Knowledge Distillation in AutoMM¶
Pretrained foundation models are becoming increasingly large. However, these models are difficult to deploy due to limited resources available in deployment scenarios. To benefit from large models under this constraint, you transfer the knowledge from the large-scale teacher models to the student model, with knowledge distillation. In this way, the small student model can be practically deployed under real-world scenarios, while the performance will be better than training the student model from scratch thanks to the teacher.
In this tutorial, we introduce how to adopt MultiModalPredictor for
knowledge distillation. For the purpose of demonstration, we use the
Question-answering NLI
dataset, which comprises 104,743 question, answer pairs sampled from
question answering datasets. We will demonstrate how to use a large
model to guide the learning and improve the performance of a small model
in AutoGluon.
Load Dataset¶
The Question-answering NLI
dataset contains sentence pairs in English. In the label column, 0
means that the sentence is not related to the question and 1 means
that the sentence is related to the question.
import datasets
from datasets import load_dataset
datasets.logging.disable_progress_bar()
dataset = load_dataset("glue", "qnli")
Downloading and preparing dataset glue/qnli to /home/ci/.cache/huggingface/datasets/glue/qnli/1.0.0/dacbe3125aa31d7f70367a07a8a9e72a5a0bfeb5fc42e75c9db75b96da6053ad...
Dataset glue downloaded and prepared to /home/ci/.cache/huggingface/datasets/glue/qnli/1.0.0/dacbe3125aa31d7f70367a07a8a9e72a5a0bfeb5fc42e75c9db75b96da6053ad. Subsequent calls will reuse this data.
dataset['train']
Dataset({
features: ['question', 'sentence', 'label', 'idx'],
num_rows: 104743
})
from sklearn.model_selection import train_test_split
train_valid_df = dataset["train"].to_pandas()[["question", "sentence", "label"]].sample(1000, random_state=123)
train_df, valid_df = train_test_split(train_valid_df, test_size=0.2, random_state=123)
test_df = dataset["validation"].to_pandas()[["question", "sentence", "label"]].sample(1000, random_state=123)
Load the Teacher Model¶
In our example, we will directly load a teacher model with the google/bert_uncased_L-12_H-768_A-12 backbone that has been trained on QNLI and distill it into a student model with the google/bert_uncased_L-6_H-768_A-12 backbone.
!wget --quiet https://automl-mm-bench.s3.amazonaws.com/unit-tests/distillation_sample_teacher.zip -O distillation_sample_teacher.zip
!unzip -q -o distillation_sample_teacher.zip -d .
from autogluon.multimodal import MultiModalPredictor
teacher_predictor = MultiModalPredictor.load("ag_distillation_sample_teacher/")
/home/ci/opt/venv/lib/python3.8/site-packages/sklearn/base.py:329: UserWarning: Trying to unpickle estimator LabelEncoder from version 1.0.2 when using version 1.1.3. This might lead to breaking code or invalid results. Use at your own risk. For more info please refer to:
https://scikit-learn.org/stable/model_persistence.html#security-maintainability-limitations
warnings.warn(
/home/ci/opt/venv/lib/python3.8/site-packages/sklearn/base.py:329: UserWarning: Trying to unpickle estimator StandardScaler from version 1.0.2 when using version 1.1.3. This might lead to breaking code or invalid results. Use at your own risk. For more info please refer to:
https://scikit-learn.org/stable/model_persistence.html#security-maintainability-limitations
warnings.warn(
Distill to Student¶
Training the student model is straight forward. You may just add the
teacher_predictor argument when calling .fit(). Internally, the
student will be trained by matching the prediction/feature map from the
teacher. It can perform better than directly finetuning the student.
student_predictor = MultiModalPredictor(label="label")
student_predictor.fit(
train_df,
tuning_data=valid_df,
teacher_predictor=teacher_predictor,
hyperparameters={
"model.hf_text.checkpoint_name": "google/bert_uncased_L-6_H-768_A-12",
"optimization.max_epochs": 2,
}
)
Global seed set to 123 /home/ci/opt/venv/lib/python3.8/site-packages/torchmetrics/utilities/prints.py:36: UserWarning: Metric AUROC will save all targets and predictions in buffer. For large datasets this may lead to large memory footprint. warnings.warn(*args, **kwargs) /home/ci/opt/venv/lib/python3.8/site-packages/pytorch_lightning/utilities/parsing.py:268: UserWarning: Attribute 'softmax_regression_loss_func' is an instance of nn.Module and is already saved during checkpointing. It is recommended to ignore them using self.save_hyperparameters(ignore=['softmax_regression_loss_func']). rank_zero_warn( Auto select gpus: [0] Using 16bit native Automatic Mixed Precision (AMP) GPU available: True (cuda), used: True TPU available: False, using: 0 TPU cores IPU available: False, using: 0 IPUs HPU available: False, using: 0 HPUs LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0] | Name | Type | Params ------------------------------------------------------------------------------ 0 | student_model | HFAutoModelForTextPrediction | 67.0 M 1 | teacher_model | HFAutoModelForTextPrediction | 109 M 2 | validation_metric | AUROC | 0 3 | hard_label_loss_func | CrossEntropyLoss | 0 4 | soft_label_loss_func | CrossEntropyLoss | 0 5 | softmax_regression_loss_func | MSELoss | 0 6 | output_feature_loss_func | MSELoss | 0 7 | output_feature_adaptor | Identity | 0 8 | rkd_loss_func | RKDLoss | 0 ------------------------------------------------------------------------------ 176 M Trainable params 0 Non-trainable params 176 M Total params 352.881 Total estimated model params size (MB) Epoch 0, global step 3: 'val_roc_auc' reached 0.63572 (best 0.63572), saving model to '/home/ci/autogluon/docs/_build/eval/tutorials/multimodal/advanced_topics/AutogluonModels/ag-20221117_032412/epoch=0-step=3.ckpt' as top 3 Epoch 0, global step 7: 'val_roc_auc' reached 0.69998 (best 0.69998), saving model to '/home/ci/autogluon/docs/_build/eval/tutorials/multimodal/advanced_topics/AutogluonModels/ag-20221117_032412/epoch=0-step=7.ckpt' as top 3 Epoch 1, global step 10: 'val_roc_auc' reached 0.70933 (best 0.70933), saving model to '/home/ci/autogluon/docs/_build/eval/tutorials/multimodal/advanced_topics/AutogluonModels/ag-20221117_032412/epoch=1-step=10.ckpt' as top 3 Epoch 1, global step 14: 'val_roc_auc' reached 0.71219 (best 0.71219), saving model to '/home/ci/autogluon/docs/_build/eval/tutorials/multimodal/advanced_topics/AutogluonModels/ag-20221117_032412/epoch=1-step=14.ckpt' as top 3 Trainer.fit stopped: max_epochs=2 reached.
<autogluon.multimodal.predictor.MultiModalPredictor at 0x7f061cf16280>
print(student_predictor.evaluate(data=test_df))
{'roc_auc': 0.7905329444571136}
More about Knowledge Distillation¶
To learn how to customize distillation and how it compares with direct finetuning, see the distillation examples and README in AutoMM Distillation Examples. Especially the multilingual distillation example with more details and customization.
Other Examples¶
You may go to AutoMM Examples to explore other examples about AutoMM.
Customization¶
To learn how to customize AutoMM, please refer to Customize AutoMM.