8.96453857421875e-05,
|
| 710 |
+
132619: 0.0001583099365234375,
|
| 711 |
+
2175: 2.1219253540039062e-05,
|
| 712 |
+
23714: -0.000247955322265625,
|
| 713 |
+
73173: -0.00021839141845703125,
|
| 714 |
+
144252: -0.0001430511474609375,
|
| 715 |
+
}
|
| 716 |
+
EMBEDDINGS_DS_AFTER_LN_MIN = {
|
| 717 |
+
3478: -1.6953125,
|
| 718 |
+
368: -1.6875,
|
| 719 |
+
109586: -1.6875,
|
| 720 |
+
35433: -2.125,
|
| 721 |
+
2: -1.390625,
|
| 722 |
+
77: -1.5390625,
|
| 723 |
+
132619: -1.875,
|
| 724 |
+
2175: -1.4609375,
|
| 725 |
+
23714: -2.296875,
|
| 726 |
+
73173: -1.3515625,
|
| 727 |
+
144252: -1.78125,
|
| 728 |
+
}
|
| 729 |
+
EMBEDDINGS_DS_AFTER_LN_MAX = {
|
| 730 |
+
3478: 2.265625,
|
| 731 |
+
368: 2.28125,
|
| 732 |
+
109586: 1.953125,
|
| 733 |
+
35433: 1.90625,
|
| 734 |
+
2: 2.703125,
|
| 735 |
+
77: 2.828125,
|
| 736 |
+
132619: 1.65625,
|
| 737 |
+
2175: 2.015625,
|
| 738 |
+
23714: 2.234375,
|
| 739 |
+
73173: 2.171875,
|
| 740 |
+
144252: 1.828125,
|
| 741 |
+
}
|
| 742 |
+
|
| 743 |
+
EMBEDDINGS_DS_AFTER_LN = {
|
| 744 |
+
"mean": EMBEDDINGS_DS_AFTER_LN_MEAN,
|
| 745 |
+
"min": EMBEDDINGS_DS_AFTER_LN_MIN,
|
| 746 |
+
"max": EMBEDDINGS_DS_AFTER_LN_MAX,
|
| 747 |
+
}
|
| 748 |
+
|
| 749 |
+
tensor_ids = torch.LongTensor([EXAMPLE_IDS])
|
| 750 |
+
with torch.no_grad():
|
| 751 |
+
embeddings = model.transformer.word_embeddings(tensor_ids)
|
| 752 |
+
embeddings_ln = model.transformer.word_embeddings_layernorm(embeddings) #
|
| 753 |
+
# first check the embeddings before LN
|
| 754 |
+
output_dict = {"min": {}, "max": {}, "mean": {}, "sum": {"value": embeddings.sum().item()}}
|
| 755 |
+
for i, idx in enumerate(EXAMPLE_IDS):
|
| 756 |
+
output_dict["min"][idx] = embeddings.min(dim=-1).values[0][i].item()
|
| 757 |
+
output_dict["max"][idx] = embeddings.max(dim=-1).values[0][i].item()
|
| 758 |
+
output_dict["mean"][idx] = embeddings.mean(dim=-1)[0][i].item()
|
| 759 |
+
|
| 760 |
+
for key in TEST_EMBEDDINGS[str(model.dtype)].keys():
|
| 761 |
+
self.assertDictEqual(TEST_EMBEDDINGS[str(model.dtype)][key], output_dict[key])
|
| 762 |
+
|
| 763 |
+
output_dict_norm = {"min": {}, "max": {}, "mean": {}}
|
| 764 |
+
for i, idx in enumerate(EXAMPLE_IDS):
|
| 765 |
+
output_dict_norm["min"][idx] = embeddings_ln.min(dim=-1).values[0][i].item()
|
| 766 |
+
output_dict_norm["max"][idx] = embeddings_ln.max(dim=-1).values[0][i].item()
|
| 767 |
+
output_dict_norm["mean"][idx] = embeddings_ln.mean(dim=-1)[0][i].item()
|
| 768 |
+
|
| 769 |
+
# This test does not pass when places = 2
|
| 770 |
+
for i, key in enumerate(output_dict_norm.keys()):
|
| 771 |
+
for j, idx in enumerate(output_dict[key].keys()):
|
| 772 |
+
self.assertAlmostEqual(EMBEDDINGS_DS_AFTER_LN[key][idx], output_dict_norm[key][idx], places=1)
|
| 773 |
+
|
| 774 |
+
@require_torch
|
| 775 |
+
def test_hidden_states_transformers(self):
|
| 776 |
+
cuda_available = torch.cuda.is_available()
|
| 777 |
+
model = BloomModel.from_pretrained(self.path_bigscience_model, use_cache=False, torch_dtype="auto").to(
|
| 778 |
+
torch_device
|
| 779 |
+
)
|
| 780 |
+
model.eval()
|
| 781 |
+
|
| 782 |
+
EXAMPLE_IDS = [3478, 368, 109586, 35433, 2, 77, 132619, 3478, 368, 109586, 35433, 2, 2175, 23714, 73173, 144252, 2, 77, 132619, 3478] # fmt: skip
|
| 783 |
+
|
| 784 |
+
MEAN_VALUE_LAST_LM = -4.3392181396484375e-05
|
| 785 |
+
MIN_MAX_DICT = {"min": -2.0625, "max": 2.75}
|
| 786 |
+
tensor_ids = torch.LongTensor([EXAMPLE_IDS])
|
| 787 |
+
|
| 788 |
+
with torch.no_grad():
|
| 789 |
+
logits = model(tensor_ids.to(torch_device))
|
| 790 |
+
output_dict = {
|
| 791 |
+
"min": logits.last_hidden_state.min(dim=-1).values[0][0].item(),
|
| 792 |
+
"max": logits.last_hidden_state.max(dim=-1).values[0][0].item(),
|
| 793 |
+
}
|
| 794 |
+
|
| 795 |
+
if cuda_available:
|
| 796 |
+
self.assertAlmostEqual(MEAN_VALUE_LAST_LM, logits.last_hidden_state.mean().item(), places=4)
|
| 797 |
+
else:
|
| 798 |
+
self.assertAlmostEqual(MEAN_VALUE_LAST_LM, logits.last_hidden_state.mean().item(), places=3)
|
| 799 |
+
|
| 800 |
+
self.assertDictEqual(MIN_MAX_DICT, output_dict)
|
| 801 |
+
|
| 802 |
+
@require_torch
|
| 803 |
+
def test_logits(self):
|
| 804 |
+
cuda_available = torch.cuda.is_available()
|
| 805 |
+
model = BloomForCausalLM.from_pretrained(self.path_bigscience_model, use_cache=False, torch_dtype="auto").to(
|
| 806 |
+
torch_device
|
| 807 |
+
) # load in bf16
|
| 808 |
+
model.eval()
|
| 809 |
+
|
| 810 |
+
EXAMPLE_IDS = [3478, 368, 109586, 35433, 2, 77, 132619, 3478, 368, 109586, 35433, 2, 2175, 23714, 73173, 144252, 2, 77, 132619, 3478] # fmt: skip
|
| 811 |
+
|
| 812 |
+
MEAN_LOGITS_GPU_1 = -1.823902130126953e-05
|
| 813 |
+
MEAN_LOGITS_GPU_2 = 1.9431114196777344e-05
|
| 814 |
+
|
| 815 |
+
tensor_ids = torch.LongTensor([EXAMPLE_IDS]).to(torch_device)
|
| 816 |
+
with torch.no_grad():
|
| 817 |
+
output = model(tensor_ids).logits
|
| 818 |
+
|
| 819 |
+
output_gpu_1, output_gpu_2 = output.split(125440, dim=-1)
|
| 820 |
+
if cuda_available:
|
| 821 |
+
self.assertAlmostEqual(output_gpu_1.mean().item(), MEAN_LOGITS_GPU_1, places=6)
|
| 822 |
+
self.assertAlmostEqual(output_gpu_2.mean().item(), MEAN_LOGITS_GPU_2, places=6)
|
| 823 |
+
else:
|
| 824 |
+
self.assertAlmostEqual(output_gpu_1.mean().item(), MEAN_LOGITS_GPU_1, places=6) # 1e-06 precision!!
|
| 825 |
+
self.assertAlmostEqual(output_gpu_2.mean().item(), MEAN_LOGITS_GPU_2, places=6)
|
tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/bloom/test_modeling_flax_bloom.py
ADDED
|
@@ -0,0 +1,251 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Copyright 2023 The HuggingFace Team. All rights reserved.
|
| 2 |
+
#
|
| 3 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
| 4 |
+
# you may not use this file except in compliance with the License.
|
| 5 |
+
# You may obtain a copy of the License at
|
| 6 |
+
#
|
| 7 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
| 8 |
+
#
|
| 9 |
+
# Unless required by applicable law or agreed to in writing, software
|
| 10 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
| 11 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
| 12 |
+
# See the License for the specific language governing permissions and
|
| 13 |
+
# limitations under the License.
|
| 14 |
+
import unittest
|
| 15 |
+
|
| 16 |
+
import numpy as np
|
| 17 |
+
|
| 18 |
+
from transformers import BloomConfig, BloomTokenizerFast, is_flax_available
|
| 19 |
+
from transformers.testing_utils import require_flax, slow
|
| 20 |
+
|
| 21 |
+
from ...generation.test_flax_utils import FlaxGenerationTesterMixin
|
| 22 |
+
from ...test_modeling_flax_common import FlaxModelTesterMixin, ids_tensor
|
| 23 |
+
|
| 24 |
+
|
| 25 |
+
if is_flax_available():
|
| 26 |
+
import os
|
| 27 |
+
|
| 28 |
+
# The slow tests are often failing with OOM error on GPU
|
| 29 |
+
# This makes JAX allocate exactly what is needed on demand, and deallocate memory that is no longer needed
|
| 30 |
+
# but will be slower as stated here https://jax.readthedocs.io/en/latest/gpu_memory_allocation.html
|
| 31 |
+
os.environ["XLA_PYTHON_CLIENT_ALLOCATOR"] = "platform"
|
| 32 |
+
|
| 33 |
+
import jax.numpy as jnp
|
| 34 |
+
|
| 35 |
+
from transformers import FlaxBloomForCausalLM, FlaxBloomModel
|
| 36 |
+
|
| 37 |
+
|
| 38 |
+
def prepare_bloom_inputs_dict(config, input_ids, attention_mask=None):
|
| 39 |
+
if attention_mask is None:
|
| 40 |
+
attention_mask = np.where(input_ids != config.pad_token_id, 1, 0)
|
| 41 |
+
return {"input_ids": input_ids, "attention_mask": attention_mask}
|
| 42 |
+
|
| 43 |
+
|
| 44 |
+
@require_flax
|
| 45 |
+
class FlaxBloomModelTester:
|
| 46 |
+
def __init__(
|
| 47 |
+
self,
|
| 48 |
+
parent,
|
| 49 |
+
batch_size=13,
|
| 50 |
+
seq_length=7,
|
| 51 |
+
is_training=True,
|
| 52 |
+
use_labels=False,
|
| 53 |
+
vocab_size=99,
|
| 54 |
+
hidden_size=16,
|
| 55 |
+
n_layer=2,
|
| 56 |
+
n_head=4,
|
| 57 |
+
hidden_act="gelu",
|
| 58 |
+
hidden_dropout=0.1,
|
| 59 |
+
attention_probs_dropout_prob=0.1,
|
| 60 |
+
eos_token_id=2,
|
| 61 |
+
pad_token_id=1,
|
| 62 |
+
bos_token_id=0,
|
| 63 |
+
initializer_range=0.02,
|
| 64 |
+
apply_residual_connection_post_layernorm=False,
|
| 65 |
+
):
|
| 66 |
+
self.parent = parent
|
| 67 |
+
self.batch_size = batch_size
|
| 68 |
+
self.seq_length = seq_length
|
| 69 |
+
self.is_training = is_training
|
| 70 |
+
self.use_labels = use_labels
|
| 71 |
+
self.vocab_size = vocab_size
|
| 72 |
+
self.hidden_size = hidden_size
|
| 73 |
+
self.num_hidden_layers = n_layer
|
| 74 |
+
self.num_attention_heads = n_head
|
| 75 |
+
self.hidden_act = hidden_act
|
| 76 |
+
self.hidden_dropout = hidden_dropout
|
| 77 |
+
self.attention_probs_dropout_prob = attention_probs_dropout_prob
|
| 78 |
+
self.eos_token_id = eos_token_id
|
| 79 |
+
self.pad_token_id = pad_token_id
|
| 80 |
+
self.bos_token_id = bos_token_id
|
| 81 |
+
self.initializer_range = initializer_range
|
| 82 |
+
self.is_encoder_decoder = False
|
| 83 |
+
self.apply_residual_connection_post_layernorm = apply_residual_connection_post_layernorm
|
| 84 |
+
|
| 85 |
+
def prepare_config_and_inputs(self):
|
| 86 |
+
input_ids = np.clip(ids_tensor([self.batch_size, self.seq_length - 1], self.vocab_size), 3, self.vocab_size)
|
| 87 |
+
input_ids = np.concatenate((input_ids, 2 * np.ones((self.batch_size, 1), dtype=np.int64)), -1)
|
| 88 |
+
|
| 89 |
+
config = BloomConfig(
|
| 90 |
+
vocab_size=self.vocab_size,
|
| 91 |
+
hidden_size=self.hidden_size,
|
| 92 |
+
n_layer=self.num_hidden_layers,
|
| 93 |
+
n_head=self.num_attention_heads,
|
| 94 |
+
hidden_dropout=self.hidden_dropout,
|
| 95 |
+
attention_dropout=self.attention_probs_dropout_prob,
|
| 96 |
+
eos_token_id=self.eos_token_id,
|
| 97 |
+
bos_token_id=self.bos_token_id,
|
| 98 |
+
pad_token_id=self.pad_token_id,
|
| 99 |
+
is_encoder_decoder=False,
|
| 100 |
+
use_cache=False,
|
| 101 |
+
)
|
| 102 |
+
inputs_dict = prepare_bloom_inputs_dict(config, input_ids)
|
| 103 |
+
return config, inputs_dict
|
| 104 |
+
|
| 105 |
+
def prepare_config_and_inputs_for_common(self):
|
| 106 |
+
config, inputs_dict = self.prepare_config_and_inputs()
|
| 107 |
+
return config, inputs_dict
|
| 108 |
+
|
| 109 |
+
def check_use_cache_forward(self, model_class_name, config, inputs_dict):
|
| 110 |
+
max_length = 20
|
| 111 |
+
model = model_class_name(config)
|
| 112 |
+
|
| 113 |
+
input_ids = inputs_dict["input_ids"]
|
| 114 |
+
attention_mask = jnp.ones((input_ids.shape[0], max_length), dtype="i4")
|
| 115 |
+
|
| 116 |
+
past_key_values = model.init_cache(input_ids.shape[0], max_length)
|
| 117 |
+
|
| 118 |
+
outputs_cache = model(
|
| 119 |
+
input_ids[:, :-1],
|
| 120 |
+
attention_mask=attention_mask,
|
| 121 |
+
past_key_values=past_key_values,
|
| 122 |
+
)
|
| 123 |
+
|
| 124 |
+
outputs_cache_next = model(
|
| 125 |
+
input_ids[:, -1:],
|
| 126 |
+
attention_mask=attention_mask,
|
| 127 |
+
past_key_values=outputs_cache.past_key_values,
|
| 128 |
+
)
|
| 129 |
+
|
| 130 |
+
outputs = model(input_ids)
|
| 131 |
+
|
| 132 |
+
diff = np.max(np.abs((outputs_cache_next[0][:, -1, :5] - outputs[0][:, -1, :5])))
|
| 133 |
+
self.parent.assertTrue(diff < 1e-3, msg=f"Max diff is {diff}")
|
| 134 |
+
|
| 135 |
+
def check_use_cache_forward_with_attn_mask(self, model_class_name, config, inputs_dict):
|
| 136 |
+
max_length = 20
|
| 137 |
+
model = model_class_name(config)
|
| 138 |
+
|
| 139 |
+
input_ids, attention_mask = (
|
| 140 |
+
inputs_dict["input_ids"],
|
| 141 |
+
inputs_dict["attention_mask"],
|
| 142 |
+
)
|
| 143 |
+
|
| 144 |
+
attention_mask_cache = jnp.concatenate(
|
| 145 |
+
[
|
| 146 |
+
attention_mask,
|
| 147 |
+
jnp.zeros((attention_mask.shape[0], max_length - attention_mask.shape[1])),
|
| 148 |
+
],
|
| 149 |
+
axis=-1,
|
| 150 |
+
)
|
| 151 |
+
|
| 152 |
+
past_key_values = model.init_cache(input_ids.shape[0], max_length)
|
| 153 |
+
|
| 154 |
+
outputs_cache = model(
|
| 155 |
+
input_ids[:, :-1],
|
| 156 |
+
attention_mask=attention_mask_cache,
|
| 157 |
+
past_key_values=past_key_values,
|
| 158 |
+
)
|
| 159 |
+
outputs_cache_next = model(
|
| 160 |
+
input_ids[:, -1:],
|
| 161 |
+
past_key_values=outputs_cache.past_key_values,
|
| 162 |
+
attention_mask=attention_mask_cache,
|
| 163 |
+
)
|
| 164 |
+
|
| 165 |
+
outputs = model(input_ids, attention_mask=attention_mask)
|
| 166 |
+
|
| 167 |
+
diff = np.max(np.abs((outputs_cache_next[0][:, -1, :5] - outputs[0][:, -1, :5])))
|
| 168 |
+
self.parent.assertTrue(diff < 1e-3, msg=f"Max diff is {diff}")
|
| 169 |
+
|
| 170 |
+
|
| 171 |
+
@require_flax
|
| 172 |
+
class FlaxBloomModelTest(FlaxModelTesterMixin, unittest.TestCase, FlaxGenerationTesterMixin):
|
| 173 |
+
all_model_classes = (FlaxBloomModel, FlaxBloomForCausalLM) if is_flax_available() else ()
|
| 174 |
+
all_generative_model_classes = () if is_flax_available() else ()
|
| 175 |
+
|
| 176 |
+
def setUp(self):
|
| 177 |
+
self.model_tester = FlaxBloomModelTester(self)
|
| 178 |
+
|
| 179 |
+
def test_use_cache_forward(self):
|
| 180 |
+
config, inputs_dict = self.model_tester.prepare_config_and_inputs()
|
| 181 |
+
for model_class in self.all_model_classes:
|
| 182 |
+
self.model_tester.check_use_cache_forward(model_class, config, inputs_dict)
|
| 183 |
+
|
| 184 |
+
def test_use_cache_forward_with_attn_mask(self):
|
| 185 |
+
config, inputs_dict = self.model_tester.prepare_config_and_inputs()
|
| 186 |
+
for model_class in self.all_model_classes:
|
| 187 |
+
self.model_tester.check_use_cache_forward_with_attn_mask(model_class, config, inputs_dict)
|
| 188 |
+
|
| 189 |
+
@slow
|
| 190 |
+
def test_model_from_pretrained(self):
|
| 191 |
+
for model_class_name in self.all_model_classes:
|
| 192 |
+
model = model_class_name.from_pretrained("bigscience/bloom-560m")
|
| 193 |
+
input_ids = np.ones((1, 1)) * model.config.eos_token_id
|
| 194 |
+
outputs = model(input_ids)
|
| 195 |
+
self.assertIsNotNone(outputs)
|
| 196 |
+
|
| 197 |
+
|
| 198 |
+
@slow
|
| 199 |
+
@require_flax
|
| 200 |
+
class FlaxBloomGenerationTest(unittest.TestCase):
|
| 201 |
+
all_model_classes = (FlaxBloomForCausalLM,) if is_flax_available() else ()
|
| 202 |
+
all_generative_model_classes = () if is_flax_available() else ()
|
| 203 |
+
|
| 204 |
+
def setUp(self):
|
| 205 |
+
self.model_id = "bigscience/bloom-560m"
|
| 206 |
+
self.tokenizer = BloomTokenizerFast.from_pretrained(self.model_id, padding_side="left")
|
| 207 |
+
self.model_tester = FlaxBloomModelTester(self)
|
| 208 |
+
self.model = FlaxBloomForCausalLM.from_pretrained(self.model_id, from_pt=True, revision="gs555750")
|
| 209 |
+
|
| 210 |
+
def test_model_batched_gen(self):
|
| 211 |
+
# tests if the model outputs the same generation for the same batched input
|
| 212 |
+
input_sentences = [
|
| 213 |
+
"Hello there is this string is definitely longer I believe that",
|
| 214 |
+
"Hello there is this string is definitely longer I believe that",
|
| 215 |
+
]
|
| 216 |
+
inputs = self.tokenizer(input_sentences, return_tensors="np", padding=True, truncation=True)
|
| 217 |
+
sequences_fx = self.model.generate(**inputs, max_length=20).sequences
|
| 218 |
+
self.assertEqual(sequences_fx[0].tolist(), sequences_fx[1].tolist())
|
| 219 |
+
|
| 220 |
+
def test_model_batched_padding_left(self):
|
| 221 |
+
# tests if the model outputs the same generation for an input that is part of a batch
|
| 222 |
+
# and a single input
|
| 223 |
+
input_sentences_batch = [
|
| 224 |
+
"Hello there is this string is definitely longer I believe that",
|
| 225 |
+
"Hi I want to order",
|
| 226 |
+
]
|
| 227 |
+
inputs = self.tokenizer(input_sentences_batch, return_tensors="np", padding=True, truncation=True)
|
| 228 |
+
sequences_fx_batch = self.model.generate(**inputs, max_length=20).sequences
|
| 229 |
+
|
| 230 |
+
input_sentence_simple = "Hi I want to order"
|
| 231 |
+
inputs_simple = self.tokenizer(input_sentence_simple, return_tensors="np")
|
| 232 |
+
sequences_fx_simple = self.model.generate(**inputs_simple, max_length=20).sequences
|
| 233 |
+
|
| 234 |
+
self.assertEqual(sequences_fx_batch[1][6:].tolist(), sequences_fx_simple[0][:-6].tolist())
|
| 235 |
+
|
| 236 |
+
def test_batch_generated_text(self):
|
| 237 |
+
input_sentences = [
|
| 238 |
+
"Hello what is",
|
| 239 |
+
"Running a quick test with the",
|
| 240 |
+
]
|
| 241 |
+
inputs = self.tokenizer(input_sentences, return_tensors="np", padding=True, truncation=True)
|
| 242 |
+
generated_ids = self.model.generate(**inputs, max_length=20).sequences
|
| 243 |
+
generated_text = self.tokenizer.batch_decode(generated_ids, skip_special_tokens=True)
|
| 244 |
+
|
| 245 |
+
# these generations match those of the PyTorch model, ensuring correctness
|
| 246 |
+
EXPECTED_GENERATIONS = [
|
| 247 |
+
"Hello what is the best way to get the data from the server? I have tried",
|
| 248 |
+
"Running a quick test with the following command:\nsudo apt-get install python3\nsudo apt-get install python2",
|
| 249 |
+
]
|
| 250 |
+
|
| 251 |
+
self.assertListEqual(generated_text, EXPECTED_GENERATIONS)
|
tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/bloom/test_tokenization_bloom.py
ADDED
|
@@ -0,0 +1,162 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# coding=utf-8
|
| 2 |
+
# Copyright 2022 The HuggingFace Team. All rights reserved.
|
| 3 |
+
#
|
| 4 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
| 5 |
+
# you may not use this file except in compliance with the License.
|
| 6 |
+
# You may obtain a copy of the License at
|
| 7 |
+
#
|
| 8 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
| 9 |
+
#
|
| 10 |
+
# Unless required by applicable law or agreed to in writing, software
|
| 11 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
| 12 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
| 13 |
+
# See the License for the specific language governing permissions and
|
| 14 |
+
# limitations under the License.
|
| 15 |
+
|
| 16 |
+
import unittest
|
| 17 |
+
|
| 18 |
+
from datasets import load_dataset
|
| 19 |
+
|
| 20 |
+
from transformers import BloomTokenizerFast
|
| 21 |
+
from transformers.testing_utils import require_jinja, require_tokenizers
|
| 22 |
+
|
| 23 |
+
from ...test_tokenization_common import TokenizerTesterMixin
|
| 24 |
+
|
| 25 |
+
|
| 26 |
+
@require_tokenizers
|
| 27 |
+
class BloomTokenizationTest(TokenizerTesterMixin, unittest.TestCase):
|
| 28 |
+
from_pretrained_id = "bigscience/tokenizer"
|
| 29 |
+
slow_tokenizer_class = None
|
| 30 |
+
rust_tokenizer_class = BloomTokenizerFast
|
| 31 |
+
tokenizer_class = BloomTokenizerFast
|
| 32 |
+
test_rust_tokenizer = True
|
| 33 |
+
test_slow_tokenizer = False
|
| 34 |
+
from_pretrained_vocab_key = "tokenizer_file"
|
| 35 |
+
special_tokens_map = {"bos_token": "<s>", "eos_token": "</s>", "unk_token": "<unk>", "pad_token": "<pad>"}
|
| 36 |
+
|
| 37 |
+
def setUp(self):
|
| 38 |
+
super().setUp()
|
| 39 |
+
tokenizer = BloomTokenizerFast.from_pretrained("bigscience/tokenizer")
|
| 40 |
+
tokenizer.save_pretrained(self.tmpdirname)
|
| 41 |
+
|
| 42 |
+
def get_rust_tokenizer(self, **kwargs):
|
| 43 |
+
kwargs.update(self.special_tokens_map)
|
| 44 |
+
return BloomTokenizerFast.from_pretrained(self.tmpdirname, **kwargs)
|
| 45 |
+
|
| 46 |
+
@unittest.skip(reason="This needs a slow tokenizer. Bloom does not have one!")
|
| 47 |
+
def test_encode_decode_with_spaces(self):
|
| 48 |
+
return
|
| 49 |
+
|
| 50 |
+
def test_encodings_from_sample_data(self):
|
| 51 |
+
"""
|
| 52 |
+
Assert that the created tokens are the same than the hard-coded ones
|
| 53 |
+
"""
|
| 54 |
+
tokenizer = self.get_rust_tokenizer()
|
| 55 |
+
|
| 56 |
+
INPUT_SENTENCES = ["The quick brown fox</s>", "jumps over the lazy dog</s>"]
|
| 57 |
+
TARGET_TOKENS = [[2175, 23714, 73173, 144252, 2], [77, 132619, 3478, 368, 109586, 35433, 2]]
|
| 58 |
+
|
| 59 |
+
computed_tokens = tokenizer.batch_encode_plus(INPUT_SENTENCES)["input_ids"]
|
| 60 |
+
self.assertListEqual(TARGET_TOKENS, computed_tokens)
|
| 61 |
+
|
| 62 |
+
decoded_tokens = tokenizer.batch_decode(computed_tokens)
|
| 63 |
+
self.assertListEqual(decoded_tokens, INPUT_SENTENCES)
|
| 64 |
+
|
| 65 |
+
def test_padding(self, max_length=6):
|
| 66 |
+
for tokenizer, pretrained_name, kwargs in self.tokenizers_list:
|
| 67 |
+
with self.subTest(f"{tokenizer.__class__.__name__} ({pretrained_name})"):
|
| 68 |
+
tokenizer_r = self.rust_tokenizer_class.from_pretrained(pretrained_name, **kwargs)
|
| 69 |
+
# tokenizer_r.pad_token = None # Hotfixing padding = None
|
| 70 |
+
# Simple input
|
| 71 |
+
s = "This is a simple input"
|
| 72 |
+
s2 = ["This is a simple input 1", "This is a simple input 2"]
|
| 73 |
+
p = ("This is a simple input", "This is a pair")
|
| 74 |
+
p2 = [
|
| 75 |
+
("This is a simple input 1", "This is a simple input 2"),
|
| 76 |
+
("This is a simple pair 1", "This is a simple pair 2"),
|
| 77 |
+
]
|
| 78 |
+
|
| 79 |
+
# Simple input tests
|
| 80 |
+
try:
|
| 81 |
+
tokenizer_r.encode(s, max_length=max_length)
|
| 82 |
+
tokenizer_r.encode_plus(s, max_length=max_length)
|
| 83 |
+
|
| 84 |
+
tokenizer_r.batch_encode_plus(s2, max_length=max_length)
|
| 85 |
+
tokenizer_r.encode(p, max_length=max_length)
|
| 86 |
+
tokenizer_r.batch_encode_plus(p2, max_length=max_length)
|
| 87 |
+
except ValueError:
|
| 88 |
+
self.fail("Bloom Tokenizer should be able to deal with padding")
|
| 89 |
+
|
| 90 |
+
tokenizer_r.pad_token = None # Hotfixing padding = None
|
| 91 |
+
self.assertRaises(ValueError, tokenizer_r.encode, s, max_length=max_length, padding="max_length")
|
| 92 |
+
|
| 93 |
+
# Simple input
|
| 94 |
+
self.assertRaises(ValueError, tokenizer_r.encode_plus, s, max_length=max_length, padding="max_length")
|
| 95 |
+
|
| 96 |
+
# Simple input
|
| 97 |
+
self.assertRaises(
|
| 98 |
+
ValueError,
|
| 99 |
+
tokenizer_r.batch_encode_plus,
|
| 100 |
+
s2,
|
| 101 |
+
max_length=max_length,
|
| 102 |
+
padding="max_length",
|
| 103 |
+
)
|
| 104 |
+
|
| 105 |
+
# Pair input
|
| 106 |
+
self.assertRaises(ValueError, tokenizer_r.encode, p, max_length=max_length, padding="max_length")
|
| 107 |
+
|
| 108 |
+
# Pair input
|
| 109 |
+
self.assertRaises(ValueError, tokenizer_r.encode_plus, p, max_length=max_length, padding="max_length")
|
| 110 |
+
|
| 111 |
+
# Pair input
|
| 112 |
+
self.assertRaises(
|
| 113 |
+
ValueError,
|
| 114 |
+
tokenizer_r.batch_encode_plus,
|
| 115 |
+
p2,
|
| 116 |
+
max_length=max_length,
|
| 117 |
+
padding="max_length",
|
| 118 |
+
)
|
| 119 |
+
|
| 120 |
+
def test_encodings_from_xnli_dataset(self):
|
| 121 |
+
"""
|
| 122 |
+
Tests the tokenizer downloaded from here:
|
| 123 |
+
- https://huggingface.co/bigscience/tokenizer/
|
| 124 |
+
"""
|
| 125 |
+
tokenizer = self.get_rust_tokenizer()
|
| 126 |
+
ds = load_dataset("facebook/xnli", "all_languages", split="test", streaming=True)
|
| 127 |
+
|
| 128 |
+
sample_data = next(iter(ds))["premise"] # pick up one data
|
| 129 |
+
input_text = list(sample_data.values())
|
| 130 |
+
|
| 131 |
+
output_tokens = list(map(tokenizer.encode, input_text))
|
| 132 |
+
predicted_text = [tokenizer.decode(x, clean_up_tokenization_spaces=False) for x in output_tokens]
|
| 133 |
+
self.assertListEqual(predicted_text, input_text)
|
| 134 |
+
|
| 135 |
+
@require_jinja
|
| 136 |
+
def test_tokenization_for_chat(self):
|
| 137 |
+
tokenizer = self.get_rust_tokenizer()
|
| 138 |
+
tokenizer.chat_template = "{% for message in messages %}" "{{ message.content }}{{ eos_token }}" "{% endfor %}"
|
| 139 |
+
test_chats = [
|
| 140 |
+
[{"role": "system", "content": "You are a helpful chatbot."}, {"role": "user", "content": "Hello!"}],
|
| 141 |
+
[
|
| 142 |
+
{"role": "system", "content": "You are a helpful chatbot."},
|
| 143 |
+
{"role": "user", "content": "Hello!"},
|
| 144 |
+
{"role": "assistant", "content": "Nice to meet you."},
|
| 145 |
+
],
|
| 146 |
+
[{"role": "assistant", "content": "Nice to meet you."}, {"role": "user", "content": "Hello!"}],
|
| 147 |
+
]
|
| 148 |
+
tokenized_chats = [tokenizer.apply_chat_template(test_chat) for test_chat in test_chats]
|
| 149 |
+
expected_tokens = [
|
| 150 |
+
[5448, 1306, 267, 66799, 44799, 37143, 17, 2, 59414, 4, 2],
|
| 151 |
+
[5448, 1306, 267, 66799, 44799, 37143, 17, 2, 59414, 4, 2, 229126, 427, 11890, 1152, 17, 2],
|
| 152 |
+
[229126, 427, 11890, 1152, 17, 2, 59414, 4, 2],
|
| 153 |
+
]
|
| 154 |
+
for tokenized_chat, expected_tokens in zip(tokenized_chats, expected_tokens):
|
| 155 |
+
self.assertListEqual(tokenized_chat, expected_tokens)
|
| 156 |
+
|
| 157 |
+
def test_add_prefix_space_fast(self):
|
| 158 |
+
tokenizer_w_prefix = self.get_rust_tokenizer(add_prefix_space=True)
|
| 159 |
+
tokenizer_wo_prefix = self.get_rust_tokenizer(add_prefix_space=False)
|
| 160 |
+
tokens_w_prefix = tokenizer_w_prefix.tokenize("Hey")
|
| 161 |
+
tokens_wo_prefix = tokenizer_wo_prefix.tokenize("Hey")
|
| 162 |
+
self.assertNotEqual(tokens_w_prefix, tokens_wo_prefix)
|
tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/bridgetower/__init__.py
ADDED
|
File without changes
|
tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/bridgetower/test_image_processing_bridgetower.py
ADDED
|
@@ -0,0 +1,158 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# coding=utf-8
|
| 2 |
+
# Copyright 2023 The Intel Labs Team Authors, The Microsoft Research Team Authors and HuggingFace Inc. team. All rights reserved.
|
| 3 |
+
#
|
| 4 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
| 5 |
+
# you may not use this file except in compliance with the License.
|
| 6 |
+
# You may obtain a copy of the License at
|
| 7 |
+
#
|
| 8 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
| 9 |
+
#
|
| 10 |
+
# Unless required by applicable law or agreed to in writing, software
|
| 11 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
| 12 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
| 13 |
+
# See the License for the specific language governing permissions and
|
| 14 |
+
# limitations under the License.
|
| 15 |
+
|
| 16 |
+
|
| 17 |
+
import unittest
|
| 18 |
+
from typing import Dict, List, Optional, Union
|
| 19 |
+
|
| 20 |
+
import numpy as np
|
| 21 |
+
|
| 22 |
+
from transformers.testing_utils import require_torch, require_vision
|
| 23 |
+
from transformers.utils import is_vision_available
|
| 24 |
+
|
| 25 |
+
from ...test_image_processing_common import ImageProcessingTestMixin, prepare_image_inputs
|
| 26 |
+
|
| 27 |
+
|
| 28 |
+
if is_vision_available():
|
| 29 |
+
from PIL import Image
|
| 30 |
+
|
| 31 |
+
from transformers import BridgeTowerImageProcessor
|
| 32 |
+
|
| 33 |
+
|
| 34 |
+
class BridgeTowerImageProcessingTester(unittest.TestCase):
|
| 35 |
+
def __init__(
|
| 36 |
+
self,
|
| 37 |
+
parent,
|
| 38 |
+
do_resize: bool = True,
|
| 39 |
+
size: Dict[str, int] = None,
|
| 40 |
+
size_divisor: int = 32,
|
| 41 |
+
do_rescale: bool = True,
|
| 42 |
+
rescale_factor: Union[int, float] = 1 / 255,
|
| 43 |
+
do_normalize: bool = True,
|
| 44 |
+
do_center_crop: bool = True,
|
| 45 |
+
image_mean: Optional[Union[float, List[float]]] = [0.48145466, 0.4578275, 0.40821073],
|
| 46 |
+
image_std: Optional[Union[float, List[float]]] = [0.26862954, 0.26130258, 0.27577711],
|
| 47 |
+
do_pad: bool = True,
|
| 48 |
+
batch_size=7,
|
| 49 |
+
min_resolution=30,
|
| 50 |
+
max_resolution=400,
|
| 51 |
+
num_channels=3,
|
| 52 |
+
):
|
| 53 |
+
super().__init__()
|
| 54 |
+
self.parent = parent
|
| 55 |
+
self.do_resize = do_resize
|
| 56 |
+
self.size = size if size is not None else {"shortest_edge": 288}
|
| 57 |
+
self.size_divisor = size_divisor
|
| 58 |
+
self.do_rescale = do_rescale
|
| 59 |
+
self.rescale_factor = rescale_factor
|
| 60 |
+
self.do_normalize = do_normalize
|
| 61 |
+
self.do_center_crop = do_center_crop
|
| 62 |
+
self.image_mean = image_mean
|
| 63 |
+
self.image_std = image_std
|
| 64 |
+
self.do_pad = do_pad
|
| 65 |
+
self.batch_size = batch_size
|
| 66 |
+
self.num_channels = num_channels
|
| 67 |
+
self.min_resolution = min_resolution
|
| 68 |
+
self.max_resolution = max_resolution
|
| 69 |
+
|
| 70 |
+
def prepare_image_processor_dict(self):
|
| 71 |
+
return {
|
| 72 |
+
"image_mean": self.image_mean,
|
| 73 |
+
"image_std": self.image_std,
|
| 74 |
+
"do_normalize": self.do_normalize,
|
| 75 |
+
"do_resize": self.do_resize,
|
| 76 |
+
"size": self.size,
|
| 77 |
+
"size_divisor": self.size_divisor,
|
| 78 |
+
}
|
| 79 |
+
|
| 80 |
+
def get_expected_values(self, image_inputs, batched=False):
|
| 81 |
+
"""
|
| 82 |
+
This function computes the expected height and width when providing images to BridgeTowerImageProcessor,
|
| 83 |
+
assuming do_resize is set to True with a scalar size and size_divisor.
|
| 84 |
+
"""
|
| 85 |
+
if not batched:
|
| 86 |
+
size = self.size["shortest_edge"]
|
| 87 |
+
image = image_inputs[0]
|
| 88 |
+
if isinstance(image, Image.Image):
|
| 89 |
+
w, h = image.size
|
| 90 |
+
elif isinstance(image, np.ndarray):
|
| 91 |
+
h, w = image.shape[0], image.shape[1]
|
| 92 |
+
else:
|
| 93 |
+
h, w = image.shape[1], image.shape[2]
|
| 94 |
+
scale = size / min(w, h)
|
| 95 |
+
if h < w:
|
| 96 |
+
newh, neww = size, scale * w
|
| 97 |
+
else:
|
| 98 |
+
newh, neww = scale * h, size
|
| 99 |
+
|
| 100 |
+
max_size = int((1333 / 800) * size)
|
| 101 |
+
if max(newh, neww) > max_size:
|
| 102 |
+
scale = max_size / max(newh, neww)
|
| 103 |
+
newh = newh * scale
|
| 104 |
+
neww = neww * scale
|
| 105 |
+
|
| 106 |
+
newh, neww = int(newh + 0.5), int(neww + 0.5)
|
| 107 |
+
expected_height, expected_width = (
|
| 108 |
+
newh // self.size_divisor * self.size_divisor,
|
| 109 |
+
neww // self.size_divisor * self.size_divisor,
|
| 110 |
+
)
|
| 111 |
+
|
| 112 |
+
else:
|
| 113 |
+
expected_values = []
|
| 114 |
+
for image in image_inputs:
|
| 115 |
+
expected_height, expected_width = self.get_expected_values([image])
|
| 116 |
+
expected_values.append((expected_height, expected_width))
|
| 117 |
+
expected_height = max(expected_values, key=lambda item: item[0])[0]
|
| 118 |
+
expected_width = max(expected_values, key=lambda item: item[1])[1]
|
| 119 |
+
|
| 120 |
+
return expected_height, expected_width
|
| 121 |
+
|
| 122 |
+
def expected_output_image_shape(self, images):
|
| 123 |
+
height, width = self.get_expected_values(images, batched=True)
|
| 124 |
+
return self.num_channels, height, width
|
| 125 |
+
|
| 126 |
+
def prepare_image_inputs(self, equal_resolution=False, numpify=False, torchify=False):
|
| 127 |
+
return prepare_image_inputs(
|
| 128 |
+
batch_size=self.batch_size,
|
| 129 |
+
num_channels=self.num_channels,
|
| 130 |
+
min_resolution=self.min_resolution,
|
| 131 |
+
max_resolution=self.max_resolution,
|
| 132 |
+
equal_resolution=equal_resolution,
|
| 133 |
+
numpify=numpify,
|
| 134 |
+
torchify=torchify,
|
| 135 |
+
)
|
| 136 |
+
|
| 137 |
+
|
| 138 |
+
@require_torch
|
| 139 |
+
@require_vision
|
| 140 |
+
class BridgeTowerImageProcessingTest(ImageProcessingTestMixin, unittest.TestCase):
|
| 141 |
+
image_processing_class = BridgeTowerImageProcessor if is_vision_available() else None
|
| 142 |
+
|
| 143 |
+
def setUp(self):
|
| 144 |
+
super().setUp()
|
| 145 |
+
self.image_processor_tester = BridgeTowerImageProcessingTester(self)
|
| 146 |
+
|
| 147 |
+
@property
|
| 148 |
+
def image_processor_dict(self):
|
| 149 |
+
return self.image_processor_tester.prepare_image_processor_dict()
|
| 150 |
+
|
| 151 |
+
def test_image_processor_properties(self):
|
| 152 |
+
image_processing = self.image_processing_class(**self.image_processor_dict)
|
| 153 |
+
self.assertTrue(hasattr(image_processing, "image_mean"))
|
| 154 |
+
self.assertTrue(hasattr(image_processing, "image_std"))
|
| 155 |
+
self.assertTrue(hasattr(image_processing, "do_normalize"))
|
| 156 |
+
self.assertTrue(hasattr(image_processing, "do_resize"))
|
| 157 |
+
self.assertTrue(hasattr(image_processing, "size"))
|
| 158 |
+
self.assertTrue(hasattr(image_processing, "size_divisor"))
|
tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/bridgetower/test_modeling_bridgetower.py
ADDED
|
@@ -0,0 +1,692 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# coding=utf-8
|
| 2 |
+
# Copyright 2023 The Intel Labs Team Authors, The Microsoft Research Team Authors and HuggingFace Inc. team. All rights reserved.
|
| 3 |
+
#
|
| 4 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
| 5 |
+
# you may not use this file except in compliance with the License.
|
| 6 |
+
# You may obtain a copy of the License at
|
| 7 |
+
#
|
| 8 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
| 9 |
+
#
|
| 10 |
+
# Unless required by applicable law or agreed to in writing, software
|
| 11 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
| 12 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
| 13 |
+
# See the License for the specific language governing permissions and
|
| 14 |
+
# limitations under the License.
|
| 15 |
+
"""Testing suite for the PyTorch BridgeTower model."""
|
| 16 |
+
|
| 17 |
+
import tempfile
|
| 18 |
+
import unittest
|
| 19 |
+
|
| 20 |
+
import numpy as np
|
| 21 |
+
|
| 22 |
+
from transformers import (
|
| 23 |
+
BridgeTowerConfig,
|
| 24 |
+
BridgeTowerTextConfig,
|
| 25 |
+
BridgeTowerVisionConfig,
|
| 26 |
+
is_torch_available,
|
| 27 |
+
is_vision_available,
|
| 28 |
+
)
|
| 29 |
+
from transformers.testing_utils import require_torch, require_vision, slow, torch_device
|
| 30 |
+
from transformers.utils import cached_property
|
| 31 |
+
|
| 32 |
+
from ...test_configuration_common import ConfigTester
|
| 33 |
+
from ...test_modeling_common import (
|
| 34 |
+
ModelTesterMixin,
|
| 35 |
+
_config_zero_init,
|
| 36 |
+
floats_tensor,
|
| 37 |
+
ids_tensor,
|
| 38 |
+
random_attention_mask,
|
| 39 |
+
)
|
| 40 |
+
from ...test_pipeline_mixin import PipelineTesterMixin
|
| 41 |
+
|
| 42 |
+
|
| 43 |
+
if is_torch_available():
|
| 44 |
+
import torch
|
| 45 |
+
|
| 46 |
+
from transformers import (
|
| 47 |
+
BridgeTowerForContrastiveLearning,
|
| 48 |
+
BridgeTowerForImageAndTextRetrieval,
|
| 49 |
+
BridgeTowerForMaskedLM,
|
| 50 |
+
BridgeTowerModel,
|
| 51 |
+
)
|
| 52 |
+
|
| 53 |
+
if is_vision_available():
|
| 54 |
+
from PIL import Image
|
| 55 |
+
|
| 56 |
+
from transformers import BridgeTowerProcessor
|
| 57 |
+
|
| 58 |
+
|
| 59 |
+
class BridgeTowerTextModelTester:
|
| 60 |
+
def __init__(
|
| 61 |
+
self,
|
| 62 |
+
parent,
|
| 63 |
+
hidden_act="gelu",
|
| 64 |
+
hidden_size=64,
|
| 65 |
+
initializer_factor=1,
|
| 66 |
+
layer_norm_eps=1e-05,
|
| 67 |
+
num_attention_heads=4,
|
| 68 |
+
num_hidden_layers=2,
|
| 69 |
+
intermediate_size=128,
|
| 70 |
+
tie_word_embeddings=False,
|
| 71 |
+
output_hidden_states=False,
|
| 72 |
+
):
|
| 73 |
+
self.parent = parent
|
| 74 |
+
self.hidden_act = hidden_act
|
| 75 |
+
self.hidden_size = hidden_size
|
| 76 |
+
self.initializer_factor = initializer_factor
|
| 77 |
+
self.layer_norm_eps = layer_norm_eps
|
| 78 |
+
self.num_attention_heads = num_attention_heads
|
| 79 |
+
self.num_hidden_layers = num_hidden_layers
|
| 80 |
+
self.intermediate_size = intermediate_size
|
| 81 |
+
self.tie_word_embeddings = tie_word_embeddings
|
| 82 |
+
self.vocab_size = 99
|
| 83 |
+
self.seq_length = 4
|
| 84 |
+
self.batch_size = 1
|
| 85 |
+
self.is_training = False
|
| 86 |
+
self.output_hidden_states = output_hidden_states
|
| 87 |
+
|
| 88 |
+
def prepare_config_and_inputs(self):
|
| 89 |
+
input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
|
| 90 |
+
attention_mask = random_attention_mask([self.batch_size, self.seq_length])
|
| 91 |
+
|
| 92 |
+
config = self.get_config()
|
| 93 |
+
|
| 94 |
+
return config, input_ids, attention_mask
|
| 95 |
+
|
| 96 |
+
def get_config(self):
|
| 97 |
+
return BridgeTowerTextConfig(
|
| 98 |
+
hidden_act=self.hidden_act,
|
| 99 |
+
hidden_size=self.hidden_size,
|
| 100 |
+
initializer_factor=self.initializer_factor,
|
| 101 |
+
layer_norm_eps=self.layer_norm_eps,
|
| 102 |
+
num_attention_heads=self.num_attention_heads,
|
| 103 |
+
num_hidden_layers=self.num_hidden_layers,
|
| 104 |
+
intermediate_size=self.intermediate_size,
|
| 105 |
+
tie_word_embeddings=self.tie_word_embeddings,
|
| 106 |
+
output_hidden_states=self.output_hidden_states,
|
| 107 |
+
vocab_size=self.vocab_size,
|
| 108 |
+
)
|
| 109 |
+
|
| 110 |
+
|
| 111 |
+
class BridgeTowerImageModelTester:
|
| 112 |
+
def __init__(
|
| 113 |
+
self,
|
| 114 |
+
parent,
|
| 115 |
+
hidden_size=64,
|
| 116 |
+
initializer_factor=1,
|
| 117 |
+
layer_norm_eps=1e-05,
|
| 118 |
+
num_hidden_layers=2,
|
| 119 |
+
init_layernorm_from_vision_encoder=False,
|
| 120 |
+
output_hidden_states=False,
|
| 121 |
+
image_size=64,
|
| 122 |
+
):
|
| 123 |
+
self.parent = parent
|
| 124 |
+
self.hidden_size = hidden_size
|
| 125 |
+
self.initializer_factor = initializer_factor
|
| 126 |
+
self.layer_norm_eps = layer_norm_eps
|
| 127 |
+
self.num_hidden_layers = num_hidden_layers
|
| 128 |
+
self.init_layernorm_from_vision_encoder = init_layernorm_from_vision_encoder
|
| 129 |
+
self.num_channels = 3
|
| 130 |
+
self.num_image_features = 17
|
| 131 |
+
self.batch_size = 1
|
| 132 |
+
self.image_size = image_size
|
| 133 |
+
self.is_training = False
|
| 134 |
+
self.output_hidden_states = output_hidden_states
|
| 135 |
+
|
| 136 |
+
def prepare_config_and_inputs(self):
|
| 137 |
+
pixel_values = floats_tensor([self.batch_size, self.num_channels, self.image_size, self.image_size])
|
| 138 |
+
pixel_mask = random_attention_mask([self.batch_size, self.image_size, self.image_size])
|
| 139 |
+
config = self.get_config()
|
| 140 |
+
|
| 141 |
+
return config, pixel_values, pixel_mask
|
| 142 |
+
|
| 143 |
+
def get_config(self):
|
| 144 |
+
return BridgeTowerVisionConfig(
|
| 145 |
+
hidden_size=self.hidden_size,
|
| 146 |
+
initializer_factor=self.initializer_factor,
|
| 147 |
+
layer_norm_eps=self.layer_norm_eps,
|
| 148 |
+
num_hidden_layers=self.num_hidden_layers,
|
| 149 |
+
init_layernorm_from_vision_encoder=self.init_layernorm_from_vision_encoder,
|
| 150 |
+
num_channels=self.num_channels,
|
| 151 |
+
num_image_features=self.num_image_features,
|
| 152 |
+
batch_size=self.batch_size,
|
| 153 |
+
image_size=self.image_size,
|
| 154 |
+
is_training=self.is_training,
|
| 155 |
+
output_hidden_states=self.output_hidden_states,
|
| 156 |
+
)
|
| 157 |
+
|
| 158 |
+
|
| 159 |
+
class BridgeTowerModelTester:
|
| 160 |
+
def __init__(
|
| 161 |
+
self,
|
| 162 |
+
parent,
|
| 163 |
+
text_kwargs=None,
|
| 164 |
+
vision_kwargs=None,
|
| 165 |
+
share_cross_modal_transformer_layers=True,
|
| 166 |
+
share_link_tower_layers=False,
|
| 167 |
+
link_tower_type="add",
|
| 168 |
+
init_layernorm_from_vision_encoder=False,
|
| 169 |
+
contrastive_hidden_size=512,
|
| 170 |
+
logit_scale_init_value=2.6592,
|
| 171 |
+
hidden_size=64,
|
| 172 |
+
num_hidden_layers=2,
|
| 173 |
+
num_attention_heads=4,
|
| 174 |
+
intermediate_size=128,
|
| 175 |
+
):
|
| 176 |
+
if text_kwargs is None:
|
| 177 |
+
text_kwargs = {}
|
| 178 |
+
if vision_kwargs is None:
|
| 179 |
+
vision_kwargs = {}
|
| 180 |
+
|
| 181 |
+
self.parent = parent
|
| 182 |
+
self.text_model_tester = BridgeTowerTextModelTester(parent, **text_kwargs)
|
| 183 |
+
self.vision_model_tester = BridgeTowerImageModelTester(parent, **vision_kwargs)
|
| 184 |
+
|
| 185 |
+
self.share_cross_modal_transformer_layers = share_cross_modal_transformer_layers
|
| 186 |
+
self.share_link_tower_layers = share_link_tower_layers
|
| 187 |
+
self.link_tower_type = link_tower_type
|
| 188 |
+
self.init_layernorm_from_vision_encoder = init_layernorm_from_vision_encoder
|
| 189 |
+
self.contrastive_hidden_size = contrastive_hidden_size
|
| 190 |
+
self.logit_scale_init_value = logit_scale_init_value
|
| 191 |
+
|
| 192 |
+
self.batch_size = 1
|
| 193 |
+
self.expected_num_hidden_layers = 8
|
| 194 |
+
self.is_training = False
|
| 195 |
+
|
| 196 |
+
self.hidden_size = hidden_size
|
| 197 |
+
self.num_hidden_layers = num_hidden_layers
|
| 198 |
+
self.num_attention_heads = num_attention_heads
|
| 199 |
+
self.intermediate_size = intermediate_size
|
| 200 |
+
|
| 201 |
+
def prepare_config_and_inputs(self):
|
| 202 |
+
text_config, input_ids, attention_mask = self.text_model_tester.prepare_config_and_inputs()
|
| 203 |
+
vision_config, pixel_values, pixel_mask = self.vision_model_tester.prepare_config_and_inputs()
|
| 204 |
+
|
| 205 |
+
config = self.get_config()
|
| 206 |
+
|
| 207 |
+
return (config, input_ids, attention_mask, pixel_values, pixel_mask)
|
| 208 |
+
|
| 209 |
+
def get_config(self):
|
| 210 |
+
return BridgeTowerConfig.from_text_vision_configs(
|
| 211 |
+
text_config=self.text_model_tester.get_config(),
|
| 212 |
+
vision_config=self.vision_model_tester.get_config(),
|
| 213 |
+
share_cross_modal_transformer_layers=self.share_cross_modal_transformer_layers,
|
| 214 |
+
share_link_tower_layers=self.share_link_tower_layers,
|
| 215 |
+
link_tower_type=self.link_tower_type,
|
| 216 |
+
init_layernorm_from_vision_encoder=self.init_layernorm_from_vision_encoder,
|
| 217 |
+
contrastive_hidden_size=self.contrastive_hidden_size,
|
| 218 |
+
logit_scale_init_value=self.logit_scale_init_value,
|
| 219 |
+
hidden_size=self.hidden_size,
|
| 220 |
+
num_hidden_layers=self.num_hidden_layers,
|
| 221 |
+
num_attention_heads=self.num_attention_heads,
|
| 222 |
+
intermediate_size=self.intermediate_size,
|
| 223 |
+
)
|
| 224 |
+
|
| 225 |
+
def create_and_check_model(
|
| 226 |
+
self,
|
| 227 |
+
config,
|
| 228 |
+
input_ids,
|
| 229 |
+
attention_mask,
|
| 230 |
+
pixel_values,
|
| 231 |
+
pixel_mask,
|
| 232 |
+
):
|
| 233 |
+
model = BridgeTowerModel(config=config)
|
| 234 |
+
model.to(torch_device)
|
| 235 |
+
model.eval()
|
| 236 |
+
result = model(input_ids, attention_mask=attention_mask, pixel_values=pixel_values, pixel_mask=pixel_mask)
|
| 237 |
+
result = model(input_ids, attention_mask=attention_mask, pixel_values=pixel_values)
|
| 238 |
+
self.parent.assertEqual(
|
| 239 |
+
result["text_features"].shape,
|
| 240 |
+
(self.batch_size, self.text_model_tester.seq_length, self.text_model_tester.hidden_size),
|
| 241 |
+
)
|
| 242 |
+
self.parent.assertEqual(
|
| 243 |
+
result["image_features"].shape,
|
| 244 |
+
(self.batch_size, self.vision_model_tester.num_image_features, self.vision_model_tester.hidden_size),
|
| 245 |
+
)
|
| 246 |
+
self.parent.assertEqual(
|
| 247 |
+
result["pooler_output"].shape,
|
| 248 |
+
(self.batch_size, self.text_model_tester.hidden_size + self.vision_model_tester.hidden_size),
|
| 249 |
+
)
|
| 250 |
+
|
| 251 |
+
def create_and_check_for_image_and_text_retrieval(
|
| 252 |
+
self,
|
| 253 |
+
config,
|
| 254 |
+
input_ids,
|
| 255 |
+
attention_mask,
|
| 256 |
+
pixel_values,
|
| 257 |
+
pixel_mask,
|
| 258 |
+
):
|
| 259 |
+
bridgetower_itm_output_last_dimension = 2
|
| 260 |
+
|
| 261 |
+
model = BridgeTowerForImageAndTextRetrieval(config)
|
| 262 |
+
model.to(torch_device)
|
| 263 |
+
model.eval()
|
| 264 |
+
result = model(input_ids, attention_mask=attention_mask, pixel_values=pixel_values, pixel_mask=pixel_mask)
|
| 265 |
+
result = model(input_ids, attention_mask=attention_mask, pixel_values=pixel_values)
|
| 266 |
+
|
| 267 |
+
self.parent.assertEqual(result.logits.shape, (self.batch_size, bridgetower_itm_output_last_dimension))
|
| 268 |
+
|
| 269 |
+
def create_and_check_for_masked_language_modeling(
|
| 270 |
+
self,
|
| 271 |
+
config,
|
| 272 |
+
input_ids,
|
| 273 |
+
attention_mask,
|
| 274 |
+
pixel_values,
|
| 275 |
+
pixel_mask,
|
| 276 |
+
):
|
| 277 |
+
model = BridgeTowerForMaskedLM(config)
|
| 278 |
+
model.to(torch_device)
|
| 279 |
+
model.eval()
|
| 280 |
+
result = model(input_ids, attention_mask=attention_mask, pixel_values=pixel_values, pixel_mask=pixel_mask)
|
| 281 |
+
result = model(input_ids, attention_mask=attention_mask, pixel_values=pixel_values)
|
| 282 |
+
|
| 283 |
+
self.parent.assertEqual(
|
| 284 |
+
result.logits.shape,
|
| 285 |
+
(self.batch_size, self.text_model_tester.seq_length, self.text_model_tester.vocab_size),
|
| 286 |
+
)
|
| 287 |
+
|
| 288 |
+
def prepare_config_and_inputs_for_common(self):
|
| 289 |
+
config_and_inputs = self.prepare_config_and_inputs()
|
| 290 |
+
(config, input_ids, attention_mask, pixel_values, pixel_mask) = config_and_inputs
|
| 291 |
+
inputs_dict = {
|
| 292 |
+
"input_ids": input_ids,
|
| 293 |
+
"attention_mask": attention_mask,
|
| 294 |
+
"pixel_values": pixel_values,
|
| 295 |
+
"pixel_mask": pixel_mask,
|
| 296 |
+
}
|
| 297 |
+
return config, inputs_dict
|
| 298 |
+
|
| 299 |
+
|
| 300 |
+
@require_torch
|
| 301 |
+
class BridgeTowerModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase):
|
| 302 |
+
all_model_classes = (
|
| 303 |
+
(
|
| 304 |
+
BridgeTowerModel,
|
| 305 |
+
BridgeTowerForImageAndTextRetrieval,
|
| 306 |
+
BridgeTowerForMaskedLM,
|
| 307 |
+
BridgeTowerForContrastiveLearning,
|
| 308 |
+
)
|
| 309 |
+
if is_torch_available()
|
| 310 |
+
else ()
|
| 311 |
+
)
|
| 312 |
+
pipeline_model_mapping = {"feature-extraction": BridgeTowerModel} if is_torch_available() else {}
|
| 313 |
+
|
| 314 |
+
is_training = False
|
| 315 |
+
test_headmasking = False
|
| 316 |
+
test_pruning = False
|
| 317 |
+
test_torchscript = False
|
| 318 |
+
test_resize_embeddings = False
|
| 319 |
+
has_attentions = False
|
| 320 |
+
|
| 321 |
+
@unittest.skip(reason="Does not work on the tiny model as we keep hitting edge cases.")
|
| 322 |
+
def test_cpu_offload(self):
|
| 323 |
+
pass
|
| 324 |
+
|
| 325 |
+
@unittest.skip(reason="Does not work on the tiny model as we keep hitting edge cases.")
|
| 326 |
+
def test_disk_offload(self):
|
| 327 |
+
pass
|
| 328 |
+
|
| 329 |
+
@unittest.skip(reason="Does not work on the tiny model as we keep hitting edge cases.")
|
| 330 |
+
def test_model_parallelism(self):
|
| 331 |
+
pass
|
| 332 |
+
|
| 333 |
+
# function to extract meaningful tensor from output per different model_class
|
| 334 |
+
def extract_output(self, outputs, model_class):
|
| 335 |
+
return outputs["pooler_output"] if model_class == "BridgeTowerModel" else outputs["logits"]
|
| 336 |
+
|
| 337 |
+
def setUp(self):
|
| 338 |
+
self.model_tester = BridgeTowerModelTester(self)
|
| 339 |
+
self.config_tester = ConfigTester(self, config_class=BridgeTowerConfig, hidden_size=37, vocab_size=99)
|
| 340 |
+
|
| 341 |
+
def test_config(self):
|
| 342 |
+
self.config_tester.run_common_tests()
|
| 343 |
+
|
| 344 |
+
def test_model(self):
|
| 345 |
+
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
| 346 |
+
self.model_tester.create_and_check_model(*config_and_inputs)
|
| 347 |
+
|
| 348 |
+
def test_for_image_and_text_retrieval(self):
|
| 349 |
+
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
| 350 |
+
self.model_tester.create_and_check_for_image_and_text_retrieval(*config_and_inputs)
|
| 351 |
+
|
| 352 |
+
def test_for_masked_language_modeling(self):
|
| 353 |
+
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
| 354 |
+
self.model_tester.create_and_check_for_masked_language_modeling(*config_and_inputs)
|
| 355 |
+
|
| 356 |
+
@slow
|
| 357 |
+
def test_model_from_pretrained(self):
|
| 358 |
+
model_name = "BridgeTower/bridgetower-base"
|
| 359 |
+
model = BridgeTowerModel.from_pretrained(model_name)
|
| 360 |
+
self.assertIsNotNone(model)
|
| 361 |
+
|
| 362 |
+
@slow
|
| 363 |
+
def test_save_load_fast_init_from_base(self):
|
| 364 |
+
# Override as it is a slow test on this model
|
| 365 |
+
super().test_save_load_fast_init_from_base()
|
| 366 |
+
|
| 367 |
+
# Override as extracting meaningful tensor from output is different for BridgeTower
|
| 368 |
+
def test_save_load(self):
|
| 369 |
+
config, input_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
| 370 |
+
for model_class in self.all_model_classes:
|
| 371 |
+
model = model_class(config)
|
| 372 |
+
model.to(torch_device)
|
| 373 |
+
model.eval()
|
| 374 |
+
with torch.no_grad():
|
| 375 |
+
outputs = model(**input_dict)
|
| 376 |
+
|
| 377 |
+
out_2 = self.extract_output(outputs, model_class.__name__)
|
| 378 |
+
out_2 = out_2.cpu().numpy()
|
| 379 |
+
out_2[np.isnan(out_2)] = 0
|
| 380 |
+
|
| 381 |
+
with tempfile.TemporaryDirectory() as tmpdirname:
|
| 382 |
+
model.save_pretrained(tmpdirname)
|
| 383 |
+
model = model_class.from_pretrained(tmpdirname)
|
| 384 |
+
model.to(torch_device)
|
| 385 |
+
with torch.no_grad():
|
| 386 |
+
after_outputs = model(**input_dict)
|
| 387 |
+
|
| 388 |
+
# Make sure we don't have nans
|
| 389 |
+
out_1 = self.extract_output(after_outputs, model_class.__name__)
|
| 390 |
+
out_1 = out_1.cpu().numpy()
|
| 391 |
+
out_1[np.isnan(out_1)] = 0
|
| 392 |
+
max_diff = np.amax(np.abs(out_1 - out_2))
|
| 393 |
+
self.assertLessEqual(max_diff, 1e-5)
|
| 394 |
+
|
| 395 |
+
# Override this as `hidden states output` is different for BridgeTower
|
| 396 |
+
def test_hidden_states_output(self):
|
| 397 |
+
def check_hidden_states_output(inputs_dict, config, model_class):
|
| 398 |
+
model = model_class(config)
|
| 399 |
+
model.to(torch_device)
|
| 400 |
+
model.eval()
|
| 401 |
+
|
| 402 |
+
with torch.no_grad():
|
| 403 |
+
outputs = model(**self._prepare_for_class(inputs_dict, model_class))
|
| 404 |
+
|
| 405 |
+
hidden_states_text, hidden_states_vision, hidden_states_cross = (
|
| 406 |
+
outputs.encoder_hidden_states if config.is_encoder_decoder else outputs.hidden_states
|
| 407 |
+
)
|
| 408 |
+
|
| 409 |
+
expected_num_layers = self.model_tester.expected_num_hidden_layers
|
| 410 |
+
self.assertEqual(
|
| 411 |
+
sum((len(hidden_states_text), len(hidden_states_vision), len(hidden_states_cross))),
|
| 412 |
+
expected_num_layers,
|
| 413 |
+
)
|
| 414 |
+
|
| 415 |
+
seq_length = self.model_tester.text_model_tester.seq_length
|
| 416 |
+
num_image_features = self.model_tester.vision_model_tester.num_image_features
|
| 417 |
+
|
| 418 |
+
self.assertListEqual(
|
| 419 |
+
list(hidden_states_text[0].shape[-2:]),
|
| 420 |
+
[seq_length, self.model_tester.text_model_tester.hidden_size],
|
| 421 |
+
)
|
| 422 |
+
self.assertListEqual(
|
| 423 |
+
list(hidden_states_vision[0].shape),
|
| 424 |
+
[num_image_features, 1, self.model_tester.vision_model_tester.hidden_size],
|
| 425 |
+
)
|
| 426 |
+
self.assertListEqual(
|
| 427 |
+
list(hidden_states_cross[0][0].shape[-2:]),
|
| 428 |
+
[seq_length, self.model_tester.text_model_tester.hidden_size],
|
| 429 |
+
)
|
| 430 |
+
self.assertListEqual(
|
| 431 |
+
list(hidden_states_cross[0][1].shape[-2:]),
|
| 432 |
+
[num_image_features, self.model_tester.vision_model_tester.hidden_size],
|
| 433 |
+
)
|
| 434 |
+
|
| 435 |
+
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
| 436 |
+
|
| 437 |
+
for model_class in self.all_model_classes:
|
| 438 |
+
inputs_dict["output_hidden_states"] = True
|
| 439 |
+
check_hidden_states_output(inputs_dict, config, model_class)
|
| 440 |
+
|
| 441 |
+
# check that output_hidden_states also work using config
|
| 442 |
+
del inputs_dict["output_hidden_states"]
|
| 443 |
+
config.output_hidden_states = True
|
| 444 |
+
check_hidden_states_output(inputs_dict, config, model_class)
|
| 445 |
+
|
| 446 |
+
# Override as `hidden states output` is different for BridgeTower
|
| 447 |
+
def test_retain_grad_hidden_states_attentions(self):
|
| 448 |
+
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
| 449 |
+
config.output_hidden_states = True
|
| 450 |
+
config.output_attentions = self.has_attentions
|
| 451 |
+
|
| 452 |
+
# no need to test all models as different heads yield the same functionality
|
| 453 |
+
model_class = self.all_model_classes[0]
|
| 454 |
+
model = model_class(config)
|
| 455 |
+
model.to(torch_device)
|
| 456 |
+
|
| 457 |
+
inputs = self._prepare_for_class(inputs_dict, model_class)
|
| 458 |
+
|
| 459 |
+
outputs = model(**inputs)
|
| 460 |
+
|
| 461 |
+
output = outputs[0]
|
| 462 |
+
|
| 463 |
+
# Encoder-/Decoder-only models
|
| 464 |
+
hidden_states = outputs.hidden_states[0][0]
|
| 465 |
+
hidden_states.retain_grad()
|
| 466 |
+
|
| 467 |
+
if self.has_attentions:
|
| 468 |
+
attentions = outputs.attentions[0][0]
|
| 469 |
+
attentions.retain_grad()
|
| 470 |
+
|
| 471 |
+
output.flatten()[0].backward(retain_graph=True)
|
| 472 |
+
|
| 473 |
+
self.assertIsNotNone(hidden_states.grad)
|
| 474 |
+
|
| 475 |
+
if self.has_attentions:
|
| 476 |
+
self.assertIsNotNone(attentions.grad)
|
| 477 |
+
|
| 478 |
+
# override as the `logit_scale` parameter initilization is different for BRIDGE TOWER
|
| 479 |
+
def test_initialization(self):
|
| 480 |
+
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
| 481 |
+
|
| 482 |
+
configs_no_init = _config_zero_init(config)
|
| 483 |
+
for model_class in self.all_model_classes:
|
| 484 |
+
model = model_class(config=configs_no_init)
|
| 485 |
+
for name, param in model.named_parameters():
|
| 486 |
+
if param.requires_grad:
|
| 487 |
+
if name == "logit_scale":
|
| 488 |
+
self.assertAlmostEqual(
|
| 489 |
+
param.data.item(),
|
| 490 |
+
config.logit_scale_init_value,
|
| 491 |
+
delta=1e-3,
|
| 492 |
+
msg=f"Parameter {name} of model {model_class} seems not properly initialized",
|
| 493 |
+
)
|
| 494 |
+
else:
|
| 495 |
+
self.assertIn(
|
| 496 |
+
((param.data.mean() * 1e9).round() / 1e9).item(),
|
| 497 |
+
[0.0, 1.0],
|
| 498 |
+
msg=f"Parameter {name} of model {model_class} seems not properly initialized",
|
| 499 |
+
)
|
| 500 |
+
|
| 501 |
+
@unittest.skip(reason="""Bridge Tower does not have input/output embeddings. So this test is not applicable.""")
|
| 502 |
+
def test_model_get_set_embeddings(self):
|
| 503 |
+
pass
|
| 504 |
+
|
| 505 |
+
@unittest.skip(reason="""Bridge Tower does not have input/output embeddings. Thus this test is not applicable.""")
|
| 506 |
+
def test_inputs_embeds(self):
|
| 507 |
+
pass
|
| 508 |
+
|
| 509 |
+
@unittest.skip(reason="Bridge Tower does not use inputs_embeds")
|
| 510 |
+
def test_inputs_embeds_matches_input_ids(self):
|
| 511 |
+
pass
|
| 512 |
+
|
| 513 |
+
|
| 514 |
+
# We will verify our results on an image of cute cats
|
| 515 |
+
def prepare_img():
|
| 516 |
+
image = Image.open("./tests/fixtures/tests_samples/COCO/000000039769.png")
|
| 517 |
+
return image
|
| 518 |
+
|
| 519 |
+
|
| 520 |
+
@require_torch
|
| 521 |
+
@require_vision
|
| 522 |
+
class BridgeTowerModelIntegrationTest(unittest.TestCase):
|
| 523 |
+
@cached_property
|
| 524 |
+
def default_processor(self):
|
| 525 |
+
return (
|
| 526 |
+
BridgeTowerProcessor.from_pretrained("BridgeTower/bridgetower-base-itm-mlm")
|
| 527 |
+
if is_vision_available()
|
| 528 |
+
else None
|
| 529 |
+
)
|
| 530 |
+
|
| 531 |
+
@slow
|
| 532 |
+
def test_image_and_text_retrieval(self):
|
| 533 |
+
model = BridgeTowerForImageAndTextRetrieval.from_pretrained("BridgeTower/bridgetower-base-itm-mlm").to(
|
| 534 |
+
torch_device
|
| 535 |
+
)
|
| 536 |
+
model.eval()
|
| 537 |
+
processor = self.default_processor
|
| 538 |
+
image = prepare_img()
|
| 539 |
+
text = "a bunch of cats laying on a tower."
|
| 540 |
+
inputs = processor(image, text, return_tensors="pt").to(torch_device)
|
| 541 |
+
|
| 542 |
+
# forward pass
|
| 543 |
+
with torch.no_grad():
|
| 544 |
+
outputs = model(**inputs)
|
| 545 |
+
|
| 546 |
+
# verify the logits
|
| 547 |
+
expected_shape = torch.Size([1, 2])
|
| 548 |
+
self.assertEqual(outputs.logits.shape, expected_shape)
|
| 549 |
+
self.assertTrue(outputs.logits[0, 1].item() > outputs.logits[0, 0].item())
|
| 550 |
+
|
| 551 |
+
# verify loss
|
| 552 |
+
inputs["labels"] = torch.ones(1, dtype=torch.long, device=torch_device)
|
| 553 |
+
inputs = inputs.to(torch_device)
|
| 554 |
+
with torch.no_grad():
|
| 555 |
+
outputs = model(**inputs)
|
| 556 |
+
self.assertAlmostEqual(outputs.loss.item(), 0.5108, places=4)
|
| 557 |
+
|
| 558 |
+
@slow
|
| 559 |
+
def test_masked_language_modeling(self):
|
| 560 |
+
model = BridgeTowerForMaskedLM.from_pretrained("BridgeTower/bridgetower-base-itm-mlm").to(torch_device)
|
| 561 |
+
model.eval()
|
| 562 |
+
processor = self.default_processor
|
| 563 |
+
image = prepare_img()
|
| 564 |
+
text = "a bunch of <mask> laying on a tower."
|
| 565 |
+
inputs = processor(image, text, return_tensors="pt").to(torch_device)
|
| 566 |
+
|
| 567 |
+
# forward pass
|
| 568 |
+
with torch.no_grad():
|
| 569 |
+
outputs = model(**inputs)
|
| 570 |
+
|
| 571 |
+
# verify the logits
|
| 572 |
+
expected_shape = torch.Size([1, 11, 50265])
|
| 573 |
+
self.assertEqual(outputs.logits.shape, expected_shape)
|
| 574 |
+
|
| 575 |
+
# verify predicted word
|
| 576 |
+
predicted_id = outputs.logits.argmax(dim=-1).squeeze(0).tolist()[4]
|
| 577 |
+
self.assertTrue(processor.decode([predicted_id]) == " cats")
|
| 578 |
+
|
| 579 |
+
# verify loss
|
| 580 |
+
inputs["labels"] = inputs["input_ids"].clone()
|
| 581 |
+
inputs = inputs.to(torch_device)
|
| 582 |
+
with torch.no_grad():
|
| 583 |
+
outputs = model(**inputs)
|
| 584 |
+
self.assertAlmostEqual(outputs.loss.item(), 5.7373, places=4)
|
| 585 |
+
|
| 586 |
+
@slow
|
| 587 |
+
def test_constrastive_learning(self):
|
| 588 |
+
model = BridgeTowerForContrastiveLearning.from_pretrained("BridgeTower/bridgetower-large-itm-mlm-itc").to(
|
| 589 |
+
torch_device
|
| 590 |
+
)
|
| 591 |
+
model.eval()
|
| 592 |
+
processor = BridgeTowerProcessor.from_pretrained("BridgeTower/bridgetower-large-itm-mlm-itc")
|
| 593 |
+
image = prepare_img()
|
| 594 |
+
text = "a bunch of cats laying on a tower."
|
| 595 |
+
inputs = processor(image, text, padding=True, return_tensors="pt").to(torch_device)
|
| 596 |
+
with torch.no_grad():
|
| 597 |
+
outputs = model(**inputs, output_hidden_states=True, return_loss=True)
|
| 598 |
+
|
| 599 |
+
# verify the logits
|
| 600 |
+
expected_shape = torch.Size([1, 3, 512])
|
| 601 |
+
self.assertEqual(outputs.logits.shape, expected_shape)
|
| 602 |
+
|
| 603 |
+
|
| 604 |
+
@slow
|
| 605 |
+
@require_torch
|
| 606 |
+
class BridgeTowerModelTrainingTest(unittest.TestCase):
|
| 607 |
+
all_training_supported_model_classes = (
|
| 608 |
+
(BridgeTowerForImageAndTextRetrieval, BridgeTowerForMaskedLM, BridgeTowerForContrastiveLearning)
|
| 609 |
+
if is_torch_available()
|
| 610 |
+
else ()
|
| 611 |
+
)
|
| 612 |
+
|
| 613 |
+
def setUp(self):
|
| 614 |
+
self.model_tester = BridgeTowerModelTester(self)
|
| 615 |
+
self.config_tester = ConfigTester(self, config_class=BridgeTowerConfig, hidden_size=37, vocab_size=99)
|
| 616 |
+
|
| 617 |
+
def _prepare_inputs_for_training(self, model_class):
|
| 618 |
+
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
| 619 |
+
if model_class == BridgeTowerForMaskedLM:
|
| 620 |
+
inputs_dict["labels"] = inputs_dict["input_ids"]
|
| 621 |
+
elif model_class == BridgeTowerForImageAndTextRetrieval:
|
| 622 |
+
inputs_dict["labels"] = ids_tensor([1], 2)
|
| 623 |
+
elif model_class == BridgeTowerForContrastiveLearning:
|
| 624 |
+
inputs_dict["return_loss"] = True
|
| 625 |
+
return config, inputs_dict
|
| 626 |
+
|
| 627 |
+
def _get_non_used_layer_names(self, model_class):
|
| 628 |
+
non_used_layer_names = ["text_model.pooler"]
|
| 629 |
+
if model_class == BridgeTowerForMaskedLM:
|
| 630 |
+
non_used_layer_names = non_used_layer_names + [
|
| 631 |
+
# This number `1` actually depends on the number of layers in `cross_modal_image_layers` (by minus 1)
|
| 632 |
+
"cross_modal_image_layers.1",
|
| 633 |
+
"cross_modal_image_pooler",
|
| 634 |
+
"cross_modal_text_pooler",
|
| 635 |
+
]
|
| 636 |
+
return non_used_layer_names
|
| 637 |
+
|
| 638 |
+
def _is_layer_used(self, model_class, layer_name):
|
| 639 |
+
non_used_layer_names = self._get_non_used_layer_names(model_class)
|
| 640 |
+
for non_used_layer_name in non_used_layer_names:
|
| 641 |
+
if non_used_layer_name in layer_name:
|
| 642 |
+
return False
|
| 643 |
+
return True
|
| 644 |
+
|
| 645 |
+
def test_training(self):
|
| 646 |
+
for model_class in self.all_training_supported_model_classes:
|
| 647 |
+
config, inputs_dict = self._prepare_inputs_for_training(model_class)
|
| 648 |
+
model = model_class(config)
|
| 649 |
+
model.to(torch_device)
|
| 650 |
+
model.train()
|
| 651 |
+
|
| 652 |
+
loss = model(**inputs_dict).loss
|
| 653 |
+
loss.backward()
|
| 654 |
+
|
| 655 |
+
# verify the gradients of used layers' weight are not None
|
| 656 |
+
for name, param in model.named_parameters():
|
| 657 |
+
if self._is_layer_used(model_class, name):
|
| 658 |
+
self.assertIsNotNone(param.grad, f"Gradients should not be None - got {param.grad} for {name}")
|
| 659 |
+
|
| 660 |
+
@slow
|
| 661 |
+
def test_inference_interpolate_pos_encoding(self):
|
| 662 |
+
# ViT models have an `interpolate_pos_encoding` argument in their forward method,
|
| 663 |
+
# allowing to interpolate the pre-trained position embeddings in order to use
|
| 664 |
+
# the model on higher resolutions. The DINO model by Facebook AI leverages this
|
| 665 |
+
# to visualize self-attention on higher resolution images.
|
| 666 |
+
model_name = "BridgeTower/bridgetower-base"
|
| 667 |
+
model = BridgeTowerModel.from_pretrained(model_name).to(torch_device)
|
| 668 |
+
|
| 669 |
+
image_processor = BridgeTowerProcessor.from_pretrained(model_name, size={"shortest_edge": 180})
|
| 670 |
+
|
| 671 |
+
image = Image.open("./tests/fixtures/tests_samples/COCO/000000039769.png")
|
| 672 |
+
inputs = image_processor(text="what's in the image", images=image, return_tensors="pt").to(torch_device)
|
| 673 |
+
|
| 674 |
+
# interpolate_pos_encodiung false should return value error
|
| 675 |
+
with self.assertRaises(ValueError, msg="doesn't match model"):
|
| 676 |
+
with torch.no_grad():
|
| 677 |
+
model(**inputs, interpolate_pos_encoding=False)
|
| 678 |
+
|
| 679 |
+
# forward pass
|
| 680 |
+
with torch.no_grad():
|
| 681 |
+
outputs = model(**inputs, interpolate_pos_encoding=True)
|
| 682 |
+
|
| 683 |
+
# verify the logits
|
| 684 |
+
expected_shape = torch.Size((1, 122, 768))
|
| 685 |
+
|
| 686 |
+
self.assertEqual(outputs.image_features.shape, expected_shape)
|
| 687 |
+
|
| 688 |
+
expected_slice = torch.tensor(
|
| 689 |
+
[[-0.6518, 0.4978, -0.4544], [-2.6672, -0.0843, -0.4210], [-2.4510, -0.1002, -0.3458]]
|
| 690 |
+
).to(torch_device)
|
| 691 |
+
|
| 692 |
+
self.assertTrue(torch.allclose(outputs.image_features[0, :3, :3], expected_slice, atol=1e-4))
|
tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/bridgetower/test_processor_bridgetower.py
ADDED
|
@@ -0,0 +1,203 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Copyright 2023 The HuggingFace Team. All rights reserved.
|
| 2 |
+
#
|
| 3 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
| 4 |
+
# you may not use this file except in compliance with the License.
|
| 5 |
+
# You may obtain a copy of the License at
|
| 6 |
+
#
|
| 7 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
| 8 |
+
#
|
| 9 |
+
# Unless required by applicable law or agreed to in writing, software
|
| 10 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
| 11 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
| 12 |
+
# See the License for the specific language governing permissions and
|
| 13 |
+
# limitations under the License.
|
| 14 |
+
import shutil
|
| 15 |
+
import tempfile
|
| 16 |
+
import unittest
|
| 17 |
+
|
| 18 |
+
from transformers.testing_utils import require_torch, require_vision
|
| 19 |
+
from transformers.utils import is_vision_available
|
| 20 |
+
|
| 21 |
+
from ...test_processing_common import ProcessorTesterMixin
|
| 22 |
+
|
| 23 |
+
|
| 24 |
+
if is_vision_available():
|
| 25 |
+
from transformers import (
|
| 26 |
+
AutoProcessor,
|
| 27 |
+
BridgeTowerImageProcessor,
|
| 28 |
+
BridgeTowerProcessor,
|
| 29 |
+
RobertaTokenizerFast,
|
| 30 |
+
)
|
| 31 |
+
|
| 32 |
+
|
| 33 |
+
@require_vision
|
| 34 |
+
class BridgeTowerProcessorTest(ProcessorTesterMixin, unittest.TestCase):
|
| 35 |
+
processor_class = BridgeTowerProcessor
|
| 36 |
+
|
| 37 |
+
def setUp(self):
|
| 38 |
+
self.tmpdirname = tempfile.mkdtemp()
|
| 39 |
+
|
| 40 |
+
image_processor = BridgeTowerImageProcessor()
|
| 41 |
+
tokenizer = RobertaTokenizerFast.from_pretrained("BridgeTower/bridgetower-large-itm-mlm-itc")
|
| 42 |
+
|
| 43 |
+
processor = BridgeTowerProcessor(image_processor, tokenizer)
|
| 44 |
+
|
| 45 |
+
processor.save_pretrained(self.tmpdirname)
|
| 46 |
+
|
| 47 |
+
def get_tokenizer(self, **kwargs):
|
| 48 |
+
return AutoProcessor.from_pretrained(self.tmpdirname, **kwargs).tokenizer
|
| 49 |
+
|
| 50 |
+
def get_image_processor(self, **kwargs):
|
| 51 |
+
return AutoProcessor.from_pretrained(self.tmpdirname, **kwargs).image_processor
|
| 52 |
+
|
| 53 |
+
def tearDown(self):
|
| 54 |
+
shutil.rmtree(self.tmpdirname)
|
| 55 |
+
|
| 56 |
+
# Some kwargs tests are overriden from common tests to handle shortest_edge
|
| 57 |
+
# and size_divisor behaviour
|
| 58 |
+
|
| 59 |
+
@require_torch
|
| 60 |
+
@require_vision
|
| 61 |
+
def test_image_processor_defaults_preserved_by_image_kwargs(self):
|
| 62 |
+
if "image_processor" not in self.processor_class.attributes:
|
| 63 |
+
self.skipTest(f"image_processor attribute not present in {self.processor_class}")
|
| 64 |
+
image_processor = self.get_component(
|
| 65 |
+
"image_processor",
|
| 66 |
+
crop_size={"shortest_edge": 234, "longest_edge": 234},
|
| 67 |
+
)
|
| 68 |
+
tokenizer = self.get_component("tokenizer", max_length=117, padding="max_length")
|
| 69 |
+
|
| 70 |
+
processor = self.processor_class(tokenizer=tokenizer, image_processor=image_processor)
|
| 71 |
+
self.skip_processor_without_typed_kwargs(processor)
|
| 72 |
+
|
| 73 |
+
input_str = "lower newer"
|
| 74 |
+
image_input = self.prepare_image_inputs()
|
| 75 |
+
|
| 76 |
+
inputs = processor(text=input_str, images=image_input)
|
| 77 |
+
self.assertEqual(len(inputs["pixel_values"][0][0]), 234)
|
| 78 |
+
|
| 79 |
+
@require_torch
|
| 80 |
+
@require_vision
|
| 81 |
+
def test_structured_kwargs_nested_from_dict(self):
|
| 82 |
+
if "image_processor" not in self.processor_class.attributes:
|
| 83 |
+
self.skipTest(f"image_processor attribute not present in {self.processor_class}")
|
| 84 |
+
|
| 85 |
+
image_processor = self.get_component("image_processor")
|
| 86 |
+
tokenizer = self.get_component("tokenizer")
|
| 87 |
+
|
| 88 |
+
processor = self.processor_class(tokenizer=tokenizer, image_processor=image_processor)
|
| 89 |
+
self.skip_processor_without_typed_kwargs(processor)
|
| 90 |
+
input_str = "lower newer"
|
| 91 |
+
image_input = self.prepare_image_inputs()
|
| 92 |
+
|
| 93 |
+
# Define the kwargs for each modality
|
| 94 |
+
all_kwargs = {
|
| 95 |
+
"common_kwargs": {"return_tensors": "pt"},
|
| 96 |
+
"images_kwargs": {
|
| 97 |
+
"crop_size": {"shortest_edge": 214},
|
| 98 |
+
},
|
| 99 |
+
"text_kwargs": {"padding": "max_length", "max_length": 76},
|
| 100 |
+
}
|
| 101 |
+
|
| 102 |
+
inputs = processor(text=input_str, images=image_input, **all_kwargs)
|
| 103 |
+
self.assertEqual(inputs["pixel_values"].shape[2], 214)
|
| 104 |
+
|
| 105 |
+
self.assertEqual(len(inputs["input_ids"][0]), 76)
|
| 106 |
+
|
| 107 |
+
@require_torch
|
| 108 |
+
@require_vision
|
| 109 |
+
def test_kwargs_overrides_default_image_processor_kwargs(self):
|
| 110 |
+
if "image_processor" not in self.processor_class.attributes:
|
| 111 |
+
self.skipTest(f"image_processor attribute not present in {self.processor_class}")
|
| 112 |
+
image_processor = self.get_component("image_processor", crop_size={"shortest_edge": 234})
|
| 113 |
+
tokenizer = self.get_component("tokenizer", max_length=117)
|
| 114 |
+
if not tokenizer.pad_token:
|
| 115 |
+
tokenizer.pad_token = "[TEST_PAD]"
|
| 116 |
+
processor = self.processor_class(tokenizer=tokenizer, image_processor=image_processor)
|
| 117 |
+
self.skip_processor_without_typed_kwargs(processor)
|
| 118 |
+
|
| 119 |
+
input_str = "lower newer"
|
| 120 |
+
image_input = self.prepare_image_inputs()
|
| 121 |
+
inputs = processor(text=input_str, images=image_input, crop_size={"shortest_edge": 224})
|
| 122 |
+
self.assertEqual(len(inputs["pixel_values"][0][0]), 224)
|
| 123 |
+
|
| 124 |
+
@require_torch
|
| 125 |
+
@require_vision
|
| 126 |
+
def test_unstructured_kwargs_batched(self):
|
| 127 |
+
if "image_processor" not in self.processor_class.attributes:
|
| 128 |
+
self.skipTest(f"image_processor attribute not present in {self.processor_class}")
|
| 129 |
+
image_processor = self.get_component("image_processor")
|
| 130 |
+
tokenizer = self.get_component("tokenizer")
|
| 131 |
+
if not tokenizer.pad_token:
|
| 132 |
+
tokenizer.pad_token = "[TEST_PAD]"
|
| 133 |
+
processor = self.processor_class(tokenizer=tokenizer, image_processor=image_processor)
|
| 134 |
+
self.skip_processor_without_typed_kwargs(processor)
|
| 135 |
+
|
| 136 |
+
input_str = ["lower newer", "upper older longer string"]
|
| 137 |
+
image_input = self.prepare_image_inputs(batch_size=2)
|
| 138 |
+
inputs = processor(
|
| 139 |
+
text=input_str,
|
| 140 |
+
images=image_input,
|
| 141 |
+
return_tensors="pt",
|
| 142 |
+
crop_size={"shortest_edge": 214},
|
| 143 |
+
padding="longest",
|
| 144 |
+
max_length=76,
|
| 145 |
+
)
|
| 146 |
+
self.assertEqual(inputs["pixel_values"].shape[2], 214)
|
| 147 |
+
|
| 148 |
+
self.assertEqual(len(inputs["input_ids"][0]), 6)
|
| 149 |
+
|
| 150 |
+
@require_torch
|
| 151 |
+
@require_vision
|
| 152 |
+
def test_unstructured_kwargs(self):
|
| 153 |
+
if "image_processor" not in self.processor_class.attributes:
|
| 154 |
+
self.skipTest(f"image_processor attribute not present in {self.processor_class}")
|
| 155 |
+
image_processor = self.get_component("image_processor")
|
| 156 |
+
tokenizer = self.get_component("tokenizer")
|
| 157 |
+
if not tokenizer.pad_token:
|
| 158 |
+
tokenizer.pad_token = "[TEST_PAD]"
|
| 159 |
+
processor = self.processor_class(tokenizer=tokenizer, image_processor=image_processor)
|
| 160 |
+
self.skip_processor_without_typed_kwargs(processor)
|
| 161 |
+
|
| 162 |
+
input_str = "lower newer"
|
| 163 |
+
image_input = self.prepare_image_inputs()
|
| 164 |
+
inputs = processor(
|
| 165 |
+
text=input_str,
|
| 166 |
+
images=image_input,
|
| 167 |
+
return_tensors="pt",
|
| 168 |
+
crop_size={"shortest_edge": 214},
|
| 169 |
+
padding="max_length",
|
| 170 |
+
max_length=76,
|
| 171 |
+
)
|
| 172 |
+
|
| 173 |
+
self.assertEqual(inputs["pixel_values"].shape[2], 214)
|
| 174 |
+
self.assertEqual(len(inputs["input_ids"][0]), 76)
|
| 175 |
+
|
| 176 |
+
@require_torch
|
| 177 |
+
@require_vision
|
| 178 |
+
def test_structured_kwargs_nested(self):
|
| 179 |
+
if "image_processor" not in self.processor_class.attributes:
|
| 180 |
+
self.skipTest(f"image_processor attribute not present in {self.processor_class}")
|
| 181 |
+
image_processor = self.get_component("image_processor")
|
| 182 |
+
tokenizer = self.get_component("tokenizer")
|
| 183 |
+
if not tokenizer.pad_token:
|
| 184 |
+
tokenizer.pad_token = "[TEST_PAD]"
|
| 185 |
+
processor = self.processor_class(tokenizer=tokenizer, image_processor=image_processor)
|
| 186 |
+
self.skip_processor_without_typed_kwargs(processor)
|
| 187 |
+
|
| 188 |
+
input_str = "lower newer"
|
| 189 |
+
image_input = self.prepare_image_inputs()
|
| 190 |
+
|
| 191 |
+
# Define the kwargs for each modality
|
| 192 |
+
all_kwargs = {
|
| 193 |
+
"common_kwargs": {"return_tensors": "pt"},
|
| 194 |
+
"images_kwargs": {"crop_size": {"shortest_edge": 214}},
|
| 195 |
+
"text_kwargs": {"padding": "max_length", "max_length": 76},
|
| 196 |
+
}
|
| 197 |
+
|
| 198 |
+
inputs = processor(text=input_str, images=image_input, **all_kwargs)
|
| 199 |
+
self.skip_processor_without_typed_kwargs(processor)
|
| 200 |
+
|
| 201 |
+
self.assertEqual(inputs["pixel_values"].shape[2], 214)
|
| 202 |
+
|
| 203 |
+
self.assertEqual(len(inputs["input_ids"][0]), 76)
|
tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/bros/__init__.py
ADDED
|
File without changes
|
tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/bros/test_modeling_bros.py
ADDED
|
@@ -0,0 +1,455 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# coding=utf-8
|
| 2 |
+
# Copyright 2023 The HuggingFace Inc. team. All rights reserved.
|
| 3 |
+
#
|
| 4 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
| 5 |
+
# you may not use this file except in compliance with the License.
|
| 6 |
+
# You may obtain a copy of the License at
|
| 7 |
+
#
|
| 8 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
| 9 |
+
#
|
| 10 |
+
# Unless required by applicable law or agreed to in writing, software
|
| 11 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
| 12 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
| 13 |
+
# See the License for the specific language governing permissions and
|
| 14 |
+
# limitations under the License.
|
| 15 |
+
"""Testing suite for the PyTorch Bros model."""
|
| 16 |
+
|
| 17 |
+
import copy
|
| 18 |
+
import unittest
|
| 19 |
+
|
| 20 |
+
from transformers.testing_utils import require_torch, require_torch_multi_gpu, slow, torch_device
|
| 21 |
+
from transformers.utils import is_torch_available
|
| 22 |
+
|
| 23 |
+
from ...test_configuration_common import ConfigTester
|
| 24 |
+
from ...test_modeling_common import ModelTesterMixin, ids_tensor, random_attention_mask
|
| 25 |
+
from ...test_pipeline_mixin import PipelineTesterMixin
|
| 26 |
+
|
| 27 |
+
|
| 28 |
+
if is_torch_available():
|
| 29 |
+
import torch
|
| 30 |
+
|
| 31 |
+
from transformers import (
|
| 32 |
+
BrosConfig,
|
| 33 |
+
BrosForTokenClassification,
|
| 34 |
+
BrosModel,
|
| 35 |
+
BrosSpadeEEForTokenClassification,
|
| 36 |
+
BrosSpadeELForTokenClassification,
|
| 37 |
+
)
|
| 38 |
+
|
| 39 |
+
|
| 40 |
+
class BrosModelTester:
|
| 41 |
+
def __init__(
|
| 42 |
+
self,
|
| 43 |
+
parent,
|
| 44 |
+
batch_size=13,
|
| 45 |
+
seq_length=7,
|
| 46 |
+
is_training=True,
|
| 47 |
+
use_input_mask=True,
|
| 48 |
+
use_token_type_ids=True,
|
| 49 |
+
use_bbox_first_token_mask=True,
|
| 50 |
+
use_labels=True,
|
| 51 |
+
vocab_size=99,
|
| 52 |
+
hidden_size=64,
|
| 53 |
+
num_hidden_layers=5,
|
| 54 |
+
num_attention_heads=4,
|
| 55 |
+
intermediate_size=37,
|
| 56 |
+
hidden_act="gelu",
|
| 57 |
+
hidden_dropout_prob=0.1,
|
| 58 |
+
attention_probs_dropout_prob=0.1,
|
| 59 |
+
max_position_embeddings=512,
|
| 60 |
+
type_vocab_size=16,
|
| 61 |
+
type_sequence_label_size=2,
|
| 62 |
+
initializer_range=0.02,
|
| 63 |
+
num_labels=3,
|
| 64 |
+
num_choices=4,
|
| 65 |
+
scope=None,
|
| 66 |
+
):
|
| 67 |
+
self.parent = parent
|
| 68 |
+
self.batch_size = batch_size
|
| 69 |
+
self.seq_length = seq_length
|
| 70 |
+
self.is_training = is_training
|
| 71 |
+
self.use_input_mask = use_input_mask
|
| 72 |
+
self.use_bbox_first_token_mask = use_bbox_first_token_mask
|
| 73 |
+
self.use_token_type_ids = use_token_type_ids
|
| 74 |
+
self.use_labels = use_labels
|
| 75 |
+
self.vocab_size = vocab_size
|
| 76 |
+
self.hidden_size = hidden_size
|
| 77 |
+
self.num_hidden_layers = num_hidden_layers
|
| 78 |
+
self.num_attention_heads = num_attention_heads
|
| 79 |
+
self.intermediate_size = intermediate_size
|
| 80 |
+
self.hidden_act = hidden_act
|
| 81 |
+
self.hidden_dropout_prob = hidden_dropout_prob
|
| 82 |
+
self.attention_probs_dropout_prob = attention_probs_dropout_prob
|
| 83 |
+
self.max_position_embeddings = max_position_embeddings
|
| 84 |
+
self.type_vocab_size = type_vocab_size
|
| 85 |
+
self.type_sequence_label_size = type_sequence_label_size
|
| 86 |
+
self.initializer_range = initializer_range
|
| 87 |
+
self.num_labels = num_labels
|
| 88 |
+
|
| 89 |
+
def prepare_config_and_inputs(self):
|
| 90 |
+
input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
|
| 91 |
+
|
| 92 |
+
bbox = ids_tensor([self.batch_size, self.seq_length, 8], 1)
|
| 93 |
+
# Ensure that bbox is legal
|
| 94 |
+
for i in range(bbox.shape[0]):
|
| 95 |
+
for j in range(bbox.shape[1]):
|
| 96 |
+
if bbox[i, j, 3] < bbox[i, j, 1]:
|
| 97 |
+
t = bbox[i, j, 3]
|
| 98 |
+
bbox[i, j, 3] = bbox[i, j, 1]
|
| 99 |
+
bbox[i, j, 1] = t
|
| 100 |
+
if bbox[i, j, 2] < bbox[i, j, 0]:
|
| 101 |
+
t = bbox[i, j, 2]
|
| 102 |
+
bbox[i, j, 2] = bbox[i, j, 0]
|
| 103 |
+
bbox[i, j, 0] = t
|
| 104 |
+
|
| 105 |
+
input_mask = None
|
| 106 |
+
if self.use_input_mask:
|
| 107 |
+
input_mask = random_attention_mask([self.batch_size, self.seq_length])
|
| 108 |
+
|
| 109 |
+
bbox_first_token_mask = None
|
| 110 |
+
if self.use_bbox_first_token_mask:
|
| 111 |
+
bbox_first_token_mask = torch.ones([self.batch_size, self.seq_length], dtype=torch.bool).to(torch_device)
|
| 112 |
+
|
| 113 |
+
token_type_ids = None
|
| 114 |
+
if self.use_token_type_ids:
|
| 115 |
+
token_type_ids = ids_tensor([self.batch_size, self.seq_length], self.type_vocab_size)
|
| 116 |
+
|
| 117 |
+
token_labels = None
|
| 118 |
+
if self.use_labels:
|
| 119 |
+
token_labels = ids_tensor([self.batch_size, self.seq_length], self.num_labels)
|
| 120 |
+
initial_token_labels = ids_tensor([self.batch_size, self.seq_length], self.num_labels)
|
| 121 |
+
subsequent_token_labels = ids_tensor([self.batch_size, self.seq_length], self.num_labels)
|
| 122 |
+
|
| 123 |
+
config = self.get_config()
|
| 124 |
+
|
| 125 |
+
return (
|
| 126 |
+
config,
|
| 127 |
+
input_ids,
|
| 128 |
+
bbox,
|
| 129 |
+
token_type_ids,
|
| 130 |
+
input_mask,
|
| 131 |
+
bbox_first_token_mask,
|
| 132 |
+
token_labels,
|
| 133 |
+
initial_token_labels,
|
| 134 |
+
subsequent_token_labels,
|
| 135 |
+
)
|
| 136 |
+
|
| 137 |
+
def get_config(self):
|
| 138 |
+
return BrosConfig(
|
| 139 |
+
vocab_size=self.vocab_size,
|
| 140 |
+
hidden_size=self.hidden_size,
|
| 141 |
+
num_hidden_layers=self.num_hidden_layers,
|
| 142 |
+
num_attention_heads=self.num_attention_heads,
|
| 143 |
+
intermediate_size=self.intermediate_size,
|
| 144 |
+
hidden_act=self.hidden_act,
|
| 145 |
+
hidden_dropout_prob=self.hidden_dropout_prob,
|
| 146 |
+
attention_probs_dropout_prob=self.attention_probs_dropout_prob,
|
| 147 |
+
max_position_embeddings=self.max_position_embeddings,
|
| 148 |
+
type_vocab_size=self.type_vocab_size,
|
| 149 |
+
is_decoder=False,
|
| 150 |
+
initializer_range=self.initializer_range,
|
| 151 |
+
)
|
| 152 |
+
|
| 153 |
+
def create_and_check_model(
|
| 154 |
+
self,
|
| 155 |
+
config,
|
| 156 |
+
input_ids,
|
| 157 |
+
bbox,
|
| 158 |
+
token_type_ids,
|
| 159 |
+
input_mask,
|
| 160 |
+
bbox_first_token_mask,
|
| 161 |
+
token_labels,
|
| 162 |
+
initial_token_labels,
|
| 163 |
+
subsequent_token_labels,
|
| 164 |
+
):
|
| 165 |
+
model = BrosModel(config=config)
|
| 166 |
+
model.to(torch_device)
|
| 167 |
+
model.eval()
|
| 168 |
+
result = model(input_ids, bbox=bbox, attention_mask=input_mask, token_type_ids=token_type_ids)
|
| 169 |
+
result = model(input_ids, bbox=bbox, token_type_ids=token_type_ids)
|
| 170 |
+
result = model(input_ids, bbox=bbox)
|
| 171 |
+
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
|
| 172 |
+
|
| 173 |
+
def create_and_check_for_token_classification(
|
| 174 |
+
self,
|
| 175 |
+
config,
|
| 176 |
+
input_ids,
|
| 177 |
+
bbox,
|
| 178 |
+
token_type_ids,
|
| 179 |
+
input_mask,
|
| 180 |
+
bbox_first_token_mask,
|
| 181 |
+
token_labels,
|
| 182 |
+
initial_token_labels,
|
| 183 |
+
subsequent_token_labels,
|
| 184 |
+
):
|
| 185 |
+
config.num_labels = self.num_labels
|
| 186 |
+
model = BrosForTokenClassification(config=config)
|
| 187 |
+
model.to(torch_device)
|
| 188 |
+
model.eval()
|
| 189 |
+
result = model(
|
| 190 |
+
input_ids, bbox=bbox, attention_mask=input_mask, token_type_ids=token_type_ids, labels=token_labels
|
| 191 |
+
)
|
| 192 |
+
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.num_labels))
|
| 193 |
+
|
| 194 |
+
def create_and_check_for_spade_ee_token_classification(
|
| 195 |
+
self,
|
| 196 |
+
config,
|
| 197 |
+
input_ids,
|
| 198 |
+
bbox,
|
| 199 |
+
token_type_ids,
|
| 200 |
+
input_mask,
|
| 201 |
+
bbox_first_token_mask,
|
| 202 |
+
token_labels,
|
| 203 |
+
initial_token_labels,
|
| 204 |
+
subsequent_token_labels,
|
| 205 |
+
):
|
| 206 |
+
config.num_labels = self.num_labels
|
| 207 |
+
model = BrosSpadeEEForTokenClassification(config=config)
|
| 208 |
+
model.to(torch_device)
|
| 209 |
+
model.eval()
|
| 210 |
+
result = model(
|
| 211 |
+
input_ids,
|
| 212 |
+
bbox=bbox,
|
| 213 |
+
attention_mask=input_mask,
|
| 214 |
+
bbox_first_token_mask=bbox_first_token_mask,
|
| 215 |
+
token_type_ids=token_type_ids,
|
| 216 |
+
initial_token_labels=token_labels,
|
| 217 |
+
subsequent_token_labels=token_labels,
|
| 218 |
+
)
|
| 219 |
+
self.parent.assertEqual(result.initial_token_logits.shape, (self.batch_size, self.seq_length, self.num_labels))
|
| 220 |
+
self.parent.assertEqual(
|
| 221 |
+
result.subsequent_token_logits.shape, (self.batch_size, self.seq_length, self.seq_length + 1)
|
| 222 |
+
)
|
| 223 |
+
|
| 224 |
+
def create_and_check_for_spade_el_token_classification(
|
| 225 |
+
self,
|
| 226 |
+
config,
|
| 227 |
+
input_ids,
|
| 228 |
+
bbox,
|
| 229 |
+
token_type_ids,
|
| 230 |
+
input_mask,
|
| 231 |
+
bbox_first_token_mask,
|
| 232 |
+
token_labels,
|
| 233 |
+
initial_token_labels,
|
| 234 |
+
subsequent_token_labels,
|
| 235 |
+
):
|
| 236 |
+
config.num_labels = self.num_labels
|
| 237 |
+
model = BrosSpadeELForTokenClassification(config=config)
|
| 238 |
+
model.to(torch_device)
|
| 239 |
+
model.eval()
|
| 240 |
+
result = model(
|
| 241 |
+
input_ids,
|
| 242 |
+
bbox=bbox,
|
| 243 |
+
attention_mask=input_mask,
|
| 244 |
+
bbox_first_token_mask=bbox_first_token_mask,
|
| 245 |
+
token_type_ids=token_type_ids,
|
| 246 |
+
labels=token_labels,
|
| 247 |
+
)
|
| 248 |
+
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.seq_length + 1))
|
| 249 |
+
|
| 250 |
+
def prepare_config_and_inputs_for_common(self):
|
| 251 |
+
config_and_inputs = self.prepare_config_and_inputs()
|
| 252 |
+
(
|
| 253 |
+
config,
|
| 254 |
+
input_ids,
|
| 255 |
+
bbox,
|
| 256 |
+
token_type_ids,
|
| 257 |
+
input_mask,
|
| 258 |
+
bbox_first_token_mask,
|
| 259 |
+
token_labels,
|
| 260 |
+
initial_token_labels,
|
| 261 |
+
subsequent_token_labels,
|
| 262 |
+
) = config_and_inputs
|
| 263 |
+
inputs_dict = {
|
| 264 |
+
"input_ids": input_ids,
|
| 265 |
+
"bbox": bbox,
|
| 266 |
+
"token_type_ids": token_type_ids,
|
| 267 |
+
"attention_mask": input_mask,
|
| 268 |
+
}
|
| 269 |
+
return config, inputs_dict
|
| 270 |
+
|
| 271 |
+
|
| 272 |
+
@require_torch
|
| 273 |
+
class BrosModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase):
|
| 274 |
+
test_pruning = False
|
| 275 |
+
test_torchscript = False
|
| 276 |
+
test_mismatched_shapes = False
|
| 277 |
+
|
| 278 |
+
all_model_classes = (
|
| 279 |
+
(
|
| 280 |
+
BrosForTokenClassification,
|
| 281 |
+
BrosSpadeEEForTokenClassification,
|
| 282 |
+
BrosSpadeELForTokenClassification,
|
| 283 |
+
BrosModel,
|
| 284 |
+
)
|
| 285 |
+
if is_torch_available()
|
| 286 |
+
else ()
|
| 287 |
+
)
|
| 288 |
+
all_generative_model_classes = () if is_torch_available() else ()
|
| 289 |
+
pipeline_model_mapping = (
|
| 290 |
+
{"feature-extraction": BrosModel, "token-classification": BrosForTokenClassification}
|
| 291 |
+
if is_torch_available()
|
| 292 |
+
else {}
|
| 293 |
+
)
|
| 294 |
+
|
| 295 |
+
# BROS requires `bbox` in the inputs which doesn't fit into the above 2 pipelines' input formats.
|
| 296 |
+
# see https://github.com/huggingface/transformers/pull/26294
|
| 297 |
+
def is_pipeline_test_to_skip(
|
| 298 |
+
self,
|
| 299 |
+
pipeline_test_case_name,
|
| 300 |
+
config_class,
|
| 301 |
+
model_architecture,
|
| 302 |
+
tokenizer_name,
|
| 303 |
+
image_processor_name,
|
| 304 |
+
feature_extractor_name,
|
| 305 |
+
processor_name,
|
| 306 |
+
):
|
| 307 |
+
return True
|
| 308 |
+
|
| 309 |
+
def setUp(self):
|
| 310 |
+
self.model_tester = BrosModelTester(self)
|
| 311 |
+
self.config_tester = ConfigTester(self, config_class=BrosConfig, hidden_size=37)
|
| 312 |
+
|
| 313 |
+
def _prepare_for_class(self, inputs_dict, model_class, return_labels=False):
|
| 314 |
+
inputs_dict = copy.deepcopy(inputs_dict)
|
| 315 |
+
|
| 316 |
+
if return_labels:
|
| 317 |
+
if model_class.__name__ in ["BrosForTokenClassification", "BrosSpadeELForTokenClassification"]:
|
| 318 |
+
inputs_dict["labels"] = torch.zeros(
|
| 319 |
+
(self.model_tester.batch_size, self.model_tester.seq_length),
|
| 320 |
+
dtype=torch.long,
|
| 321 |
+
device=torch_device,
|
| 322 |
+
)
|
| 323 |
+
inputs_dict["bbox_first_token_mask"] = torch.ones(
|
| 324 |
+
[self.model_tester.batch_size, self.model_tester.seq_length],
|
| 325 |
+
dtype=torch.bool,
|
| 326 |
+
device=torch_device,
|
| 327 |
+
)
|
| 328 |
+
elif model_class.__name__ in ["BrosSpadeEEForTokenClassification"]:
|
| 329 |
+
inputs_dict["initial_token_labels"] = torch.zeros(
|
| 330 |
+
(self.model_tester.batch_size, self.model_tester.seq_length),
|
| 331 |
+
dtype=torch.long,
|
| 332 |
+
device=torch_device,
|
| 333 |
+
)
|
| 334 |
+
inputs_dict["subsequent_token_labels"] = torch.zeros(
|
| 335 |
+
(self.model_tester.batch_size, self.model_tester.seq_length),
|
| 336 |
+
dtype=torch.long,
|
| 337 |
+
device=torch_device,
|
| 338 |
+
)
|
| 339 |
+
inputs_dict["bbox_first_token_mask"] = torch.ones(
|
| 340 |
+
[self.model_tester.batch_size, self.model_tester.seq_length],
|
| 341 |
+
dtype=torch.bool,
|
| 342 |
+
device=torch_device,
|
| 343 |
+
)
|
| 344 |
+
|
| 345 |
+
return inputs_dict
|
| 346 |
+
|
| 347 |
+
def test_config(self):
|
| 348 |
+
self.config_tester.run_common_tests()
|
| 349 |
+
|
| 350 |
+
def test_model(self):
|
| 351 |
+
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
| 352 |
+
self.model_tester.create_and_check_model(*config_and_inputs)
|
| 353 |
+
|
| 354 |
+
@require_torch_multi_gpu
|
| 355 |
+
def test_multi_gpu_data_parallel_forward(self):
|
| 356 |
+
super().test_multi_gpu_data_parallel_forward()
|
| 357 |
+
|
| 358 |
+
def test_model_various_embeddings(self):
|
| 359 |
+
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
| 360 |
+
for type in ["absolute", "relative_key", "relative_key_query"]:
|
| 361 |
+
config_and_inputs[0].position_embedding_type = type
|
| 362 |
+
self.model_tester.create_and_check_model(*config_and_inputs)
|
| 363 |
+
|
| 364 |
+
def test_for_token_classification(self):
|
| 365 |
+
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
| 366 |
+
self.model_tester.create_and_check_for_token_classification(*config_and_inputs)
|
| 367 |
+
|
| 368 |
+
def test_for_spade_ee_token_classification(self):
|
| 369 |
+
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
| 370 |
+
self.model_tester.create_and_check_for_spade_ee_token_classification(*config_and_inputs)
|
| 371 |
+
|
| 372 |
+
def test_for_spade_el_token_classification(self):
|
| 373 |
+
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
| 374 |
+
self.model_tester.create_and_check_for_spade_el_token_classification(*config_and_inputs)
|
| 375 |
+
|
| 376 |
+
@slow
|
| 377 |
+
def test_model_from_pretrained(self):
|
| 378 |
+
model_name = "jinho8345/bros-base-uncased"
|
| 379 |
+
model = BrosModel.from_pretrained(model_name)
|
| 380 |
+
self.assertIsNotNone(model)
|
| 381 |
+
|
| 382 |
+
|
| 383 |
+
def prepare_bros_batch_inputs():
|
| 384 |
+
attention_mask = torch.tensor([[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]])
|
| 385 |
+
|
| 386 |
+
bbox = torch.tensor(
|
| 387 |
+
[
|
| 388 |
+
[
|
| 389 |
+
[0.0000, 0.0000, 0.0000, 0.0000],
|
| 390 |
+
[0.5223, 0.5590, 0.5787, 0.5720],
|
| 391 |
+
[0.5853, 0.5590, 0.6864, 0.5720],
|
| 392 |
+
[0.5853, 0.5590, 0.6864, 0.5720],
|
| 393 |
+
[0.1234, 0.5700, 0.2192, 0.5840],
|
| 394 |
+
[0.2231, 0.5680, 0.2782, 0.5780],
|
| 395 |
+
[0.2874, 0.5670, 0.3333, 0.5780],
|
| 396 |
+
[0.3425, 0.5640, 0.4344, 0.5750],
|
| 397 |
+
[0.0866, 0.7770, 0.1181, 0.7870],
|
| 398 |
+
[0.1168, 0.7770, 0.1522, 0.7850],
|
| 399 |
+
[0.1535, 0.7750, 0.1864, 0.7850],
|
| 400 |
+
[0.1890, 0.7750, 0.2572, 0.7850],
|
| 401 |
+
[1.0000, 1.0000, 1.0000, 1.0000],
|
| 402 |
+
],
|
| 403 |
+
[
|
| 404 |
+
[0.0000, 0.0000, 0.0000, 0.0000],
|
| 405 |
+
[0.4396, 0.6720, 0.4659, 0.6850],
|
| 406 |
+
[0.4698, 0.6720, 0.4843, 0.6850],
|
| 407 |
+
[0.1575, 0.6870, 0.2021, 0.6980],
|
| 408 |
+
[0.2047, 0.6870, 0.2730, 0.7000],
|
| 409 |
+
[0.1299, 0.7010, 0.1430, 0.7140],
|
| 410 |
+
[0.1299, 0.7010, 0.1430, 0.7140],
|
| 411 |
+
[0.1562, 0.7010, 0.2441, 0.7120],
|
| 412 |
+
[0.1562, 0.7010, 0.2441, 0.7120],
|
| 413 |
+
[0.2454, 0.7010, 0.3150, 0.7120],
|
| 414 |
+
[0.3176, 0.7010, 0.3320, 0.7110],
|
| 415 |
+
[0.3333, 0.7000, 0.4029, 0.7140],
|
| 416 |
+
[1.0000, 1.0000, 1.0000, 1.0000],
|
| 417 |
+
],
|
| 418 |
+
]
|
| 419 |
+
)
|
| 420 |
+
input_ids = torch.tensor(
|
| 421 |
+
[
|
| 422 |
+
[101, 1055, 8910, 1012, 5719, 3296, 5366, 3378, 2146, 2846, 10807, 13494, 102],
|
| 423 |
+
[101, 2112, 1997, 3671, 6364, 1019, 1012, 5057, 1011, 4646, 2030, 2974, 102],
|
| 424 |
+
]
|
| 425 |
+
)
|
| 426 |
+
|
| 427 |
+
return input_ids, bbox, attention_mask
|
| 428 |
+
|
| 429 |
+
|
| 430 |
+
@require_torch
|
| 431 |
+
class BrosModelIntegrationTest(unittest.TestCase):
|
| 432 |
+
@slow
|
| 433 |
+
def test_inference_no_head(self):
|
| 434 |
+
model = BrosModel.from_pretrained("jinho8345/bros-base-uncased").to(torch_device)
|
| 435 |
+
|
| 436 |
+
input_ids, bbox, attention_mask = prepare_bros_batch_inputs()
|
| 437 |
+
|
| 438 |
+
with torch.no_grad():
|
| 439 |
+
outputs = model(
|
| 440 |
+
input_ids.to(torch_device),
|
| 441 |
+
bbox.to(torch_device),
|
| 442 |
+
attention_mask=attention_mask.to(torch_device),
|
| 443 |
+
return_dict=True,
|
| 444 |
+
)
|
| 445 |
+
|
| 446 |
+
# verify the logits
|
| 447 |
+
expected_shape = torch.Size((2, 13, 768))
|
| 448 |
+
self.assertEqual(outputs.last_hidden_state.shape, expected_shape)
|
| 449 |
+
|
| 450 |
+
expected_slice = torch.tensor(
|
| 451 |
+
[[-0.3074, 0.1363, 0.3143], [0.0925, -0.1155, 0.1050], [0.0221, 0.0003, 0.1285]]
|
| 452 |
+
).to(torch_device)
|
| 453 |
+
torch.set_printoptions(sci_mode=False)
|
| 454 |
+
|
| 455 |
+
self.assertTrue(torch.allclose(outputs.last_hidden_state[0, :3, :3], expected_slice, atol=1e-4))
|
tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/byt5/__init__.py
ADDED
|
File without changes
|
tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/byt5/test_tokenization_byt5.py
ADDED
|
@@ -0,0 +1,362 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# coding=utf-8
|
| 2 |
+
# Copyright 2020 Google T5 Authors and HuggingFace Inc. team.
|
| 3 |
+
#
|
| 4 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
| 5 |
+
# you may not use this file except in compliance with the License.
|
| 6 |
+
# You may obtain a copy of the License at
|
| 7 |
+
#
|
| 8 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
| 9 |
+
#
|
| 10 |
+
# Unless required by applicable law or agreed to in writing, software
|
| 11 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
| 12 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
| 13 |
+
# See the License for the specific language governing permissions and
|
| 14 |
+
# limitations under the License.
|
| 15 |
+
|
| 16 |
+
import json
|
| 17 |
+
import os
|
| 18 |
+
import re
|
| 19 |
+
import shutil
|
| 20 |
+
import tempfile
|
| 21 |
+
import unittest
|
| 22 |
+
from typing import Tuple
|
| 23 |
+
|
| 24 |
+
from transformers import AddedToken, BatchEncoding, ByT5Tokenizer
|
| 25 |
+
from transformers.utils import cached_property, is_tf_available, is_torch_available
|
| 26 |
+
|
| 27 |
+
from ...test_tokenization_common import TokenizerTesterMixin
|
| 28 |
+
|
| 29 |
+
|
| 30 |
+
if is_torch_available():
|
| 31 |
+
FRAMEWORK = "pt"
|
| 32 |
+
elif is_tf_available():
|
| 33 |
+
FRAMEWORK = "tf"
|
| 34 |
+
else:
|
| 35 |
+
FRAMEWORK = "jax"
|
| 36 |
+
|
| 37 |
+
|
| 38 |
+
class ByT5TokenizationTest(TokenizerTesterMixin, unittest.TestCase):
|
| 39 |
+
tokenizer_class = ByT5Tokenizer
|
| 40 |
+
test_rust_tokenizer = False
|
| 41 |
+
|
| 42 |
+
def setUp(self):
|
| 43 |
+
super().setUp()
|
| 44 |
+
tokenizer = ByT5Tokenizer()
|
| 45 |
+
tokenizer.save_pretrained(self.tmpdirname)
|
| 46 |
+
|
| 47 |
+
@cached_property
|
| 48 |
+
def t5_base_tokenizer(self):
|
| 49 |
+
return ByT5Tokenizer.from_pretrained("google/byt5-small")
|
| 50 |
+
|
| 51 |
+
def get_tokenizer(self, **kwargs) -> ByT5Tokenizer:
|
| 52 |
+
return self.tokenizer_class.from_pretrained(self.tmpdirname, **kwargs)
|
| 53 |
+
|
| 54 |
+
def get_clean_sequence(self, tokenizer, with_prefix_space=False, max_length=20, min_length=5) -> Tuple[str, list]:
|
| 55 |
+
# XXX The default common tokenizer tests assume that every ID is decodable on its own.
|
| 56 |
+
# This assumption is invalid for ByT5 because single bytes might not be
|
| 57 |
+
# valid utf-8 (byte 128 for instance).
|
| 58 |
+
# Here we're overriding the smallest possible method to provide
|
| 59 |
+
# a clean sequence without making the same assumption.
|
| 60 |
+
|
| 61 |
+
toks = []
|
| 62 |
+
for i in range(len(tokenizer)):
|
| 63 |
+
try:
|
| 64 |
+
tok = tokenizer.decode([i], clean_up_tokenization_spaces=False)
|
| 65 |
+
except UnicodeDecodeError:
|
| 66 |
+
pass
|
| 67 |
+
toks.append((i, tok))
|
| 68 |
+
|
| 69 |
+
toks = list(filter(lambda t: re.match(r"^[ a-zA-Z]+$", t[1]), toks))
|
| 70 |
+
toks = list(filter(lambda t: [t[0]] == tokenizer.encode(t[1], add_special_tokens=False), toks))
|
| 71 |
+
if max_length is not None and len(toks) > max_length:
|
| 72 |
+
toks = toks[:max_length]
|
| 73 |
+
if min_length is not None and len(toks) < min_length and len(toks) > 0:
|
| 74 |
+
while len(toks) < min_length:
|
| 75 |
+
toks = toks + toks
|
| 76 |
+
# toks_str = [t[1] for t in toks]
|
| 77 |
+
toks_ids = [t[0] for t in toks]
|
| 78 |
+
|
| 79 |
+
# Ensure consistency
|
| 80 |
+
output_txt = tokenizer.decode(toks_ids, clean_up_tokenization_spaces=False)
|
| 81 |
+
if " " not in output_txt and len(toks_ids) > 1:
|
| 82 |
+
output_txt = (
|
| 83 |
+
tokenizer.decode([toks_ids[0]], clean_up_tokenization_spaces=False)
|
| 84 |
+
+ " "
|
| 85 |
+
+ tokenizer.decode(toks_ids[1:], clean_up_tokenization_spaces=False)
|
| 86 |
+
)
|
| 87 |
+
if with_prefix_space:
|
| 88 |
+
output_txt = " " + output_txt
|
| 89 |
+
output_ids = tokenizer.encode(output_txt, add_special_tokens=False)
|
| 90 |
+
return output_txt, output_ids
|
| 91 |
+
|
| 92 |
+
def test_eos_treatment(self):
|
| 93 |
+
tokenizer = self.t5_base_tokenizer
|
| 94 |
+
batch_with_eos_added = tokenizer(["hi</s>", "I went to the gym</s>", "</s>"])
|
| 95 |
+
batch_without_eos_added = tokenizer(["hi", "I went to the gym", ""])
|
| 96 |
+
self.assertListEqual(batch_with_eos_added["input_ids"], batch_without_eos_added["input_ids"])
|
| 97 |
+
|
| 98 |
+
def test_multibytes_char(self):
|
| 99 |
+
tokenizer = self.t5_base_tokenizer
|
| 100 |
+
src_text = "Unicode €."
|
| 101 |
+
encoded = tokenizer(src_text)
|
| 102 |
+
encoded_ids = [88, 113, 108, 102, 114, 103, 104, 35, 229, 133, 175, 49, 1]
|
| 103 |
+
self.assertEqual(encoded["input_ids"], encoded_ids)
|
| 104 |
+
|
| 105 |
+
# decoding
|
| 106 |
+
decoded = tokenizer.decode(encoded_ids)
|
| 107 |
+
self.assertEqual(decoded, "Unicode €.</s>")
|
| 108 |
+
|
| 109 |
+
encoded = tokenizer("e è é ê ë")
|
| 110 |
+
encoded_ids = [104, 35, 198, 171, 35, 198, 172, 35, 198, 173, 35, 198, 174, 1]
|
| 111 |
+
self.assertEqual(encoded["input_ids"], encoded_ids)
|
| 112 |
+
# decoding
|
| 113 |
+
decoded = tokenizer.decode(encoded_ids)
|
| 114 |
+
self.assertEqual(decoded, "e è é ê ë</s>")
|
| 115 |
+
|
| 116 |
+
# encode/decode, but with `encode` instead of `__call__`
|
| 117 |
+
self.assertEqual(tokenizer.decode(tokenizer.encode("e è é ê ë")), "e è é ê ë</s>")
|
| 118 |
+
|
| 119 |
+
def test_prepare_batch_integration(self):
|
| 120 |
+
tokenizer = self.t5_base_tokenizer
|
| 121 |
+
src_text = ["A long paragraph for summarization.", "Another paragraph for summarization."]
|
| 122 |
+
expected_src_tokens = [68, 35, 111, 114, 113, 106, 35, 115, 100, 117, 100, 106, 117, 100, 115, 107, 35, 105, 114, 117, 35, 118, 120, 112, 112, 100, 117, 108, 125, 100, 119, 108, 114, 113, 49, 1, 0] # fmt: skip
|
| 123 |
+
batch = tokenizer(src_text, padding=True, return_tensors=FRAMEWORK)
|
| 124 |
+
self.assertIsInstance(batch, BatchEncoding)
|
| 125 |
+
|
| 126 |
+
if FRAMEWORK != "jax":
|
| 127 |
+
result = list(batch.input_ids.numpy()[0])
|
| 128 |
+
else:
|
| 129 |
+
result = list(batch.input_ids.tolist()[0])
|
| 130 |
+
|
| 131 |
+
self.assertListEqual(expected_src_tokens, result)
|
| 132 |
+
|
| 133 |
+
self.assertEqual((2, 37), batch.input_ids.shape)
|
| 134 |
+
self.assertEqual((2, 37), batch.attention_mask.shape)
|
| 135 |
+
|
| 136 |
+
def test_empty_target_text(self):
|
| 137 |
+
tokenizer = self.t5_base_tokenizer
|
| 138 |
+
src_text = ["A long paragraph for summarization.", "Another paragraph for summarization."]
|
| 139 |
+
batch = tokenizer(src_text, padding=True, return_tensors=FRAMEWORK)
|
| 140 |
+
# check if input_ids are returned and no decoder_input_ids
|
| 141 |
+
self.assertIn("input_ids", batch)
|
| 142 |
+
self.assertIn("attention_mask", batch)
|
| 143 |
+
self.assertNotIn("decoder_input_ids", batch)
|
| 144 |
+
self.assertNotIn("decoder_attention_mask", batch)
|
| 145 |
+
|
| 146 |
+
def test_max_length_integration(self):
|
| 147 |
+
tokenizer = self.t5_base_tokenizer
|
| 148 |
+
tgt_text = [
|
| 149 |
+
"Summary of the text.",
|
| 150 |
+
"Another summary.",
|
| 151 |
+
]
|
| 152 |
+
targets = tokenizer(
|
| 153 |
+
text_target=tgt_text, max_length=32, padding="max_length", truncation=True, return_tensors=FRAMEWORK
|
| 154 |
+
)
|
| 155 |
+
self.assertEqual(32, targets["input_ids"].shape[1])
|
| 156 |
+
|
| 157 |
+
def test_eos_in_input(self):
|
| 158 |
+
tokenizer = self.t5_base_tokenizer
|
| 159 |
+
src_text = ["A long paragraph for summarization. </s>"]
|
| 160 |
+
tgt_text = ["Summary of the text. </s>"]
|
| 161 |
+
expected_src_tokens = [68, 35, 111, 114, 113, 106, 35, 115, 100, 117, 100, 106, 117, 100, 115, 107, 35, 105, 114, 117, 35, 118, 120, 112, 112, 100, 117, 108, 125, 100, 119, 108, 114, 113, 49, 35, 1] # fmt: skip
|
| 162 |
+
expected_tgt_tokens = [86, 120, 112, 112, 100, 117, 124, 35, 114, 105, 35, 119, 107, 104, 35, 119, 104, 123, 119, 49, 35, 1] # fmt: skip
|
| 163 |
+
|
| 164 |
+
batch = tokenizer(src_text, text_target=tgt_text)
|
| 165 |
+
|
| 166 |
+
self.assertEqual(expected_src_tokens, batch["input_ids"][0])
|
| 167 |
+
self.assertEqual(expected_tgt_tokens, batch["labels"][0])
|
| 168 |
+
|
| 169 |
+
# cannot use default save_and_load_tokenizer test method because tokenizer has no vocab
|
| 170 |
+
def test_save_and_load_tokenizer(self):
|
| 171 |
+
# safety check on max_len default value so we are sure the test works
|
| 172 |
+
tokenizers = self.get_tokenizers()
|
| 173 |
+
for tokenizer in tokenizers:
|
| 174 |
+
with self.subTest(f"{tokenizer.__class__.__name__}"):
|
| 175 |
+
self.assertNotEqual(tokenizer.model_max_length, 42)
|
| 176 |
+
|
| 177 |
+
# Now let's start the test
|
| 178 |
+
tokenizers = self.get_tokenizers()
|
| 179 |
+
for tokenizer in tokenizers:
|
| 180 |
+
with self.subTest(f"{tokenizer.__class__.__name__}"):
|
| 181 |
+
# Isolate this from the other tests because we save additional tokens/etc
|
| 182 |
+
tmpdirname = tempfile.mkdtemp()
|
| 183 |
+
|
| 184 |
+
sample_text = " He is very happy, UNwant\u00e9d,running"
|
| 185 |
+
before_tokens = tokenizer.encode(sample_text, add_special_tokens=False)
|
| 186 |
+
tokenizer.save_pretrained(tmpdirname)
|
| 187 |
+
|
| 188 |
+
after_tokenizer = tokenizer.__class__.from_pretrained(tmpdirname)
|
| 189 |
+
after_tokens = after_tokenizer.encode(sample_text, add_special_tokens=False)
|
| 190 |
+
self.assertListEqual(before_tokens, after_tokens)
|
| 191 |
+
|
| 192 |
+
shutil.rmtree(tmpdirname)
|
| 193 |
+
|
| 194 |
+
tokenizers = self.get_tokenizers(model_max_length=42)
|
| 195 |
+
for tokenizer in tokenizers:
|
| 196 |
+
with self.subTest(f"{tokenizer.__class__.__name__}"):
|
| 197 |
+
# Isolate this from the other tests because we save additional tokens/etc
|
| 198 |
+
tmpdirname = tempfile.mkdtemp()
|
| 199 |
+
|
| 200 |
+
sample_text = " He is very happy, UNwant\u00e9d,running"
|
| 201 |
+
tokenizer.add_tokens(["bim", "bambam"])
|
| 202 |
+
additional_special_tokens = tokenizer.additional_special_tokens
|
| 203 |
+
additional_special_tokens.append("new_additional_special_token")
|
| 204 |
+
tokenizer.add_special_tokens(
|
| 205 |
+
{"additional_special_tokens": additional_special_tokens}, replace_additional_special_tokens=False
|
| 206 |
+
)
|
| 207 |
+
before_tokens = tokenizer.encode(sample_text, add_special_tokens=False)
|
| 208 |
+
tokenizer.save_pretrained(tmpdirname)
|
| 209 |
+
|
| 210 |
+
after_tokenizer = tokenizer.__class__.from_pretrained(tmpdirname)
|
| 211 |
+
after_tokens = after_tokenizer.encode(sample_text, add_special_tokens=False)
|
| 212 |
+
self.assertListEqual(before_tokens, after_tokens)
|
| 213 |
+
self.assertIn("new_additional_special_token", after_tokenizer.additional_special_tokens)
|
| 214 |
+
self.assertEqual(after_tokenizer.model_max_length, 42)
|
| 215 |
+
|
| 216 |
+
tokenizer = tokenizer.__class__.from_pretrained(tmpdirname, model_max_length=43)
|
| 217 |
+
self.assertEqual(tokenizer.model_max_length, 43)
|
| 218 |
+
|
| 219 |
+
shutil.rmtree(tmpdirname)
|
| 220 |
+
|
| 221 |
+
# There is a conflict between the default value of extra_ids and adding a new special token through additional_special_tokens
|
| 222 |
+
# We need to add the extra_ids in the list of the arg additional_special_tokens
|
| 223 |
+
def test_special_tokens_initialization_with_non_empty_additional_special_tokens(self):
|
| 224 |
+
tokenizer_list = []
|
| 225 |
+
if self.test_slow_tokenizer:
|
| 226 |
+
tokenizer_list.append((self.tokenizer_class, self.get_tokenizer()))
|
| 227 |
+
|
| 228 |
+
if self.test_rust_tokenizer:
|
| 229 |
+
tokenizer_list.append((self.rust_tokenizer_class, self.get_rust_tokenizer()))
|
| 230 |
+
|
| 231 |
+
for tokenizer_class, tokenizer_utils in tokenizer_list:
|
| 232 |
+
with tempfile.TemporaryDirectory() as tmp_dir:
|
| 233 |
+
tokenizer_utils.save_pretrained(tmp_dir)
|
| 234 |
+
|
| 235 |
+
with open(os.path.join(tmp_dir, "special_tokens_map.json"), encoding="utf-8") as json_file:
|
| 236 |
+
special_tokens_map = json.load(json_file)
|
| 237 |
+
|
| 238 |
+
with open(os.path.join(tmp_dir, "tokenizer_config.json"), encoding="utf-8") as json_file:
|
| 239 |
+
tokenizer_config = json.load(json_file)
|
| 240 |
+
|
| 241 |
+
added_tokens_extra_ids = [f"<extra_id_{i}>" for i in range(125)]
|
| 242 |
+
|
| 243 |
+
special_tokens_map["additional_special_tokens"] = added_tokens_extra_ids + [
|
| 244 |
+
"an_additional_special_token"
|
| 245 |
+
]
|
| 246 |
+
tokenizer_config["additional_special_tokens"] = added_tokens_extra_ids + [
|
| 247 |
+
"an_additional_special_token"
|
| 248 |
+
]
|
| 249 |
+
|
| 250 |
+
with open(os.path.join(tmp_dir, "special_tokens_map.json"), "w", encoding="utf-8") as outfile:
|
| 251 |
+
json.dump(special_tokens_map, outfile)
|
| 252 |
+
with open(os.path.join(tmp_dir, "tokenizer_config.json"), "w", encoding="utf-8") as outfile:
|
| 253 |
+
json.dump(tokenizer_config, outfile)
|
| 254 |
+
|
| 255 |
+
# the following checks allow us to verify that our test works as expected, i.e. that the tokenizer takes
|
| 256 |
+
# into account the new value of additional_special_tokens given in the "tokenizer_config.json" and
|
| 257 |
+
# "special_tokens_map.json" files
|
| 258 |
+
tokenizer_without_change_in_init = tokenizer_class.from_pretrained(
|
| 259 |
+
tmp_dir,
|
| 260 |
+
)
|
| 261 |
+
self.assertIn(
|
| 262 |
+
"an_additional_special_token", tokenizer_without_change_in_init.additional_special_tokens
|
| 263 |
+
)
|
| 264 |
+
# self.assertIn("an_additional_special_token",tokenizer_without_change_in_init.get_vocab()) # ByT5Tokenization no vocab
|
| 265 |
+
self.assertEqual(
|
| 266 |
+
["an_additional_special_token"],
|
| 267 |
+
tokenizer_without_change_in_init.convert_ids_to_tokens(
|
| 268 |
+
tokenizer_without_change_in_init.convert_tokens_to_ids(["an_additional_special_token"])
|
| 269 |
+
),
|
| 270 |
+
)
|
| 271 |
+
|
| 272 |
+
# Now we test that we can change the value of additional_special_tokens in the from_pretrained
|
| 273 |
+
new_added_tokens = added_tokens_extra_ids + [AddedToken("a_new_additional_special_token", lstrip=True)]
|
| 274 |
+
tokenizer = tokenizer_class.from_pretrained(
|
| 275 |
+
tmp_dir,
|
| 276 |
+
additional_special_tokens=new_added_tokens,
|
| 277 |
+
)
|
| 278 |
+
|
| 279 |
+
self.assertIn("a_new_additional_special_token", tokenizer.additional_special_tokens)
|
| 280 |
+
self.assertEqual(
|
| 281 |
+
["a_new_additional_special_token"],
|
| 282 |
+
tokenizer.convert_ids_to_tokens(
|
| 283 |
+
tokenizer.convert_tokens_to_ids(["a_new_additional_special_token"])
|
| 284 |
+
),
|
| 285 |
+
)
|
| 286 |
+
|
| 287 |
+
def test_decode_single_bytes(self):
|
| 288 |
+
tokenizer_list = []
|
| 289 |
+
if self.test_slow_tokenizer:
|
| 290 |
+
tokenizer_list.append((self.tokenizer_class, self.get_tokenizer()))
|
| 291 |
+
|
| 292 |
+
if self.test_rust_tokenizer:
|
| 293 |
+
tokenizer_list.append((self.rust_tokenizer_class, self.get_rust_tokenizer()))
|
| 294 |
+
|
| 295 |
+
for tokenizer_class, tokenizer_utils in tokenizer_list:
|
| 296 |
+
with tempfile.TemporaryDirectory() as tmp_dir:
|
| 297 |
+
tokenizer_utils.save_pretrained(tmp_dir)
|
| 298 |
+
|
| 299 |
+
tokenizer = tokenizer_class.from_pretrained(tmp_dir)
|
| 300 |
+
|
| 301 |
+
self.assertTrue(tokenizer.decode([255]) == "")
|
| 302 |
+
|
| 303 |
+
@unittest.skip(reason="ByT5Tokenizer does not have a vocabulary")
|
| 304 |
+
def test_get_vocab(self):
|
| 305 |
+
pass
|
| 306 |
+
|
| 307 |
+
@unittest.skip(reason="inputs cannot be pretokenized as ids depend on whole input string")
|
| 308 |
+
def test_pretokenized_inputs(self):
|
| 309 |
+
pass
|
| 310 |
+
|
| 311 |
+
@unittest.skip(reason="ByT5Tokenizer does not have a vocabulary")
|
| 312 |
+
def test_conversion_reversible(self):
|
| 313 |
+
pass
|
| 314 |
+
|
| 315 |
+
def test_convert_tokens_to_string_format(self):
|
| 316 |
+
# The default common tokenizer tests uses invalid tokens for ByT5 that can only accept one-character strings
|
| 317 |
+
# and special added tokens as tokens
|
| 318 |
+
tokenizers = self.get_tokenizers(fast=True, do_lower_case=True)
|
| 319 |
+
for tokenizer in tokenizers:
|
| 320 |
+
with self.subTest(f"{tokenizer.__class__.__name__}"):
|
| 321 |
+
tokens = ["t", "h", "i", "s", " ", "i", "s", " ", "a", " ", "t", "e", "x", "t", "</s>"]
|
| 322 |
+
string = tokenizer.convert_tokens_to_string(tokens)
|
| 323 |
+
|
| 324 |
+
self.assertIsInstance(string, str)
|
| 325 |
+
|
| 326 |
+
# We need a different implementation of the test of the same name defined in TokenizerTesterMixin because this tokenizer
|
| 327 |
+
# doesn't have a vocab
|
| 328 |
+
def test_tokenizers_common_ids_setters(self):
|
| 329 |
+
tokenizers = self.get_tokenizers()
|
| 330 |
+
for tokenizer in tokenizers:
|
| 331 |
+
with self.subTest(f"{tokenizer.__class__.__name__}"):
|
| 332 |
+
attributes_list = [
|
| 333 |
+
"bos_token",
|
| 334 |
+
"eos_token",
|
| 335 |
+
"unk_token",
|
| 336 |
+
"sep_token",
|
| 337 |
+
"pad_token",
|
| 338 |
+
"cls_token",
|
| 339 |
+
"mask_token",
|
| 340 |
+
]
|
| 341 |
+
|
| 342 |
+
token_id_to_test_setters = 0
|
| 343 |
+
token_to_test_setters = tokenizer.convert_ids_to_tokens(
|
| 344 |
+
token_id_to_test_setters, skip_special_tokens=False
|
| 345 |
+
)
|
| 346 |
+
|
| 347 |
+
for attr in attributes_list:
|
| 348 |
+
setattr(tokenizer, attr + "_id", None)
|
| 349 |
+
self.assertEqual(getattr(tokenizer, attr), None)
|
| 350 |
+
self.assertEqual(getattr(tokenizer, attr + "_id"), None)
|
| 351 |
+
|
| 352 |
+
setattr(tokenizer, attr + "_id", token_id_to_test_setters)
|
| 353 |
+
self.assertEqual(getattr(tokenizer, attr), token_to_test_setters)
|
| 354 |
+
self.assertEqual(getattr(tokenizer, attr + "_id"), token_id_to_test_setters)
|
| 355 |
+
|
| 356 |
+
setattr(tokenizer, "additional_special_tokens_ids", [])
|
| 357 |
+
self.assertListEqual(getattr(tokenizer, "additional_special_tokens"), [])
|
| 358 |
+
self.assertListEqual(getattr(tokenizer, "additional_special_tokens_ids"), [])
|
| 359 |
+
|
| 360 |
+
setattr(tokenizer, "additional_special_tokens_ids", [token_id_to_test_setters])
|
| 361 |
+
self.assertListEqual(getattr(tokenizer, "additional_special_tokens"), [token_to_test_setters])
|
| 362 |
+
self.assertListEqual(getattr(tokenizer, "additional_special_tokens_ids"), [token_id_to_test_setters])
|
tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/camembert/__init__.py
ADDED
|
File without changes
|
tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/camembert/test_modeling_camembert.py
ADDED
|
@@ -0,0 +1,84 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# coding=utf-8
|
| 2 |
+
# Copyright 2020 The HuggingFace Team. All rights reserved.
|
| 3 |
+
#
|
| 4 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
| 5 |
+
# you may not use this file except in compliance with the License.
|
| 6 |
+
# You may obtain a copy of the License at
|
| 7 |
+
#
|
| 8 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
| 9 |
+
#
|
| 10 |
+
# Unless required by applicable law or agreed to in writing, software
|
| 11 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
| 12 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
| 13 |
+
# See the License for the specific language governing permissions and
|
| 14 |
+
# limitations under the License.
|
| 15 |
+
|
| 16 |
+
import unittest
|
| 17 |
+
|
| 18 |
+
from transformers import is_torch_available
|
| 19 |
+
from transformers.testing_utils import (
|
| 20 |
+
require_sentencepiece,
|
| 21 |
+
require_tokenizers,
|
| 22 |
+
require_torch,
|
| 23 |
+
require_torch_sdpa,
|
| 24 |
+
slow,
|
| 25 |
+
torch_device,
|
| 26 |
+
)
|
| 27 |
+
|
| 28 |
+
|
| 29 |
+
if is_torch_available():
|
| 30 |
+
import torch
|
| 31 |
+
|
| 32 |
+
from transformers import CamembertModel
|
| 33 |
+
|
| 34 |
+
|
| 35 |
+
@require_torch
|
| 36 |
+
@require_sentencepiece
|
| 37 |
+
@require_tokenizers
|
| 38 |
+
class CamembertModelIntegrationTest(unittest.TestCase):
|
| 39 |
+
@slow
|
| 40 |
+
def test_output_embeds_base_model(self):
|
| 41 |
+
model = CamembertModel.from_pretrained("almanach/camembert-base", attn_implementation="eager")
|
| 42 |
+
model.to(torch_device)
|
| 43 |
+
|
| 44 |
+
input_ids = torch.tensor(
|
| 45 |
+
[[5, 121, 11, 660, 16, 730, 25543, 110, 83, 6]],
|
| 46 |
+
device=torch_device,
|
| 47 |
+
dtype=torch.long,
|
| 48 |
+
) # J'aime le camembert !
|
| 49 |
+
with torch.no_grad():
|
| 50 |
+
output = model(input_ids)["last_hidden_state"]
|
| 51 |
+
expected_shape = torch.Size((1, 10, 768))
|
| 52 |
+
self.assertEqual(output.shape, expected_shape)
|
| 53 |
+
# compare the actual values for a slice.
|
| 54 |
+
expected_slice = torch.tensor(
|
| 55 |
+
[[[-0.0254, 0.0235, 0.1027], [0.0606, -0.1811, -0.0418], [-0.1561, -0.1127, 0.2687]]],
|
| 56 |
+
device=torch_device,
|
| 57 |
+
dtype=torch.float,
|
| 58 |
+
)
|
| 59 |
+
# camembert = torch.hub.load('pytorch/fairseq', 'camembert.v0')
|
| 60 |
+
# camembert.eval()
|
| 61 |
+
# expected_slice = roberta.model.forward(input_ids)[0][:, :3, :3].detach()
|
| 62 |
+
|
| 63 |
+
self.assertTrue(torch.allclose(output[:, :3, :3], expected_slice, atol=1e-4))
|
| 64 |
+
|
| 65 |
+
@slow
|
| 66 |
+
@require_torch_sdpa
|
| 67 |
+
def test_output_embeds_base_model_sdpa(self):
|
| 68 |
+
input_ids = torch.tensor(
|
| 69 |
+
[[5, 121, 11, 660, 16, 730, 25543, 110, 83, 6]],
|
| 70 |
+
device=torch_device,
|
| 71 |
+
dtype=torch.long,
|
| 72 |
+
) # J'aime le camembert !
|
| 73 |
+
|
| 74 |
+
expected_slice = torch.tensor(
|
| 75 |
+
[[[-0.0254, 0.0235, 0.1027], [0.0606, -0.1811, -0.0418], [-0.1561, -0.1127, 0.2687]]],
|
| 76 |
+
device=torch_device,
|
| 77 |
+
dtype=torch.float,
|
| 78 |
+
)
|
| 79 |
+
|
| 80 |
+
model = CamembertModel.from_pretrained("almanach/camembert-base", attn_implementation="sdpa").to(torch_device)
|
| 81 |
+
with torch.no_grad():
|
| 82 |
+
output = model(input_ids)["last_hidden_state"].detach()
|
| 83 |
+
|
| 84 |
+
self.assertTrue(torch.allclose(output[:, :3, :3], expected_slice, atol=1e-4))
|
tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/camembert/test_modeling_tf_camembert.py
ADDED
|
@@ -0,0 +1,56 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# coding=utf-8
|
| 2 |
+
# Copyright 2020 The HuggingFace Team. All rights reserved.
|
| 3 |
+
#
|
| 4 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
| 5 |
+
# you may not use this file except in compliance with the License.
|
| 6 |
+
# You may obtain a copy of the License at
|
| 7 |
+
#
|
| 8 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
| 9 |
+
#
|
| 10 |
+
# Unless required by applicable law or agreed to in writing, software
|
| 11 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
| 12 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
| 13 |
+
# See the License for the specific language governing permissions and
|
| 14 |
+
# limitations under the License.
|
| 15 |
+
|
| 16 |
+
from __future__ import annotations
|
| 17 |
+
|
| 18 |
+
import unittest
|
| 19 |
+
|
| 20 |
+
from transformers import is_tf_available
|
| 21 |
+
from transformers.testing_utils import require_sentencepiece, require_tf, require_tokenizers, slow
|
| 22 |
+
|
| 23 |
+
|
| 24 |
+
if is_tf_available():
|
| 25 |
+
import numpy as np
|
| 26 |
+
import tensorflow as tf
|
| 27 |
+
|
| 28 |
+
from transformers import TFCamembertModel
|
| 29 |
+
|
| 30 |
+
|
| 31 |
+
@require_tf
|
| 32 |
+
@require_sentencepiece
|
| 33 |
+
@require_tokenizers
|
| 34 |
+
class TFCamembertModelIntegrationTest(unittest.TestCase):
|
| 35 |
+
@slow
|
| 36 |
+
def test_output_embeds_base_model(self):
|
| 37 |
+
model = TFCamembertModel.from_pretrained("jplu/tf-camembert-base")
|
| 38 |
+
|
| 39 |
+
input_ids = tf.convert_to_tensor(
|
| 40 |
+
[[5, 121, 11, 660, 16, 730, 25543, 110, 83, 6]],
|
| 41 |
+
dtype=tf.int32,
|
| 42 |
+
) # J'aime le camembert !"
|
| 43 |
+
|
| 44 |
+
output = model(input_ids)["last_hidden_state"]
|
| 45 |
+
expected_shape = tf.TensorShape((1, 10, 768))
|
| 46 |
+
self.assertEqual(output.shape, expected_shape)
|
| 47 |
+
# compare the actual values for a slice.
|
| 48 |
+
expected_slice = tf.convert_to_tensor(
|
| 49 |
+
[[[-0.0254, 0.0235, 0.1027], [0.0606, -0.1811, -0.0418], [-0.1561, -0.1127, 0.2687]]],
|
| 50 |
+
dtype=tf.float32,
|
| 51 |
+
)
|
| 52 |
+
# camembert = torch.hub.load('pytorch/fairseq', 'camembert.v0')
|
| 53 |
+
# camembert.eval()
|
| 54 |
+
# expected_slice = roberta.model.forward(input_ids)[0][:, :3, :3].detach()
|
| 55 |
+
|
| 56 |
+
self.assertTrue(np.allclose(output[:, :3, :3].numpy(), expected_slice.numpy(), atol=1e-4))
|
tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/camembert/test_tokenization_camembert.py
ADDED
|
@@ -0,0 +1,220 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# coding=utf-8
|
| 2 |
+
# Copyright 2018 HuggingFace Inc. team.
|
| 3 |
+
#
|
| 4 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
| 5 |
+
# you may not use this file except in compliance with the License.
|
| 6 |
+
# You may obtain a copy of the License at
|
| 7 |
+
#
|
| 8 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
| 9 |
+
#
|
| 10 |
+
# Unless required by applicable law or agreed to in writing, software
|
| 11 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
| 12 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
| 13 |
+
# See the License for the specific language governing permissions and
|
| 14 |
+
# limitations under the License.
|
| 15 |
+
|
| 16 |
+
import tempfile
|
| 17 |
+
import unittest
|
| 18 |
+
|
| 19 |
+
from transformers import AddedToken, CamembertTokenizer, CamembertTokenizerFast
|
| 20 |
+
from transformers.testing_utils import get_tests_dir, require_sentencepiece, require_tokenizers, slow
|
| 21 |
+
from transformers.utils import is_torch_available
|
| 22 |
+
|
| 23 |
+
from ...test_tokenization_common import TokenizerTesterMixin
|
| 24 |
+
|
| 25 |
+
|
| 26 |
+
SAMPLE_VOCAB = get_tests_dir("fixtures/test_sentencepiece.model")
|
| 27 |
+
SAMPLE_BPE_VOCAB = get_tests_dir("fixtures/test_sentencepiece_bpe.model")
|
| 28 |
+
|
| 29 |
+
FRAMEWORK = "pt" if is_torch_available() else "tf"
|
| 30 |
+
|
| 31 |
+
|
| 32 |
+
@require_sentencepiece
|
| 33 |
+
@require_tokenizers
|
| 34 |
+
class CamembertTokenizationTest(TokenizerTesterMixin, unittest.TestCase):
|
| 35 |
+
from_pretrained_id = "almanach/camembert-base"
|
| 36 |
+
tokenizer_class = CamembertTokenizer
|
| 37 |
+
rust_tokenizer_class = CamembertTokenizerFast
|
| 38 |
+
test_rust_tokenizer = True
|
| 39 |
+
test_sentencepiece = True
|
| 40 |
+
|
| 41 |
+
def setUp(self):
|
| 42 |
+
super().setUp()
|
| 43 |
+
|
| 44 |
+
# We have a SentencePiece fixture for testing
|
| 45 |
+
tokenizer = CamembertTokenizer(SAMPLE_VOCAB)
|
| 46 |
+
tokenizer.save_pretrained(self.tmpdirname)
|
| 47 |
+
|
| 48 |
+
@unittest.skip(
|
| 49 |
+
"Token maps are not equal because someone set the probability of ('<unk>NOTUSED', -100), so it's never encoded for fast"
|
| 50 |
+
)
|
| 51 |
+
def test_special_tokens_map_equal(self):
|
| 52 |
+
return
|
| 53 |
+
|
| 54 |
+
def test_convert_token_and_id(self):
|
| 55 |
+
"""Test ``_convert_token_to_id`` and ``_convert_id_to_token``."""
|
| 56 |
+
token = "<pad>"
|
| 57 |
+
token_id = 1 # 1 is the offset id, but in the spm vocab it's 3
|
| 58 |
+
|
| 59 |
+
self.assertEqual(self.get_tokenizer().convert_tokens_to_ids(token), token_id)
|
| 60 |
+
self.assertEqual(self.get_tokenizer().convert_ids_to_tokens(token_id), token)
|
| 61 |
+
|
| 62 |
+
def test_get_vocab(self):
|
| 63 |
+
vocab_keys = list(self.get_tokenizer().get_vocab().keys())
|
| 64 |
+
|
| 65 |
+
self.assertEqual(vocab_keys[0], "<s>NOTUSED")
|
| 66 |
+
self.assertEqual(vocab_keys[1], "<pad>")
|
| 67 |
+
self.assertEqual(vocab_keys[-1], "<mask>")
|
| 68 |
+
self.assertEqual(len(vocab_keys), 1_005)
|
| 69 |
+
|
| 70 |
+
def test_vocab_size(self):
|
| 71 |
+
self.assertEqual(self.get_tokenizer().vocab_size, 1_000)
|
| 72 |
+
|
| 73 |
+
def test_rust_and_python_bpe_tokenizers(self):
|
| 74 |
+
tokenizer = CamembertTokenizer(SAMPLE_BPE_VOCAB)
|
| 75 |
+
tokenizer.save_pretrained(self.tmpdirname)
|
| 76 |
+
rust_tokenizer = CamembertTokenizerFast.from_pretrained(self.tmpdirname)
|
| 77 |
+
|
| 78 |
+
sequence = "I was born in 92000, and this is falsé."
|
| 79 |
+
|
| 80 |
+
ids = tokenizer.encode(sequence)
|
| 81 |
+
rust_ids = rust_tokenizer.encode(sequence)
|
| 82 |
+
self.assertListEqual(ids, rust_ids)
|
| 83 |
+
|
| 84 |
+
ids = tokenizer.encode(sequence, add_special_tokens=False)
|
| 85 |
+
rust_ids = rust_tokenizer.encode(sequence, add_special_tokens=False)
|
| 86 |
+
self.assertListEqual(ids, rust_ids)
|
| 87 |
+
|
| 88 |
+
# <unk> tokens are not the same for `rust` than for `slow`.
|
| 89 |
+
# Because spm gives back raw token instead of `unk` in EncodeAsPieces
|
| 90 |
+
# tokens = tokenizer.tokenize(sequence)
|
| 91 |
+
tokens = tokenizer.convert_ids_to_tokens(ids)
|
| 92 |
+
rust_tokens = rust_tokenizer.tokenize(sequence)
|
| 93 |
+
self.assertListEqual(tokens, rust_tokens)
|
| 94 |
+
|
| 95 |
+
def test_rust_and_python_full_tokenizers(self):
|
| 96 |
+
if not self.test_rust_tokenizer:
|
| 97 |
+
self.skipTest(reason="test_rust_tokenizer is set to False")
|
| 98 |
+
|
| 99 |
+
tokenizer = self.get_tokenizer()
|
| 100 |
+
rust_tokenizer = self.get_rust_tokenizer()
|
| 101 |
+
|
| 102 |
+
sequence = "I was born in 92000, and this is falsé."
|
| 103 |
+
|
| 104 |
+
tokens = tokenizer.tokenize(sequence)
|
| 105 |
+
rust_tokens = rust_tokenizer.tokenize(sequence)
|
| 106 |
+
self.assertListEqual(tokens, rust_tokens)
|
| 107 |
+
|
| 108 |
+
ids = tokenizer.encode(sequence, add_special_tokens=False)
|
| 109 |
+
rust_ids = rust_tokenizer.encode(sequence, add_special_tokens=False)
|
| 110 |
+
self.assertListEqual(ids, rust_ids)
|
| 111 |
+
|
| 112 |
+
rust_tokenizer = self.get_rust_tokenizer()
|
| 113 |
+
ids = tokenizer.encode(sequence)
|
| 114 |
+
rust_ids = rust_tokenizer.encode(sequence)
|
| 115 |
+
self.assertListEqual(ids, rust_ids)
|
| 116 |
+
|
| 117 |
+
@slow
|
| 118 |
+
def test_tokenizer_integration(self):
|
| 119 |
+
expected_encoding = {'input_ids': [[5, 54, 7196, 297, 30, 23, 776, 18, 11, 3215, 3705, 8252, 22, 3164, 1181, 2116, 29, 16, 813, 25, 791, 3314, 20, 3446, 38, 27575, 120, 6, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [5, 468, 17, 11, 9088, 20, 1517, 8, 22804, 18818, 10, 38, 629, 607, 607, 142, 19, 7196, 867, 56, 10326, 24, 2267, 20, 416, 5072, 15612, 233, 734, 7, 2399, 27, 16, 3015, 1649, 7, 24, 20, 4338, 2399, 27, 13, 3400, 14, 13, 6189, 8, 930, 9, 6]], 'attention_mask': [[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]]} # fmt: skip
|
| 120 |
+
|
| 121 |
+
# camembert is a french model. So we also use french texts.
|
| 122 |
+
sequences = [
|
| 123 |
+
"Le transformeur est un modèle d'apprentissage profond introduit en 2017, "
|
| 124 |
+
"utilisé principalement dans le domaine du traitement automatique des langues (TAL).",
|
| 125 |
+
"À l'instar des réseaux de neurones récurrents (RNN), les transformeurs sont conçus "
|
| 126 |
+
"pour gérer des données séquentielles, telles que le langage naturel, pour des tâches "
|
| 127 |
+
"telles que la traduction et la synthèse de texte.",
|
| 128 |
+
]
|
| 129 |
+
|
| 130 |
+
self.tokenizer_integration_test_util(
|
| 131 |
+
expected_encoding=expected_encoding,
|
| 132 |
+
model_name="almanach/camembert-base",
|
| 133 |
+
revision="3a0641d9a1aeb7e848a74299e7e4c4bca216b4cf",
|
| 134 |
+
sequences=sequences,
|
| 135 |
+
)
|
| 136 |
+
|
| 137 |
+
# Overwritten because we have to use from slow (online pretrained is wrong, the tokenizer.json has a whole)
|
| 138 |
+
def test_added_tokens_serialization(self):
|
| 139 |
+
self.maxDiff = None
|
| 140 |
+
|
| 141 |
+
# Utility to test the added vocab
|
| 142 |
+
def _test_added_vocab_and_eos(expected, tokenizer_class, expected_eos, temp_dir):
|
| 143 |
+
tokenizer = tokenizer_class.from_pretrained(temp_dir)
|
| 144 |
+
self.assertTrue(str(expected_eos) not in tokenizer.additional_special_tokens)
|
| 145 |
+
self.assertIn(new_eos, tokenizer.added_tokens_decoder.values())
|
| 146 |
+
self.assertEqual(tokenizer.added_tokens_decoder[tokenizer.eos_token_id], new_eos)
|
| 147 |
+
self.assertTrue(all(item in tokenizer.added_tokens_decoder.items() for item in expected.items()))
|
| 148 |
+
return tokenizer
|
| 149 |
+
|
| 150 |
+
new_eos = AddedToken("[NEW_EOS]", rstrip=False, lstrip=True, normalized=False)
|
| 151 |
+
for tokenizer, pretrained_name, kwargs in self.tokenizers_list:
|
| 152 |
+
with self.subTest(f"{tokenizer.__class__.__name__} ({pretrained_name})"):
|
| 153 |
+
# Load a slow tokenizer from the hub, init with the new token for fast to also include it
|
| 154 |
+
tokenizer = self.tokenizer_class.from_pretrained(pretrained_name, eos_token=new_eos)
|
| 155 |
+
EXPECTED_ADDED_TOKENS_DECODER = tokenizer.added_tokens_decoder
|
| 156 |
+
with self.subTest("Hub -> Slow: Test loading a slow tokenizer from the hub)"):
|
| 157 |
+
self.assertEqual(tokenizer._special_tokens_map["eos_token"], new_eos)
|
| 158 |
+
self.assertIn(new_eos, list(tokenizer.added_tokens_decoder.values()))
|
| 159 |
+
|
| 160 |
+
with tempfile.TemporaryDirectory() as tmp_dir_2:
|
| 161 |
+
tokenizer.save_pretrained(tmp_dir_2)
|
| 162 |
+
with self.subTest(
|
| 163 |
+
"Hub -> Slow -> Slow: Test saving this slow tokenizer and reloading it in the fast class"
|
| 164 |
+
):
|
| 165 |
+
_test_added_vocab_and_eos(
|
| 166 |
+
EXPECTED_ADDED_TOKENS_DECODER, self.tokenizer_class, new_eos, tmp_dir_2
|
| 167 |
+
)
|
| 168 |
+
|
| 169 |
+
if self.rust_tokenizer_class is not None:
|
| 170 |
+
with self.subTest(
|
| 171 |
+
"Hub -> Slow -> Fast: Test saving this slow tokenizer and reloading it in the fast class"
|
| 172 |
+
):
|
| 173 |
+
tokenizer_fast = _test_added_vocab_and_eos(
|
| 174 |
+
EXPECTED_ADDED_TOKENS_DECODER, self.rust_tokenizer_class, new_eos, tmp_dir_2
|
| 175 |
+
)
|
| 176 |
+
with tempfile.TemporaryDirectory() as tmp_dir_3:
|
| 177 |
+
tokenizer_fast.save_pretrained(tmp_dir_3)
|
| 178 |
+
with self.subTest(
|
| 179 |
+
"Hub -> Slow -> Fast -> Fast: Test saving this fast tokenizer and reloading it in the fast class"
|
| 180 |
+
):
|
| 181 |
+
_test_added_vocab_and_eos(
|
| 182 |
+
EXPECTED_ADDED_TOKENS_DECODER, self.rust_tokenizer_class, new_eos, tmp_dir_3
|
| 183 |
+
)
|
| 184 |
+
|
| 185 |
+
with self.subTest(
|
| 186 |
+
"Hub -> Slow -> Fast -> Slow: Test saving this slow tokenizer and reloading it in the slow class"
|
| 187 |
+
):
|
| 188 |
+
_test_added_vocab_and_eos(
|
| 189 |
+
EXPECTED_ADDED_TOKENS_DECODER, self.rust_tokenizer_class, new_eos, tmp_dir_3
|
| 190 |
+
)
|
| 191 |
+
|
| 192 |
+
with self.subTest("Hub -> Fast: Test loading a fast tokenizer from the hub)"):
|
| 193 |
+
if self.rust_tokenizer_class is not None:
|
| 194 |
+
tokenizer_fast = self.rust_tokenizer_class.from_pretrained(
|
| 195 |
+
pretrained_name, eos_token=new_eos, from_slow=True
|
| 196 |
+
)
|
| 197 |
+
self.assertEqual(tokenizer_fast._special_tokens_map["eos_token"], new_eos)
|
| 198 |
+
self.assertIn(new_eos, list(tokenizer_fast.added_tokens_decoder.values()))
|
| 199 |
+
# We can't test the following because for BC we kept the default rstrip lstrip in slow not fast. Will comment once normalization is alright
|
| 200 |
+
with self.subTest("Hub -> Fast == Hub -> Slow: make sure slow and fast tokenizer match"):
|
| 201 |
+
with self.subTest("Hub -> Fast == Hub -> Slow: make sure slow and fast tokenizer match"):
|
| 202 |
+
self.assertTrue(
|
| 203 |
+
all(
|
| 204 |
+
item in tokenizer.added_tokens_decoder.items()
|
| 205 |
+
for item in EXPECTED_ADDED_TOKENS_DECODER.items()
|
| 206 |
+
)
|
| 207 |
+
)
|
| 208 |
+
|
| 209 |
+
EXPECTED_ADDED_TOKENS_DECODER = tokenizer_fast.added_tokens_decoder
|
| 210 |
+
with tempfile.TemporaryDirectory() as tmp_dir_4:
|
| 211 |
+
tokenizer_fast.save_pretrained(tmp_dir_4)
|
| 212 |
+
with self.subTest("Hub -> Fast -> Fast: saving Fast1 locally and loading"):
|
| 213 |
+
_test_added_vocab_and_eos(
|
| 214 |
+
EXPECTED_ADDED_TOKENS_DECODER, self.rust_tokenizer_class, new_eos, tmp_dir_4
|
| 215 |
+
)
|
| 216 |
+
|
| 217 |
+
with self.subTest("Hub -> Fast -> Slow: saving Fast1 locally and loading"):
|
| 218 |
+
_test_added_vocab_and_eos(
|
| 219 |
+
EXPECTED_ADDED_TOKENS_DECODER, self.tokenizer_class, new_eos, tmp_dir_4
|
| 220 |
+
)
|
tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/canine/__init__.py
ADDED
|
File without changes
|
tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/canine/test_modeling_canine.py
ADDED
|
@@ -0,0 +1,573 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# coding=utf-8
|
| 2 |
+
# Copyright 2021 The HuggingFace Inc. team. All rights reserved.
|
| 3 |
+
#
|
| 4 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
| 5 |
+
# you may not use this file except in compliance with the License.
|
| 6 |
+
# You may obtain a copy of the License at
|
| 7 |
+
#
|
| 8 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
| 9 |
+
#
|
| 10 |
+
# Unless required by applicable law or agreed to in writing, software
|
| 11 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
| 12 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
| 13 |
+
# See the License for the specific language governing permissions and
|
| 14 |
+
# limitations under the License.
|
| 15 |
+
"""Testing suite for the PyTorch CANINE model."""
|
| 16 |
+
|
| 17 |
+
import unittest
|
| 18 |
+
from typing import List, Tuple
|
| 19 |
+
|
| 20 |
+
from transformers import CanineConfig, is_torch_available
|
| 21 |
+
from transformers.testing_utils import require_torch, slow, torch_device
|
| 22 |
+
|
| 23 |
+
from ...test_configuration_common import ConfigTester
|
| 24 |
+
from ...test_modeling_common import ModelTesterMixin, _config_zero_init, global_rng, ids_tensor, random_attention_mask
|
| 25 |
+
from ...test_pipeline_mixin import PipelineTesterMixin
|
| 26 |
+
|
| 27 |
+
|
| 28 |
+
if is_torch_available():
|
| 29 |
+
import torch
|
| 30 |
+
|
| 31 |
+
from transformers import (
|
| 32 |
+
CanineForMultipleChoice,
|
| 33 |
+
CanineForQuestionAnswering,
|
| 34 |
+
CanineForSequenceClassification,
|
| 35 |
+
CanineForTokenClassification,
|
| 36 |
+
CanineModel,
|
| 37 |
+
)
|
| 38 |
+
|
| 39 |
+
|
| 40 |
+
class CanineModelTester:
|
| 41 |
+
def __init__(
|
| 42 |
+
self,
|
| 43 |
+
parent,
|
| 44 |
+
batch_size=13,
|
| 45 |
+
seq_length=7,
|
| 46 |
+
is_training=True,
|
| 47 |
+
use_input_mask=True,
|
| 48 |
+
use_token_type_ids=True,
|
| 49 |
+
use_labels=True,
|
| 50 |
+
# let's use a vocab size that's way bigger than BERT's one
|
| 51 |
+
# NOTE: this is not a model parameter, just an input
|
| 52 |
+
vocab_size=100000,
|
| 53 |
+
hidden_size=32,
|
| 54 |
+
num_hidden_layers=2,
|
| 55 |
+
num_attention_heads=4,
|
| 56 |
+
intermediate_size=37,
|
| 57 |
+
hidden_act="gelu",
|
| 58 |
+
hidden_dropout_prob=0.1,
|
| 59 |
+
attention_probs_dropout_prob=0.1,
|
| 60 |
+
max_position_embeddings=512,
|
| 61 |
+
type_vocab_size=16,
|
| 62 |
+
type_sequence_label_size=2,
|
| 63 |
+
initializer_range=0.02,
|
| 64 |
+
num_labels=3,
|
| 65 |
+
num_choices=4,
|
| 66 |
+
num_hash_buckets=16,
|
| 67 |
+
scope=None,
|
| 68 |
+
):
|
| 69 |
+
self.parent = parent
|
| 70 |
+
self.batch_size = batch_size
|
| 71 |
+
self.seq_length = seq_length
|
| 72 |
+
self.is_training = is_training
|
| 73 |
+
self.use_input_mask = use_input_mask
|
| 74 |
+
self.use_token_type_ids = use_token_type_ids
|
| 75 |
+
self.use_labels = use_labels
|
| 76 |
+
self.vocab_size = vocab_size
|
| 77 |
+
self.hidden_size = hidden_size
|
| 78 |
+
self.num_hidden_layers = num_hidden_layers
|
| 79 |
+
self.num_attention_heads = num_attention_heads
|
| 80 |
+
self.intermediate_size = intermediate_size
|
| 81 |
+
self.hidden_act = hidden_act
|
| 82 |
+
self.hidden_dropout_prob = hidden_dropout_prob
|
| 83 |
+
self.attention_probs_dropout_prob = attention_probs_dropout_prob
|
| 84 |
+
self.max_position_embeddings = max_position_embeddings
|
| 85 |
+
self.type_vocab_size = type_vocab_size
|
| 86 |
+
self.type_sequence_label_size = type_sequence_label_size
|
| 87 |
+
self.initializer_range = initializer_range
|
| 88 |
+
self.num_labels = num_labels
|
| 89 |
+
self.num_choices = num_choices
|
| 90 |
+
self.num_hash_buckets = num_hash_buckets
|
| 91 |
+
self.scope = scope
|
| 92 |
+
|
| 93 |
+
def prepare_config_and_inputs(self):
|
| 94 |
+
input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
|
| 95 |
+
|
| 96 |
+
input_mask = None
|
| 97 |
+
if self.use_input_mask:
|
| 98 |
+
input_mask = random_attention_mask([self.batch_size, self.seq_length])
|
| 99 |
+
|
| 100 |
+
token_type_ids = None
|
| 101 |
+
if self.use_token_type_ids:
|
| 102 |
+
token_type_ids = ids_tensor(input_ids.shape, self.type_vocab_size)
|
| 103 |
+
|
| 104 |
+
sequence_labels = None
|
| 105 |
+
token_labels = None
|
| 106 |
+
choice_labels = None
|
| 107 |
+
if self.use_labels:
|
| 108 |
+
sequence_labels = ids_tensor([self.batch_size], self.type_sequence_label_size)
|
| 109 |
+
token_labels = ids_tensor([self.batch_size, self.seq_length], self.num_labels)
|
| 110 |
+
choice_labels = ids_tensor([self.batch_size], self.num_choices)
|
| 111 |
+
|
| 112 |
+
config = self.get_config()
|
| 113 |
+
|
| 114 |
+
return config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
|
| 115 |
+
|
| 116 |
+
def get_config(self):
|
| 117 |
+
return CanineConfig(
|
| 118 |
+
hidden_size=self.hidden_size,
|
| 119 |
+
num_hidden_layers=self.num_hidden_layers,
|
| 120 |
+
num_attention_heads=self.num_attention_heads,
|
| 121 |
+
intermediate_size=self.intermediate_size,
|
| 122 |
+
hidden_act=self.hidden_act,
|
| 123 |
+
hidden_dropout_prob=self.hidden_dropout_prob,
|
| 124 |
+
attention_probs_dropout_prob=self.attention_probs_dropout_prob,
|
| 125 |
+
max_position_embeddings=self.max_position_embeddings,
|
| 126 |
+
type_vocab_size=self.type_vocab_size,
|
| 127 |
+
is_decoder=False,
|
| 128 |
+
initializer_range=self.initializer_range,
|
| 129 |
+
num_hash_buckets=self.num_hash_buckets,
|
| 130 |
+
)
|
| 131 |
+
|
| 132 |
+
def create_and_check_model(
|
| 133 |
+
self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
|
| 134 |
+
):
|
| 135 |
+
model = CanineModel(config=config)
|
| 136 |
+
model.to(torch_device)
|
| 137 |
+
model.eval()
|
| 138 |
+
result = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids)
|
| 139 |
+
result = model(input_ids, token_type_ids=token_type_ids)
|
| 140 |
+
result = model(input_ids)
|
| 141 |
+
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
|
| 142 |
+
|
| 143 |
+
def create_and_check_for_question_answering(
|
| 144 |
+
self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
|
| 145 |
+
):
|
| 146 |
+
model = CanineForQuestionAnswering(config=config)
|
| 147 |
+
model.to(torch_device)
|
| 148 |
+
model.eval()
|
| 149 |
+
result = model(
|
| 150 |
+
input_ids,
|
| 151 |
+
attention_mask=input_mask,
|
| 152 |
+
token_type_ids=token_type_ids,
|
| 153 |
+
start_positions=sequence_labels,
|
| 154 |
+
end_positions=sequence_labels,
|
| 155 |
+
)
|
| 156 |
+
self.parent.assertEqual(result.start_logits.shape, (self.batch_size, self.seq_length))
|
| 157 |
+
self.parent.assertEqual(result.end_logits.shape, (self.batch_size, self.seq_length))
|
| 158 |
+
|
| 159 |
+
def create_and_check_for_sequence_classification(
|
| 160 |
+
self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
|
| 161 |
+
):
|
| 162 |
+
config.num_labels = self.num_labels
|
| 163 |
+
model = CanineForSequenceClassification(config)
|
| 164 |
+
model.to(torch_device)
|
| 165 |
+
model.eval()
|
| 166 |
+
result = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids, labels=sequence_labels)
|
| 167 |
+
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.num_labels))
|
| 168 |
+
|
| 169 |
+
def create_and_check_for_token_classification(
|
| 170 |
+
self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
|
| 171 |
+
):
|
| 172 |
+
config.num_labels = self.num_labels
|
| 173 |
+
model = CanineForTokenClassification(config=config)
|
| 174 |
+
model.to(torch_device)
|
| 175 |
+
model.eval()
|
| 176 |
+
result = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids, labels=token_labels)
|
| 177 |
+
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.num_labels))
|
| 178 |
+
|
| 179 |
+
def create_and_check_for_multiple_choice(
|
| 180 |
+
self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
|
| 181 |
+
):
|
| 182 |
+
config.num_choices = self.num_choices
|
| 183 |
+
model = CanineForMultipleChoice(config=config)
|
| 184 |
+
model.to(torch_device)
|
| 185 |
+
model.eval()
|
| 186 |
+
multiple_choice_inputs_ids = input_ids.unsqueeze(1).expand(-1, self.num_choices, -1).contiguous()
|
| 187 |
+
multiple_choice_token_type_ids = token_type_ids.unsqueeze(1).expand(-1, self.num_choices, -1).contiguous()
|
| 188 |
+
multiple_choice_input_mask = input_mask.unsqueeze(1).expand(-1, self.num_choices, -1).contiguous()
|
| 189 |
+
result = model(
|
| 190 |
+
multiple_choice_inputs_ids,
|
| 191 |
+
attention_mask=multiple_choice_input_mask,
|
| 192 |
+
token_type_ids=multiple_choice_token_type_ids,
|
| 193 |
+
labels=choice_labels,
|
| 194 |
+
)
|
| 195 |
+
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.num_choices))
|
| 196 |
+
|
| 197 |
+
def prepare_config_and_inputs_for_common(self):
|
| 198 |
+
config_and_inputs = self.prepare_config_and_inputs()
|
| 199 |
+
(
|
| 200 |
+
config,
|
| 201 |
+
input_ids,
|
| 202 |
+
token_type_ids,
|
| 203 |
+
input_mask,
|
| 204 |
+
sequence_labels,
|
| 205 |
+
token_labels,
|
| 206 |
+
choice_labels,
|
| 207 |
+
) = config_and_inputs
|
| 208 |
+
inputs_dict = {"input_ids": input_ids, "token_type_ids": token_type_ids, "attention_mask": input_mask}
|
| 209 |
+
return config, inputs_dict
|
| 210 |
+
|
| 211 |
+
|
| 212 |
+
@require_torch
|
| 213 |
+
class CanineModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase):
|
| 214 |
+
all_model_classes = (
|
| 215 |
+
(
|
| 216 |
+
CanineModel,
|
| 217 |
+
CanineForMultipleChoice,
|
| 218 |
+
CanineForQuestionAnswering,
|
| 219 |
+
CanineForSequenceClassification,
|
| 220 |
+
CanineForTokenClassification,
|
| 221 |
+
)
|
| 222 |
+
if is_torch_available()
|
| 223 |
+
else ()
|
| 224 |
+
)
|
| 225 |
+
pipeline_model_mapping = (
|
| 226 |
+
{
|
| 227 |
+
"feature-extraction": CanineModel,
|
| 228 |
+
"question-answering": CanineForQuestionAnswering,
|
| 229 |
+
"text-classification": CanineForSequenceClassification,
|
| 230 |
+
"token-classification": CanineForTokenClassification,
|
| 231 |
+
"zero-shot": CanineForSequenceClassification,
|
| 232 |
+
}
|
| 233 |
+
if is_torch_available()
|
| 234 |
+
else {}
|
| 235 |
+
)
|
| 236 |
+
|
| 237 |
+
test_mismatched_shapes = False
|
| 238 |
+
test_resize_embeddings = False
|
| 239 |
+
test_pruning = False
|
| 240 |
+
|
| 241 |
+
def setUp(self):
|
| 242 |
+
self.model_tester = CanineModelTester(self)
|
| 243 |
+
# we set has_text_modality to False as the config has no vocab_size attribute
|
| 244 |
+
self.config_tester = ConfigTester(self, config_class=CanineConfig, has_text_modality=False, hidden_size=37)
|
| 245 |
+
|
| 246 |
+
def test_config(self):
|
| 247 |
+
self.config_tester.run_common_tests()
|
| 248 |
+
|
| 249 |
+
def test_model(self):
|
| 250 |
+
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
| 251 |
+
self.model_tester.create_and_check_model(*config_and_inputs)
|
| 252 |
+
|
| 253 |
+
def test_for_multiple_choice(self):
|
| 254 |
+
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
| 255 |
+
self.model_tester.create_and_check_for_multiple_choice(*config_and_inputs)
|
| 256 |
+
|
| 257 |
+
def test_for_question_answering(self):
|
| 258 |
+
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
| 259 |
+
self.model_tester.create_and_check_for_question_answering(*config_and_inputs)
|
| 260 |
+
|
| 261 |
+
def test_for_sequence_classification(self):
|
| 262 |
+
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
| 263 |
+
self.model_tester.create_and_check_for_sequence_classification(*config_and_inputs)
|
| 264 |
+
|
| 265 |
+
def test_for_token_classification(self):
|
| 266 |
+
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
| 267 |
+
self.model_tester.create_and_check_for_token_classification(*config_and_inputs)
|
| 268 |
+
|
| 269 |
+
def test_hidden_states_output(self):
|
| 270 |
+
def check_hidden_states_output(inputs_dict, config, model_class):
|
| 271 |
+
model = model_class(config)
|
| 272 |
+
model.to(torch_device)
|
| 273 |
+
model.eval()
|
| 274 |
+
|
| 275 |
+
with torch.no_grad():
|
| 276 |
+
outputs = model(**self._prepare_for_class(inputs_dict, model_class))
|
| 277 |
+
|
| 278 |
+
hidden_states = outputs.hidden_states
|
| 279 |
+
# expected_num_layers equals num_hidden_layers of the deep encoder + 1, + 2 for the first shallow encoder, + 2
|
| 280 |
+
# for the final shallow encoder
|
| 281 |
+
expected_num_layers = self.model_tester.num_hidden_layers + 1 + 2 + 2
|
| 282 |
+
self.assertEqual(len(hidden_states), expected_num_layers)
|
| 283 |
+
|
| 284 |
+
seq_length = self.model_tester.seq_length
|
| 285 |
+
for i in range(expected_num_layers):
|
| 286 |
+
if (i < 2) or ((expected_num_layers - i) < 3):
|
| 287 |
+
# the expected length of the hidden_states of the first and final shallow encoders
|
| 288 |
+
# is equal to the seq_length
|
| 289 |
+
self.assertListEqual(
|
| 290 |
+
list(hidden_states[i].shape[-2:]),
|
| 291 |
+
[seq_length, self.model_tester.hidden_size],
|
| 292 |
+
)
|
| 293 |
+
else:
|
| 294 |
+
# the expected length of the hidden_states of the deep encoder need to be updated
|
| 295 |
+
# for CANINE since the seq length is downsampled
|
| 296 |
+
self.assertListEqual(
|
| 297 |
+
list(hidden_states[i].shape[-2:]),
|
| 298 |
+
[seq_length // config.downsampling_rate, self.model_tester.hidden_size],
|
| 299 |
+
)
|
| 300 |
+
|
| 301 |
+
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
| 302 |
+
|
| 303 |
+
for model_class in self.all_model_classes:
|
| 304 |
+
inputs_dict["output_hidden_states"] = True
|
| 305 |
+
check_hidden_states_output(inputs_dict, config, model_class)
|
| 306 |
+
|
| 307 |
+
# check that output_hidden_states also work using config
|
| 308 |
+
del inputs_dict["output_hidden_states"]
|
| 309 |
+
config.output_hidden_states = True
|
| 310 |
+
|
| 311 |
+
check_hidden_states_output(inputs_dict, config, model_class)
|
| 312 |
+
|
| 313 |
+
def test_attention_outputs(self):
|
| 314 |
+
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
| 315 |
+
config.return_dict = True
|
| 316 |
+
|
| 317 |
+
seq_len = getattr(self.model_tester, "seq_length", None)
|
| 318 |
+
|
| 319 |
+
for model_class in self.all_model_classes:
|
| 320 |
+
inputs_dict["output_attentions"] = True
|
| 321 |
+
inputs_dict["output_hidden_states"] = False
|
| 322 |
+
config.return_dict = True
|
| 323 |
+
model = model_class(config)
|
| 324 |
+
model.to(torch_device)
|
| 325 |
+
model.eval()
|
| 326 |
+
with torch.no_grad():
|
| 327 |
+
outputs = model(**self._prepare_for_class(inputs_dict, model_class))
|
| 328 |
+
attentions = outputs.attentions
|
| 329 |
+
# we add + 2 due to the 2 shallow encoders
|
| 330 |
+
self.assertEqual(len(attentions), self.model_tester.num_hidden_layers + 2)
|
| 331 |
+
|
| 332 |
+
# check that output_attentions also work using config
|
| 333 |
+
del inputs_dict["output_attentions"]
|
| 334 |
+
config.output_attentions = True
|
| 335 |
+
model = model_class(config)
|
| 336 |
+
model.to(torch_device)
|
| 337 |
+
model.eval()
|
| 338 |
+
with torch.no_grad():
|
| 339 |
+
outputs = model(**self._prepare_for_class(inputs_dict, model_class))
|
| 340 |
+
attentions = outputs.attentions
|
| 341 |
+
# we add + 2 due to the 2 shallow encoders
|
| 342 |
+
self.assertEqual(len(attentions), self.model_tester.num_hidden_layers + 2)
|
| 343 |
+
|
| 344 |
+
self.assertListEqual(
|
| 345 |
+
list(attentions[0].shape[-3:]),
|
| 346 |
+
[self.model_tester.num_attention_heads, seq_len, seq_len],
|
| 347 |
+
)
|
| 348 |
+
out_len = len(outputs)
|
| 349 |
+
|
| 350 |
+
# Check attention is always last and order is fine
|
| 351 |
+
inputs_dict["output_attentions"] = True
|
| 352 |
+
inputs_dict["output_hidden_states"] = True
|
| 353 |
+
model = model_class(config)
|
| 354 |
+
model.to(torch_device)
|
| 355 |
+
model.eval()
|
| 356 |
+
with torch.no_grad():
|
| 357 |
+
outputs = model(**self._prepare_for_class(inputs_dict, model_class))
|
| 358 |
+
|
| 359 |
+
if hasattr(self.model_tester, "num_hidden_states_types"):
|
| 360 |
+
added_hidden_states = self.model_tester.num_hidden_states_types
|
| 361 |
+
else:
|
| 362 |
+
added_hidden_states = 1
|
| 363 |
+
self.assertEqual(out_len + added_hidden_states, len(outputs))
|
| 364 |
+
|
| 365 |
+
self_attentions = outputs.attentions
|
| 366 |
+
|
| 367 |
+
self.assertEqual(len(self_attentions), self.model_tester.num_hidden_layers + 2)
|
| 368 |
+
self.assertListEqual(
|
| 369 |
+
list(self_attentions[0].shape[-3:]),
|
| 370 |
+
[self.model_tester.num_attention_heads, seq_len, seq_len],
|
| 371 |
+
)
|
| 372 |
+
|
| 373 |
+
def test_model_outputs_equivalence(self):
|
| 374 |
+
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
| 375 |
+
|
| 376 |
+
def set_nan_tensor_to_zero(t):
|
| 377 |
+
t[t != t] = 0
|
| 378 |
+
return t
|
| 379 |
+
|
| 380 |
+
def check_equivalence(model, tuple_inputs, dict_inputs, additional_kwargs={}):
|
| 381 |
+
with torch.no_grad():
|
| 382 |
+
tuple_output = model(**tuple_inputs, return_dict=False, **additional_kwargs)
|
| 383 |
+
dict_output = model(**dict_inputs, return_dict=True, **additional_kwargs).to_tuple()
|
| 384 |
+
|
| 385 |
+
def recursive_check(tuple_object, dict_object):
|
| 386 |
+
if isinstance(tuple_object, (List, Tuple)):
|
| 387 |
+
for tuple_iterable_value, dict_iterable_value in zip(tuple_object, dict_object):
|
| 388 |
+
recursive_check(tuple_iterable_value, dict_iterable_value)
|
| 389 |
+
elif tuple_object is None:
|
| 390 |
+
return
|
| 391 |
+
else:
|
| 392 |
+
self.assertTrue(
|
| 393 |
+
torch.allclose(
|
| 394 |
+
set_nan_tensor_to_zero(tuple_object), set_nan_tensor_to_zero(dict_object), atol=1e-5
|
| 395 |
+
),
|
| 396 |
+
msg=(
|
| 397 |
+
"Tuple and dict output are not equal. Difference:"
|
| 398 |
+
f" {torch.max(torch.abs(tuple_object - dict_object))}. Tuple has `nan`:"
|
| 399 |
+
f" {torch.isnan(tuple_object).any()} and `inf`: {torch.isinf(tuple_object)}. Dict has"
|
| 400 |
+
f" `nan`: {torch.isnan(dict_object).any()} and `inf`: {torch.isinf(dict_object)}."
|
| 401 |
+
),
|
| 402 |
+
)
|
| 403 |
+
|
| 404 |
+
recursive_check(tuple_output, dict_output)
|
| 405 |
+
|
| 406 |
+
for model_class in self.all_model_classes:
|
| 407 |
+
print(model_class)
|
| 408 |
+
model = model_class(config)
|
| 409 |
+
model.to(torch_device)
|
| 410 |
+
model.eval()
|
| 411 |
+
|
| 412 |
+
tuple_inputs = self._prepare_for_class(inputs_dict, model_class)
|
| 413 |
+
dict_inputs = self._prepare_for_class(inputs_dict, model_class)
|
| 414 |
+
check_equivalence(model, tuple_inputs, dict_inputs)
|
| 415 |
+
|
| 416 |
+
tuple_inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
|
| 417 |
+
dict_inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
|
| 418 |
+
check_equivalence(model, tuple_inputs, dict_inputs)
|
| 419 |
+
|
| 420 |
+
tuple_inputs = self._prepare_for_class(inputs_dict, model_class)
|
| 421 |
+
dict_inputs = self._prepare_for_class(inputs_dict, model_class)
|
| 422 |
+
check_equivalence(model, tuple_inputs, dict_inputs, {"output_hidden_states": True})
|
| 423 |
+
|
| 424 |
+
tuple_inputs = self._prepare_for_class(inputs_dict, model_class)
|
| 425 |
+
dict_inputs = self._prepare_for_class(inputs_dict, model_class)
|
| 426 |
+
check_equivalence(model, tuple_inputs, dict_inputs, {"output_attentions": True})
|
| 427 |
+
|
| 428 |
+
tuple_inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
|
| 429 |
+
dict_inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
|
| 430 |
+
check_equivalence(model, tuple_inputs, dict_inputs, {"output_hidden_states": True})
|
| 431 |
+
|
| 432 |
+
tuple_inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
|
| 433 |
+
dict_inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
|
| 434 |
+
check_equivalence(model, tuple_inputs, dict_inputs, {"output_attentions": True})
|
| 435 |
+
|
| 436 |
+
tuple_inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
|
| 437 |
+
dict_inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
|
| 438 |
+
check_equivalence(
|
| 439 |
+
model, tuple_inputs, dict_inputs, {"output_hidden_states": True, "output_attentions": True}
|
| 440 |
+
)
|
| 441 |
+
|
| 442 |
+
def test_headmasking(self):
|
| 443 |
+
if not self.test_head_masking:
|
| 444 |
+
self.skipTest(reason="test_head_masking is set to False")
|
| 445 |
+
|
| 446 |
+
global_rng.seed(42)
|
| 447 |
+
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
| 448 |
+
global_rng.seed()
|
| 449 |
+
|
| 450 |
+
inputs_dict["output_attentions"] = True
|
| 451 |
+
config.output_hidden_states = True
|
| 452 |
+
configs_no_init = _config_zero_init(config) # To be sure we have no Nan
|
| 453 |
+
for model_class in self.all_model_classes:
|
| 454 |
+
model = model_class(config=configs_no_init)
|
| 455 |
+
model.to(torch_device)
|
| 456 |
+
model.eval()
|
| 457 |
+
|
| 458 |
+
# Prepare head_mask
|
| 459 |
+
# Set require_grad after having prepared the tensor to avoid error (leaf variable has been moved into the graph interior)
|
| 460 |
+
head_mask = torch.ones(
|
| 461 |
+
self.model_tester.num_hidden_layers,
|
| 462 |
+
self.model_tester.num_attention_heads,
|
| 463 |
+
device=torch_device,
|
| 464 |
+
)
|
| 465 |
+
head_mask[0, 0] = 0
|
| 466 |
+
head_mask[-1, :-1] = 0
|
| 467 |
+
head_mask.requires_grad_(requires_grad=True)
|
| 468 |
+
inputs = self._prepare_for_class(inputs_dict, model_class).copy()
|
| 469 |
+
inputs["head_mask"] = head_mask
|
| 470 |
+
|
| 471 |
+
outputs = model(**inputs, return_dict=True)
|
| 472 |
+
|
| 473 |
+
# Test that we can get a gradient back for importance score computation
|
| 474 |
+
output = sum(t.sum() for t in outputs[0])
|
| 475 |
+
output = output.sum()
|
| 476 |
+
output.backward()
|
| 477 |
+
multihead_outputs = head_mask.grad
|
| 478 |
+
|
| 479 |
+
self.assertIsNotNone(multihead_outputs)
|
| 480 |
+
self.assertEqual(len(multihead_outputs), self.model_tester.num_hidden_layers)
|
| 481 |
+
|
| 482 |
+
def check_attentions_validity(attentions):
|
| 483 |
+
# Remove Nan
|
| 484 |
+
for t in attentions:
|
| 485 |
+
self.assertLess(
|
| 486 |
+
torch.sum(torch.isnan(t)), t.numel() / 4
|
| 487 |
+
) # Check we don't have more than 25% nans (arbitrary)
|
| 488 |
+
attentions = [
|
| 489 |
+
t.masked_fill(torch.isnan(t), 0.0) for t in attentions
|
| 490 |
+
] # remove them (the test is less complete)
|
| 491 |
+
|
| 492 |
+
self.assertAlmostEqual(attentions[1][..., 0, :, :].flatten().sum().item(), 0.0)
|
| 493 |
+
self.assertNotEqual(attentions[1][..., -1, :, :].flatten().sum().item(), 0.0)
|
| 494 |
+
self.assertAlmostEqual(attentions[-2][..., -2, :, :].flatten().sum().item(), 0.0)
|
| 495 |
+
self.assertNotEqual(attentions[-2][..., -1, :, :].flatten().sum().item(), 0.0)
|
| 496 |
+
|
| 497 |
+
check_attentions_validity(outputs.attentions)
|
| 498 |
+
|
| 499 |
+
@unittest.skip(reason="CANINE does not have a get_input_embeddings() method.")
|
| 500 |
+
def test_inputs_embeds(self):
|
| 501 |
+
# ViT does not use inputs_embeds
|
| 502 |
+
pass
|
| 503 |
+
|
| 504 |
+
@unittest.skip(reason="Canine Tower does not use inputs_embeds")
|
| 505 |
+
def test_inputs_embeds_matches_input_ids(self):
|
| 506 |
+
pass
|
| 507 |
+
|
| 508 |
+
@unittest.skip(reason="CANINE does not have a get_input_embeddings() method.")
|
| 509 |
+
def test_model_get_set_embeddings(self):
|
| 510 |
+
pass
|
| 511 |
+
|
| 512 |
+
@unittest.skip(
|
| 513 |
+
reason="This architecure seem to not compute gradients properly when using GC, check: https://github.com/huggingface/transformers/pull/27124"
|
| 514 |
+
)
|
| 515 |
+
def test_training_gradient_checkpointing(self):
|
| 516 |
+
pass
|
| 517 |
+
|
| 518 |
+
@unittest.skip(
|
| 519 |
+
reason="This architecure seem to not compute gradients properly when using GC, check: https://github.com/huggingface/transformers/pull/27124"
|
| 520 |
+
)
|
| 521 |
+
def test_training_gradient_checkpointing_use_reentrant(self):
|
| 522 |
+
pass
|
| 523 |
+
|
| 524 |
+
@unittest.skip(
|
| 525 |
+
reason="This architecure seem to not compute gradients properly when using GC, check: https://github.com/huggingface/transformers/pull/27124"
|
| 526 |
+
)
|
| 527 |
+
def test_training_gradient_checkpointing_use_reentrant_false(self):
|
| 528 |
+
pass
|
| 529 |
+
|
| 530 |
+
@slow
|
| 531 |
+
def test_model_from_pretrained(self):
|
| 532 |
+
model_name = "google/canine-s"
|
| 533 |
+
model = CanineModel.from_pretrained(model_name)
|
| 534 |
+
self.assertIsNotNone(model)
|
| 535 |
+
|
| 536 |
+
|
| 537 |
+
@require_torch
|
| 538 |
+
class CanineModelIntegrationTest(unittest.TestCase):
|
| 539 |
+
@slow
|
| 540 |
+
def test_inference_no_head(self):
|
| 541 |
+
model = CanineModel.from_pretrained("google/canine-s")
|
| 542 |
+
# this one corresponds to the first example of the TydiQA dev set (in Swahili)
|
| 543 |
+
# fmt: off
|
| 544 |
+
input_ids = [57344, 57349, 85, 107, 117, 98, 119, 97, 32, 119, 97, 32, 82, 105, 106, 105, 108, 105, 32, 75, 97, 110, 116, 111, 114, 105, 32, 110, 105, 32, 107, 105, 97, 115, 105, 32, 103, 97, 110, 105, 63, 57345, 57350, 32, 82, 105, 106, 105, 108, 105, 32, 75, 97, 110, 116, 111, 114, 105, 32, 44, 32, 82, 105, 106, 105, 108, 105, 32, 75, 97, 110, 116, 97, 114, 117, 115, 105, 32, 97, 117, 32, 105, 110, 103, 46, 32, 65, 108, 112, 104, 97, 32, 67, 101, 110, 116, 97, 117, 114, 105, 32, 40, 112, 105, 97, 58, 32, 84, 111, 108, 105, 109, 97, 110, 32, 97, 117, 32, 82, 105, 103, 105, 108, 32, 75, 101, 110, 116, 97, 117, 114, 117, 115, 41, 32, 110, 105, 32, 110, 121, 111, 116, 97, 32, 105, 110, 97, 121, 111, 110, 103, 39, 97, 97, 32, 115, 97, 110, 97, 32, 107, 97, 116, 105, 107, 97, 32, 97, 110, 103, 97, 32, 121, 97, 32, 107, 117, 115, 105, 110, 105, 32, 107, 119, 101, 110, 121, 101, 32, 107, 117, 110, 100, 105, 110, 121, 111, 116, 97, 32, 121, 97, 32, 75, 97, 110, 116, 97, 114, 117, 115, 105, 32, 40, 112, 105, 97, 58, 32, 105, 110, 103, 46, 32, 67, 101, 110, 116, 97, 117, 114, 117, 115, 41, 46, 32, 78, 105, 32, 110, 121, 111, 116, 97, 32, 121, 97, 32, 107, 117, 110, 103, 97, 97, 32, 115, 97, 110, 97, 32, 121, 97, 32, 110, 110, 101, 32, 97, 110, 103, 97, 110, 105, 32, 108, 97, 107, 105, 110, 105, 32, 104, 97, 105, 111, 110, 101, 107, 97, 110, 105, 32, 107, 119, 101, 110, 121, 101, 32, 110, 117, 115, 117, 100, 117, 110, 105, 97, 32, 121, 97, 32, 107, 97, 115, 107, 97, 122, 105, 110, 105, 46, 32, 57351, 32, 65, 108, 112, 104, 97, 32, 67, 101, 110, 116, 97, 117, 114, 105, 32, 110, 105, 32, 110, 121, 111, 116, 97, 32, 121, 97, 32, 112, 101, 107, 101, 101, 32, 107, 119, 97, 32, 115, 97, 98, 97, 98, 117, 32, 110, 105, 32, 110, 121, 111, 116, 97, 32, 121, 101, 116, 117, 32, 106, 105, 114, 97, 110, 105, 32, 107, 97, 116, 105, 107, 97, 32, 97, 110, 103, 97, 32, 105, 110, 97, 32, 117, 109, 98, 97, 108, 105, 32, 119, 97, 32, 109, 105, 97, 107, 97, 32, 121, 97, 32, 110, 117, 114, 117, 32, 52, 46, 50, 46, 32, 73, 110, 97, 111, 110, 101, 107, 97, 110, 97, 32, 97, 110, 103, 97, 110, 105, 32, 107, 97, 114, 105, 98, 117, 32, 110, 97, 32, 107, 117, 110, 100, 105, 110, 121, 111, 116, 97, 32, 121, 97, 32, 83, 97, 108, 105, 98, 117, 32, 40, 67, 114, 117, 120, 41, 46, 32, 57352, 32, 82, 105, 106, 105, 108, 105, 32, 75, 97, 110, 116, 97, 114, 117, 115, 105, 32, 40, 65, 108, 112, 104, 97, 32, 67, 101, 110, 116, 97, 117, 114, 105, 41, 32, 105, 110, 97, 111, 110, 101, 107, 97, 110, 97, 32, 107, 97, 109, 97, 32, 110, 121, 111, 116, 97, 32, 109, 111, 106, 97, 32, 108, 97, 107, 105, 110, 105, 32, 107, 119, 97, 32, 100, 97, 114, 117, 98, 105, 110, 105, 32, 107, 117, 98, 119, 97, 32, 105, 110, 97, 111, 110, 101, 107, 97, 110, 97, 32, 107, 117, 119, 97, 32, 109, 102, 117, 109, 111, 32, 119, 97, 32, 110, 121, 111, 116, 97, 32, 116, 97, 116, 117, 32, 122, 105, 110, 97, 122, 111, 107, 97, 97, 32, 107, 97, 114, 105, 98, 117, 32, 110, 97, 32, 107, 117, 115, 104, 105, 107, 97, 109, 97, 110, 97, 32, 107, 97, 116, 105, 32, 121, 97, 111, 46, 32, 78, 121, 111, 116, 97, 32, 109, 97, 112, 97, 99, 104, 97, 32, 122, 97, 32, 65, 108, 112, 104, 97, 32, 67, 101, 110, 116, 97, 117, 114, 105, 32, 65, 32, 110, 97, 32, 65, 108, 112, 104, 97, 32, 67, 101, 110, 116, 97, 117, 114, 105, 32, 66, 32, 122, 105, 107, 111, 32, 109, 105, 97, 107, 97, 32, 121, 97, 32, 110, 117, 114, 117, 32, 52, 46, 51, 54, 32, 107, 117, 116, 111, 107, 97, 32, 107, 119, 101, 116, 117, 32, 110, 97, 32, 110, 121, 111, 116, 97, 32, 121, 97, 32, 116, 97, 116, 117, 32, 65, 108, 112, 104, 97, 32, 67, 101, 110, 116, 97, 117, 114, 105, 32, 67, 32, 97, 117, 32, 80, 114, 111, 120, 105, 109, 97, 32, 67, 101, 110, 116, 97, 117, 114, 105, 32, 105, 110, 97, 32, 117, 109, 98, 97, 108, 105, 32, 119, 97, 32, 109, 105, 97, 107, 97, 32, 121, 97, 32, 110, 117, 114, 117, 32, 52, 46, 50, 50, 46, 32, 57353, 32, 80, 114, 111, 120, 105, 109, 97, 32, 67, 101, 110, 116, 97, 117, 114, 105, 32, 40, 121, 97, 97, 110, 105, 32, 110, 121, 111, 116, 97, 32, 121, 97, 32, 75, 97, 110, 116, 97, 114, 117, 115, 105, 32, 105, 108, 105, 121, 111, 32, 107, 97, 114, 105, 98, 117, 32, 122, 97, 105, 100, 105, 32, 110, 97, 115, 105, 41, 32, 105, 109, 101, 103, 117, 110, 100, 117, 108, 105, 119, 97, 32, 107, 117, 119, 97, 32, 110, 97, 32, 115, 97, 121, 97, 114, 105, 32, 109, 111, 106, 97, 46, 32, 86, 105, 112, 105, 109, 111, 32, 118, 105, 110, 97, 118, 121, 111, 112, 97, 116, 105, 107, 97, 110, 97, 32, 104, 97, 100, 105, 32, 115, 97, 115, 97, 32, 122, 105, 110, 97, 111, 110, 121, 101, 115, 104, 97, 32, 117, 119, 101, 122, 101, 107, 97, 110, 111, 32, 109, 107, 117, 98, 119, 97, 32, 121, 97, 32, 107, 119, 97, 109, 98, 97, 32, 115, 97, 121, 97, 114, 105, 32, 104, 105, 105, 32, 110, 105, 32, 121, 97, 32, 109, 119, 97, 109, 98, 97, 32, 40, 107, 97, 109, 97, 32, 100, 117, 110, 105, 97, 32, 121, 101, 116, 117, 44, 32, 77, 105, 114, 105, 104, 105, 32, 97, 117, 32, 90, 117, 104, 117, 114, 97, 41, 32, 110, 97, 32, 105, 110, 97, 119, 101, 122, 97, 32, 107, 117, 119, 97, 32, 110, 97, 32, 97, 110, 103, 97, 104, 101, 119, 97, 44, 32, 116, 101, 110, 97, 32, 107, 97, 116, 105, 107, 97, 32, 117, 112, 101, 111, 32, 119, 97, 32, 106, 111, 116, 111, 32, 117, 110, 97, 111, 114, 117, 104, 117, 115, 117, 32, 107, 117, 119, 101, 112, 111, 32, 107, 119, 97, 32, 117, 104, 97, 105, 46, 32, 91, 49, 93, 57345, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]
|
| 545 |
+
attention_mask = [1 if x != 0 else 0 for x in input_ids]
|
| 546 |
+
token_type_ids = [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]
|
| 547 |
+
# fmt: on
|
| 548 |
+
input_ids = torch.tensor([input_ids])
|
| 549 |
+
attention_mask = torch.tensor([attention_mask])
|
| 550 |
+
token_type_ids = torch.tensor([token_type_ids])
|
| 551 |
+
outputs = model(input_ids, attention_mask, token_type_ids)
|
| 552 |
+
|
| 553 |
+
# verify sequence output
|
| 554 |
+
expected_shape = torch.Size((1, 2048, 768))
|
| 555 |
+
self.assertEqual(outputs.last_hidden_state.shape, expected_shape)
|
| 556 |
+
|
| 557 |
+
expected_slice = torch.tensor(
|
| 558 |
+
[
|
| 559 |
+
[-0.161433131, 0.395568609, 0.0407391489],
|
| 560 |
+
[-0.108025983, 0.362060368, -0.544592619],
|
| 561 |
+
[-0.141537309, 0.180541009, 0.076907],
|
| 562 |
+
]
|
| 563 |
+
)
|
| 564 |
+
|
| 565 |
+
self.assertTrue(torch.allclose(outputs.last_hidden_state[0, :3, :3], expected_slice, atol=1e-2))
|
| 566 |
+
|
| 567 |
+
# verify pooled output
|
| 568 |
+
expected_shape = torch.Size((1, 768))
|
| 569 |
+
self.assertEqual(outputs.pooler_output.shape, expected_shape)
|
| 570 |
+
|
| 571 |
+
expected_slice = torch.tensor([-0.884311497, -0.529064834, 0.723164916])
|
| 572 |
+
|
| 573 |
+
self.assertTrue(torch.allclose(outputs.pooler_output[0, :3], expected_slice, atol=1e-2))
|
tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/canine/test_tokenization_canine.py
ADDED
|
@@ -0,0 +1,334 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# coding=utf-8
|
| 2 |
+
# Copyright 2021 Google AI and HuggingFace Inc. team.
|
| 3 |
+
#
|
| 4 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
| 5 |
+
# you may not use this file except in compliance with the License.
|
| 6 |
+
# You may obtain a copy of the License at
|
| 7 |
+
#
|
| 8 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
| 9 |
+
#
|
| 10 |
+
# Unless required by applicable law or agreed to in writing, software
|
| 11 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
| 12 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
| 13 |
+
# See the License for the specific language governing permissions and
|
| 14 |
+
# limitations under the License.
|
| 15 |
+
|
| 16 |
+
import json
|
| 17 |
+
import os
|
| 18 |
+
import shutil
|
| 19 |
+
import tempfile
|
| 20 |
+
import unittest
|
| 21 |
+
|
| 22 |
+
from transformers import BatchEncoding, CanineTokenizer
|
| 23 |
+
from transformers.testing_utils import require_tokenizers, require_torch
|
| 24 |
+
from transformers.tokenization_utils import AddedToken
|
| 25 |
+
from transformers.utils import cached_property
|
| 26 |
+
|
| 27 |
+
from ...test_tokenization_common import TokenizerTesterMixin
|
| 28 |
+
|
| 29 |
+
|
| 30 |
+
class CanineTokenizationTest(TokenizerTesterMixin, unittest.TestCase):
|
| 31 |
+
from_pretrained_id = "nielsr/canine-s"
|
| 32 |
+
tokenizer_class = CanineTokenizer
|
| 33 |
+
test_rust_tokenizer = False
|
| 34 |
+
|
| 35 |
+
def setUp(self):
|
| 36 |
+
super().setUp()
|
| 37 |
+
tokenizer = CanineTokenizer()
|
| 38 |
+
tokenizer.save_pretrained(self.tmpdirname)
|
| 39 |
+
|
| 40 |
+
@cached_property
|
| 41 |
+
def canine_tokenizer(self):
|
| 42 |
+
return CanineTokenizer.from_pretrained("google/canine-s")
|
| 43 |
+
|
| 44 |
+
def get_tokenizer(self, **kwargs) -> CanineTokenizer:
|
| 45 |
+
tokenizer = self.tokenizer_class.from_pretrained(self.tmpdirname, **kwargs)
|
| 46 |
+
tokenizer._unicode_vocab_size = 1024
|
| 47 |
+
return tokenizer
|
| 48 |
+
|
| 49 |
+
@require_torch
|
| 50 |
+
def test_prepare_batch_integration(self):
|
| 51 |
+
tokenizer = self.canine_tokenizer
|
| 52 |
+
src_text = ["Life is like a box of chocolates.", "You never know what you're gonna get."]
|
| 53 |
+
expected_src_tokens = [57344, 76, 105, 102, 101, 32, 105, 115, 32, 108, 105, 107, 101, 32, 97, 32, 98, 111, 120, 32, 111, 102, 32, 99, 104, 111, 99, 111, 108, 97, 116, 101, 115, 46, 57345, 0, 0, 0, 0] # fmt: skip
|
| 54 |
+
batch = tokenizer(src_text, padding=True, return_tensors="pt")
|
| 55 |
+
self.assertIsInstance(batch, BatchEncoding)
|
| 56 |
+
|
| 57 |
+
result = list(batch.input_ids.numpy()[0])
|
| 58 |
+
|
| 59 |
+
self.assertListEqual(expected_src_tokens, result)
|
| 60 |
+
|
| 61 |
+
self.assertEqual((2, 39), batch.input_ids.shape)
|
| 62 |
+
self.assertEqual((2, 39), batch.attention_mask.shape)
|
| 63 |
+
|
| 64 |
+
@require_torch
|
| 65 |
+
def test_encoding_keys(self):
|
| 66 |
+
tokenizer = self.canine_tokenizer
|
| 67 |
+
src_text = ["Once there was a man.", "He wrote a test in HuggingFace Tranformers."]
|
| 68 |
+
batch = tokenizer(src_text, padding=True, return_tensors="pt")
|
| 69 |
+
# check if input_ids, attention_mask and token_type_ids are returned
|
| 70 |
+
self.assertIn("input_ids", batch)
|
| 71 |
+
self.assertIn("attention_mask", batch)
|
| 72 |
+
self.assertIn("token_type_ids", batch)
|
| 73 |
+
|
| 74 |
+
@require_torch
|
| 75 |
+
def test_max_length_integration(self):
|
| 76 |
+
tokenizer = self.canine_tokenizer
|
| 77 |
+
tgt_text = [
|
| 78 |
+
"What's the weater?",
|
| 79 |
+
"It's about 25 degrees.",
|
| 80 |
+
]
|
| 81 |
+
targets = tokenizer(
|
| 82 |
+
text_target=tgt_text, max_length=32, padding="max_length", truncation=True, return_tensors="pt"
|
| 83 |
+
)
|
| 84 |
+
self.assertEqual(32, targets["input_ids"].shape[1])
|
| 85 |
+
|
| 86 |
+
# cannot use default save_and_load_tokenizer test method because tokenizer has no vocab
|
| 87 |
+
def test_save_and_load_tokenizer(self):
|
| 88 |
+
# safety check on max_len default value so we are sure the test works
|
| 89 |
+
tokenizers = self.get_tokenizers()
|
| 90 |
+
for tokenizer in tokenizers:
|
| 91 |
+
with self.subTest(f"{tokenizer.__class__.__name__}"):
|
| 92 |
+
self.assertNotEqual(tokenizer.model_max_length, 42)
|
| 93 |
+
|
| 94 |
+
# Now let's start the test
|
| 95 |
+
tokenizers = self.get_tokenizers()
|
| 96 |
+
for tokenizer in tokenizers:
|
| 97 |
+
with self.subTest(f"{tokenizer.__class__.__name__}"):
|
| 98 |
+
# Isolate this from the other tests because we save additional tokens/etc
|
| 99 |
+
tmpdirname = tempfile.mkdtemp()
|
| 100 |
+
|
| 101 |
+
sample_text = " He is very happy, UNwant\u00e9d,running"
|
| 102 |
+
before_tokens = tokenizer.encode(sample_text, add_special_tokens=False)
|
| 103 |
+
tokenizer.save_pretrained(tmpdirname)
|
| 104 |
+
|
| 105 |
+
after_tokenizer = tokenizer.__class__.from_pretrained(tmpdirname)
|
| 106 |
+
after_tokens = after_tokenizer.encode(sample_text, add_special_tokens=False)
|
| 107 |
+
self.assertListEqual(before_tokens, after_tokens)
|
| 108 |
+
|
| 109 |
+
shutil.rmtree(tmpdirname)
|
| 110 |
+
|
| 111 |
+
tokenizers = self.get_tokenizers(model_max_length=42)
|
| 112 |
+
for tokenizer in tokenizers:
|
| 113 |
+
with self.subTest(f"{tokenizer.__class__.__name__}"):
|
| 114 |
+
# Isolate this from the other tests because we save additional tokens/etc
|
| 115 |
+
tmpdirname = tempfile.mkdtemp()
|
| 116 |
+
|
| 117 |
+
sample_text = " He is very happy, UNwant\u00e9d,running"
|
| 118 |
+
|
| 119 |
+
additional_special_tokens = tokenizer.additional_special_tokens
|
| 120 |
+
|
| 121 |
+
# We can add a new special token for Canine as follows:
|
| 122 |
+
new_additional_special_token = chr(0xE007)
|
| 123 |
+
additional_special_tokens.append(new_additional_special_token)
|
| 124 |
+
tokenizer.add_special_tokens(
|
| 125 |
+
{"additional_special_tokens": additional_special_tokens}, replace_additional_special_tokens=False
|
| 126 |
+
)
|
| 127 |
+
before_tokens = tokenizer.encode(sample_text, add_special_tokens=False)
|
| 128 |
+
tokenizer.save_pretrained(tmpdirname)
|
| 129 |
+
|
| 130 |
+
after_tokenizer = tokenizer.__class__.from_pretrained(tmpdirname)
|
| 131 |
+
after_tokens = after_tokenizer.encode(sample_text, add_special_tokens=False)
|
| 132 |
+
self.assertListEqual(before_tokens, after_tokens)
|
| 133 |
+
self.assertIn(new_additional_special_token, after_tokenizer.additional_special_tokens)
|
| 134 |
+
self.assertEqual(after_tokenizer.model_max_length, 42)
|
| 135 |
+
|
| 136 |
+
tokenizer = tokenizer.__class__.from_pretrained(tmpdirname, model_max_length=43)
|
| 137 |
+
self.assertEqual(tokenizer.model_max_length, 43)
|
| 138 |
+
|
| 139 |
+
shutil.rmtree(tmpdirname)
|
| 140 |
+
|
| 141 |
+
def test_add_special_tokens(self):
|
| 142 |
+
tokenizers = self.get_tokenizers(do_lower_case=False)
|
| 143 |
+
for tokenizer in tokenizers:
|
| 144 |
+
with self.subTest(f"{tokenizer.__class__.__name__}"):
|
| 145 |
+
input_text, ids = self.get_clean_sequence(tokenizer)
|
| 146 |
+
|
| 147 |
+
# a special token for Canine can be defined as follows:
|
| 148 |
+
SPECIAL_TOKEN = 0xE005
|
| 149 |
+
special_token = chr(SPECIAL_TOKEN)
|
| 150 |
+
|
| 151 |
+
tokenizer.add_special_tokens({"cls_token": special_token})
|
| 152 |
+
encoded_special_token = tokenizer.encode(special_token, add_special_tokens=False)
|
| 153 |
+
self.assertEqual(len(encoded_special_token), 1)
|
| 154 |
+
|
| 155 |
+
text = tokenizer.decode(ids + encoded_special_token, clean_up_tokenization_spaces=False)
|
| 156 |
+
encoded = tokenizer.encode(text, add_special_tokens=False)
|
| 157 |
+
|
| 158 |
+
input_encoded = tokenizer.encode(input_text, add_special_tokens=False)
|
| 159 |
+
special_token_id = tokenizer.encode(special_token, add_special_tokens=False)
|
| 160 |
+
self.assertEqual(encoded, input_encoded + special_token_id)
|
| 161 |
+
|
| 162 |
+
decoded = tokenizer.decode(encoded, skip_special_tokens=True)
|
| 163 |
+
self.assertTrue(special_token not in decoded)
|
| 164 |
+
|
| 165 |
+
def test_tokenize_special_tokens(self):
|
| 166 |
+
tokenizers = self.get_tokenizers(do_lower_case=True)
|
| 167 |
+
for tokenizer in tokenizers:
|
| 168 |
+
with self.subTest(f"{tokenizer.__class__.__name__}"):
|
| 169 |
+
SPECIAL_TOKEN_1 = chr(0xE005)
|
| 170 |
+
SPECIAL_TOKEN_2 = chr(0xE006)
|
| 171 |
+
tokenizer.add_tokens([SPECIAL_TOKEN_1], special_tokens=True)
|
| 172 |
+
tokenizer.add_special_tokens({"additional_special_tokens": [SPECIAL_TOKEN_2]})
|
| 173 |
+
|
| 174 |
+
token_1 = tokenizer.tokenize(SPECIAL_TOKEN_1)
|
| 175 |
+
token_2 = tokenizer.tokenize(SPECIAL_TOKEN_2)
|
| 176 |
+
|
| 177 |
+
self.assertEqual(len(token_1), 1)
|
| 178 |
+
self.assertEqual(len(token_2), 1)
|
| 179 |
+
self.assertEqual(token_1[0], SPECIAL_TOKEN_1)
|
| 180 |
+
self.assertEqual(token_2[0], SPECIAL_TOKEN_2)
|
| 181 |
+
|
| 182 |
+
@require_tokenizers
|
| 183 |
+
def test_added_token_serializable(self):
|
| 184 |
+
tokenizers = self.get_tokenizers(do_lower_case=False)
|
| 185 |
+
for tokenizer in tokenizers:
|
| 186 |
+
with self.subTest(f"{tokenizer.__class__.__name__}"):
|
| 187 |
+
# a special token for Canine can be defined as follows:
|
| 188 |
+
NEW_TOKEN = 0xE006
|
| 189 |
+
new_token = chr(NEW_TOKEN)
|
| 190 |
+
|
| 191 |
+
new_token = AddedToken(new_token, lstrip=True)
|
| 192 |
+
tokenizer.add_special_tokens({"additional_special_tokens": [new_token]})
|
| 193 |
+
|
| 194 |
+
with tempfile.TemporaryDirectory() as tmp_dir_name:
|
| 195 |
+
tokenizer.save_pretrained(tmp_dir_name)
|
| 196 |
+
tokenizer.from_pretrained(tmp_dir_name)
|
| 197 |
+
|
| 198 |
+
def test_special_tokens_initialization_with_non_empty_additional_special_tokens(self):
|
| 199 |
+
tokenizer_list = []
|
| 200 |
+
if self.test_slow_tokenizer:
|
| 201 |
+
tokenizer_list.append((self.tokenizer_class, self.get_tokenizer()))
|
| 202 |
+
|
| 203 |
+
if self.test_rust_tokenizer:
|
| 204 |
+
tokenizer_list.append((self.rust_tokenizer_class, self.get_rust_tokenizer()))
|
| 205 |
+
|
| 206 |
+
for tokenizer_class, tokenizer_utils in tokenizer_list:
|
| 207 |
+
with tempfile.TemporaryDirectory() as tmp_dir:
|
| 208 |
+
tokenizer_utils.save_pretrained(tmp_dir)
|
| 209 |
+
|
| 210 |
+
with open(os.path.join(tmp_dir, "special_tokens_map.json"), encoding="utf-8") as json_file:
|
| 211 |
+
special_tokens_map = json.load(json_file)
|
| 212 |
+
|
| 213 |
+
with open(os.path.join(tmp_dir, "tokenizer_config.json"), encoding="utf-8") as json_file:
|
| 214 |
+
tokenizer_config = json.load(json_file)
|
| 215 |
+
|
| 216 |
+
# a special token for Canine can be defined as follows:
|
| 217 |
+
NEW_TOKEN = 0xE006
|
| 218 |
+
new_token_1 = chr(NEW_TOKEN)
|
| 219 |
+
|
| 220 |
+
special_tokens_map["additional_special_tokens"] = [new_token_1]
|
| 221 |
+
tokenizer_config["additional_special_tokens"] = [new_token_1]
|
| 222 |
+
|
| 223 |
+
with open(os.path.join(tmp_dir, "special_tokens_map.json"), "w", encoding="utf-8") as outfile:
|
| 224 |
+
json.dump(special_tokens_map, outfile)
|
| 225 |
+
with open(os.path.join(tmp_dir, "tokenizer_config.json"), "w", encoding="utf-8") as outfile:
|
| 226 |
+
json.dump(tokenizer_config, outfile)
|
| 227 |
+
|
| 228 |
+
# the following checks allow us to verify that our test works as expected, i.e. that the tokenizer takes
|
| 229 |
+
# into account the new value of additional_special_tokens given in the "tokenizer_config.json" and
|
| 230 |
+
# "special_tokens_map.json" files
|
| 231 |
+
tokenizer_without_change_in_init = tokenizer_class.from_pretrained(tmp_dir, extra_ids=0)
|
| 232 |
+
self.assertIn(new_token_1, tokenizer_without_change_in_init.additional_special_tokens)
|
| 233 |
+
# self.assertIn("an_additional_special_token",tokenizer_without_change_in_init.get_vocab()) # ByT5Tokenization no vocab
|
| 234 |
+
self.assertEqual(
|
| 235 |
+
[new_token_1],
|
| 236 |
+
tokenizer_without_change_in_init.convert_ids_to_tokens(
|
| 237 |
+
tokenizer_without_change_in_init.convert_tokens_to_ids([new_token_1])
|
| 238 |
+
),
|
| 239 |
+
)
|
| 240 |
+
|
| 241 |
+
NEW_TOKEN = 0xE007
|
| 242 |
+
new_token_2 = chr(NEW_TOKEN)
|
| 243 |
+
# Now we test that we can change the value of additional_special_tokens in the from_pretrained
|
| 244 |
+
new_added_tokens = [AddedToken(new_token_2, lstrip=True)]
|
| 245 |
+
tokenizer = tokenizer_class.from_pretrained(
|
| 246 |
+
tmp_dir, additional_special_tokens=new_added_tokens, extra_ids=0
|
| 247 |
+
)
|
| 248 |
+
|
| 249 |
+
self.assertIn(new_token_2, tokenizer.additional_special_tokens)
|
| 250 |
+
# self.assertIn(new_token_2,tokenizer.get_vocab()) # ByT5Tokenization no vocab
|
| 251 |
+
self.assertEqual(
|
| 252 |
+
[new_token_2], tokenizer.convert_ids_to_tokens(tokenizer.convert_tokens_to_ids([new_token_2]))
|
| 253 |
+
)
|
| 254 |
+
|
| 255 |
+
@require_tokenizers
|
| 256 |
+
def test_encode_decode_with_spaces(self):
|
| 257 |
+
tokenizers = self.get_tokenizers(do_lower_case=False)
|
| 258 |
+
for tokenizer in tokenizers:
|
| 259 |
+
with self.subTest(f"{tokenizer.__class__.__name__}"):
|
| 260 |
+
input = "hello world"
|
| 261 |
+
if self.space_between_special_tokens:
|
| 262 |
+
output = "[CLS] hello world [SEP]"
|
| 263 |
+
else:
|
| 264 |
+
output = input
|
| 265 |
+
encoded = tokenizer.encode(input, add_special_tokens=False)
|
| 266 |
+
decoded = tokenizer.decode(encoded, spaces_between_special_tokens=self.space_between_special_tokens)
|
| 267 |
+
self.assertIn(decoded, [output, output.lower()])
|
| 268 |
+
|
| 269 |
+
# cannot use default `test_tokenizers_common_ids_setters` method because tokenizer has no vocab
|
| 270 |
+
def test_tokenizers_common_ids_setters(self):
|
| 271 |
+
tokenizers = self.get_tokenizers()
|
| 272 |
+
for tokenizer in tokenizers:
|
| 273 |
+
with self.subTest(f"{tokenizer.__class__.__name__}"):
|
| 274 |
+
attributes_list = [
|
| 275 |
+
"bos_token",
|
| 276 |
+
"eos_token",
|
| 277 |
+
"unk_token",
|
| 278 |
+
"sep_token",
|
| 279 |
+
"pad_token",
|
| 280 |
+
"cls_token",
|
| 281 |
+
"mask_token",
|
| 282 |
+
]
|
| 283 |
+
|
| 284 |
+
token_to_test_setters = "a"
|
| 285 |
+
token_id_to_test_setters = ord(token_to_test_setters)
|
| 286 |
+
|
| 287 |
+
for attr in attributes_list:
|
| 288 |
+
setattr(tokenizer, attr + "_id", None)
|
| 289 |
+
self.assertEqual(getattr(tokenizer, attr), None)
|
| 290 |
+
self.assertEqual(getattr(tokenizer, attr + "_id"), None)
|
| 291 |
+
|
| 292 |
+
setattr(tokenizer, attr + "_id", token_id_to_test_setters)
|
| 293 |
+
self.assertEqual(getattr(tokenizer, attr), token_to_test_setters)
|
| 294 |
+
self.assertEqual(getattr(tokenizer, attr + "_id"), token_id_to_test_setters)
|
| 295 |
+
|
| 296 |
+
setattr(tokenizer, "additional_special_tokens_ids", [])
|
| 297 |
+
self.assertListEqual(getattr(tokenizer, "additional_special_tokens"), [])
|
| 298 |
+
self.assertListEqual(getattr(tokenizer, "additional_special_tokens_ids"), [])
|
| 299 |
+
|
| 300 |
+
additional_special_token_id = 0xE006
|
| 301 |
+
additional_special_token = chr(additional_special_token_id)
|
| 302 |
+
setattr(tokenizer, "additional_special_tokens_ids", [additional_special_token_id])
|
| 303 |
+
self.assertListEqual(getattr(tokenizer, "additional_special_tokens"), [additional_special_token])
|
| 304 |
+
self.assertListEqual(getattr(tokenizer, "additional_special_tokens_ids"), [additional_special_token_id])
|
| 305 |
+
|
| 306 |
+
@unittest.skip(reason="tokenizer has a fixed vocab_size (namely all possible unicode code points)")
|
| 307 |
+
def test_add_tokens_tokenizer(self):
|
| 308 |
+
pass
|
| 309 |
+
|
| 310 |
+
# CanineTokenizer does not support do_lower_case = True, as each character has its own Unicode code point
|
| 311 |
+
# ("b" and "B" for example have different Unicode code points)
|
| 312 |
+
@unittest.skip(reason="CanineTokenizer does not support do_lower_case = True")
|
| 313 |
+
def test_added_tokens_do_lower_case(self):
|
| 314 |
+
pass
|
| 315 |
+
|
| 316 |
+
@unittest.skip(reason="CanineModel does not support the get_input_embeddings nor the get_vocab method")
|
| 317 |
+
def test_np_encode_plus_sent_to_model(self):
|
| 318 |
+
pass
|
| 319 |
+
|
| 320 |
+
@unittest.skip(reason="CanineModel does not support the get_input_embeddings nor the get_vocab method")
|
| 321 |
+
def test_torch_encode_plus_sent_to_model(self):
|
| 322 |
+
pass
|
| 323 |
+
|
| 324 |
+
@unittest.skip(reason="CanineTokenizer does not have vocabulary")
|
| 325 |
+
def test_get_vocab(self):
|
| 326 |
+
pass
|
| 327 |
+
|
| 328 |
+
@unittest.skip(reason="inputs cannot be pretokenized since ids depend on whole input string")
|
| 329 |
+
def test_pretokenized_inputs(self):
|
| 330 |
+
pass
|
| 331 |
+
|
| 332 |
+
@unittest.skip(reason="CanineTokenizer does not have vocabulary")
|
| 333 |
+
def test_conversion_reversible(self):
|
| 334 |
+
pass
|
tasks/tasksmith-5dd11b34cce1/tests/source/tests/models/chameleon/__init__.py
ADDED
|
File without changes
|