Spaces:
Paused
Paused
File size: 4,924 Bytes
2f86d9b | 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 | # Copyright 2026 DeepMind Technologies Limited.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
r"""Example of Gemma finetuning for an image captioning task.
Example:
Prompt:
```
<start_of_turn>user
<start_of_image><end_of_turn>
<start_of_turn>model
```
Target:
```
A diagram showing a circuit with a battery, lamp, and switch.<end_of_turn>
```
Here, the prompt only contains the `<start_of_image>` to indicate an image
is inserted.
Train locally with:
```sh
python -m kauldron.main \
--cfg=examples/multimodal.py \
--cfg.workdir=/tmp/kauldron_oss/workdir
```
"""
from kauldron import konfig
# pylint: disable=g-import-not-at-top
with konfig.imports():
import jax.numpy as jnp
from gemma import gm
from kauldron import kd
import optax
# pylint: enable=g-import-not-at-top
def get_config():
batch_size = 32
max_length = 200
return kd.train.Trainer(
seed=42,
# Dataset
train_ds=_make_dataset(
training=True,
batch_size=batch_size,
max_length=max_length,
),
# Model definition
model=gm.nn.Gemma3_4B(
tokens="batch.input",
images="batch.image",
),
# Load the weights from the pretrained checkpoint
init_transform=gm.ckpts.LoadCheckpoint(
path=gm.ckpts.CheckpointPath.GEMMA3_4B_IT,
),
# Training
num_train_steps=10_000,
train_losses={
"xentropy": kd.losses.SoftmaxCrossEntropyWithIntLabels(
logits="preds.logits",
labels="batch.target",
mask="batch.loss_mask",
),
},
train_summaries={
"image": kd.summaries.ShowImages(images="batch.image", num_images=5),
},
optimizer=optax.adafactor(learning_rate=1e-3),
checkpointer=kd.ckpts.Checkpointer(
save_interval_steps=500,
),
# Evaluation
evals={
"test": kd.evals.Evaluator(
run=kd.evals.EveryNSteps(1000),
ds=_make_dataset(
training=False,
batch_size=4,
max_length=max_length,
),
),
# The sampler evaluator run inference on a few prompts from the
# test set.
"sampling": gm.evals.SamplerEvaluator(
run=kd.evals.EveryNSteps(1000),
max_new_tokens=50, # Sampling parameters
num_batches=3,
ds=_make_dataset(training=False, sampling=True),
summaries={
"image": kd.summaries.ShowImages(
images="batch.image", num_images=5
),
},
),
},
)
def _make_dataset(
*,
training: bool,
sampling: bool = False,
batch_size: int | None = None,
max_length: int | None = None,
):
tokenizer = gm.text.Gemma3Tokenizer()
return kd.data.py.Tfds(
name="ai2dcaption",
split="llava_15" if training else "test",
shuffle=True if training else False,
num_epochs=None if training else 1,
batch_size=None if sampling else batch_size,
num_workers=4,
transforms=[
# Only keep the fields we need.See fields at:
# https://www.tensorflow.org/datasets/catalog/ai2dcaption
kd.data.Elements(keep=["image", "caption"]),
# Create a new constant field
kd.data.AddConstants({"prompt": "<start_of_image>"}),
# Create the model inputs/targets/loss_mask.
gm.data.Seq2SeqTask(
# Select which field from the dataset to use.
in_prompt="prompt",
in_response="caption",
# Output batch is {"input": ..., "target": ..., "loss_mask": ...}
out_input="input",
out_target="target",
out_target_mask="loss_mask",
tokenizer=tokenizer,
# Padding parameters
max_length=None if sampling else max_length, # pyrefly: ignore[bad-argument-type]
# In this dataset, ~1% of examples are longer than 512 tokens.
truncate=True,
sampling=sampling,
),
kd.data.py.Resize(key="image", size=(800, 800)),
# TODO(epot): Make the `num_images` dimension optional
kd.data.Rearrange(key="image", pattern="... h w c -> ... 1 h w c"),
kd.data.Cast(key="image", dtype=jnp.uint8),
],
)
|