File size: 8,040 Bytes
087e4a0 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 | from __future__ import annotations
import importlib
import importlib.util
import unittest
from unittest.mock import patch
class CocoEvaluatorRegressionTest(unittest.TestCase):
@classmethod
def setUpClass(cls) -> None:
package = "hcl"
cls.evaluator_module = importlib.import_module(f"{package}.evaluator")
cls.pipeline_module = importlib.import_module(f"{package}.pipeline")
@staticmethod
def _example(task_id: str, task_type: str, answer: object) -> dict[str, object]:
return {
"task_id": task_id,
"task_name": task_type,
"task_type": task_type,
"answer": answer,
"visible_context": {
"width": 100,
"height": 100,
"bbox_coordinate_system": "qwen_0_1000",
},
"metadata": {
"hcl_task": task_type,
"answer_bbox_coordinate_system": "pixel_xyxy",
"prediction_bbox_coordinate_system": "qwen_0_1000",
},
}
@unittest.skipUnless(importlib.util.find_spec("pycocotools"), "pycocotools is optional")
def test_detection_uses_micro_f1_and_also_reports_coco_ap(self) -> None:
example = self._example(
"det-1",
"coco_detection",
[{"label": "person", "bbox": [10, 10, 50, 50]}],
)
prediction = {
"task_id": "det-1",
"answer": [
{"label": "person", "bbox": [600, 600, 900, 900], "score": 0.9},
{"label": "person", "bbox": [100, 100, 500, 500], "score": 0.8},
],
}
metrics = self.evaluator_module.ExactMatchEvaluator().evaluate(
[example], [prediction]
)["metrics"]
self.assertNotIn("accuracy", metrics)
self.assertEqual(metrics["detection_metric_source"], "pycocotools.COCOeval")
self.assertAlmostEqual(metrics["detection_coco_ap"], 0.5)
self.assertAlmostEqual(metrics["detection_coco_ap50"], 0.5)
self.assertAlmostEqual(metrics["detection_micro_precision_iou50"], 0.5)
self.assertAlmostEqual(metrics["detection_micro_recall_iou50"], 1.0)
self.assertAlmostEqual(metrics["detection_micro_f1_iou50"], 2.0 / 3.0)
self.assertEqual(metrics["primary_metric_name"], "detection_micro_f1_iou50")
self.assertAlmostEqual(metrics["primary_score"], 2.0 / 3.0)
def test_grounding_converts_qwen_coordinates_and_reports_mean_iou(self) -> None:
example = self._example(
"ref-1", "refcoco_grounding", {"bbox": [10, 20, 50, 80]}
)
prediction = {"task_id": "ref-1", "answer": {"bbox": [100, 200, 500, 800]}}
metrics = self.evaluator_module.ExactMatchEvaluator().evaluate(
[example], [prediction]
)["metrics"]
self.assertNotIn("accuracy", metrics)
self.assertEqual(metrics["grounding_accuracy_iou50"], 1.0)
self.assertAlmostEqual(metrics["grounding_mean_iou"], 1.0)
self.assertEqual(metrics["primary_metric_name"], "grounding_accuracy_iou50")
@unittest.skipUnless(importlib.util.find_spec("pycocotools"), "pycocotools is optional")
def test_gate_top_level_is_current_task_and_anchors_stay_per_task(self) -> None:
current = self._example(
"ref-1", "refcoco_grounding", {"bbox": [10, 20, 50, 80]}
)
anchor = self._example(
"det-1",
"coco_detection",
[{"label": "person", "bbox": [10, 10, 50, 50]}],
)
predictions = [
{"task_id": "ref-1", "answer": {"bbox": [100, 200, 500, 800]}},
{
"task_id": "det-1",
"answer": [
{"label": "person", "bbox": [100, 100, 500, 500]},
{"label": "dog", "bbox": [600, 600, 900, 900]},
],
},
]
pipeline = self.pipeline_module.HCLPipeline.__new__(self.pipeline_module.HCLPipeline)
pipeline.evaluator = self.evaluator_module.ExactMatchEvaluator()
result = pipeline._evaluate_gate(
[current, anchor],
predictions,
current_val_examples=[current],
historical_anchor_examples=[anchor],
split="val",
)
self.assertEqual(result["metric_scope"], "current_task")
self.assertEqual(result["metrics"]["primary_score"], 1.0)
self.assertEqual(
result["combined_gate_metrics"]["primary_metric_name"],
"mixed_tasks_not_comparable",
)
self.assertIn("coco_detection", result["historical_anchor_task_metrics"])
self.assertEqual(
result["historical_anchor_metrics"]["primary_metric_name"],
"historical_tasks_reported_separately",
)
self.assertIsNone(result["historical_anchor_metrics"]["primary_score"])
self.assertIsNone(result["combined_gate_metrics"]["mean_example_score"])
def test_caption_composite_is_named_and_not_exposed_as_accuracy(self) -> None:
example = self._example("cap-1", "coco_caption", "a red bus")
prediction = {"task_id": "cap-1", "answer": "a bus"}
fake_caption_eval = (
{"cap-1": 0.42},
{
"caption_metric_source": "test",
"caption_count": 1,
"caption_primary_score": 0.42,
"caption_CIDEr": 0.7,
},
)
with patch.object(
self.evaluator_module,
"_official_caption_eval_scores",
return_value=fake_caption_eval,
):
metrics = self.evaluator_module.ExactMatchEvaluator().evaluate(
[example], [prediction]
)["metrics"]
self.assertNotIn("accuracy", metrics)
self.assertEqual(metrics["primary_metric_name"], "caption_composite_score")
self.assertEqual(metrics["primary_score"], 0.42)
self.assertEqual(metrics["caption_CIDEr"], 0.7)
def test_checkpoint_signatures_include_evaluation_schema(self) -> None:
version = self.evaluator_module.EVALUATION_SCHEMA_VERSION
train_signature = self.pipeline_module._train_checkpoint_signature(
split="train",
examples=[],
val_examples=[],
batch_count=0,
batchsize=None,
)
test_signature = self.pipeline_module._test_checkpoint_signature(
split="test", examples=[]
)
continual_signature = self.pipeline_module._continual_checkpoint_signature(
task_stream_path="stream.json",
task_order=[],
train_split="train",
val_split="validation",
test_split="test",
train_limit_per_task=None,
validation_limit_per_task=None,
test_limit_per_task=None,
batchsize=None,
)
self.assertEqual(train_signature["evaluation_schema_version"], version)
self.assertEqual(test_signature["evaluation_schema_version"], version)
self.assertEqual(continual_signature["evaluation_schema_version"], version)
def test_paper_anchor_thresholds_and_schema_validity(self) -> None:
caption = self._example("cap", "coco_caption", "a red bus")
vqa = self._example("vqa", "vqav2", "red")
vqa["metadata"]["answers"] = ["red", "red", "blue"]
self.assertEqual(self.evaluator_module._correct_threshold(caption), 0.5)
self.assertEqual(self.evaluator_module._correct_threshold(vqa), 1.0)
malformed_grounding = self._example(
"ground", "refcoco_grounding", {"bbox": [10, 20, 50, 80]}
)
result = self.evaluator_module.ExactMatchEvaluator().evaluate(
[malformed_grounding],
[{"task_id": "ground", "answer": "not a box"}],
)["metrics"]
self.assertEqual(result["correct"], 0)
self.assertEqual(result["format_compliant_count"], 0)
if __name__ == "__main__":
unittest.main()
|