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")

---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
Cell In[2], line 6
from datasets import load_dataset
datasets.logging.disable_progress_bar()
----> 6 dataset = load_dataset("glue", "qnli")

File ~/opt/venv/lib/python3.12/site-packages/datasets/load.py:2112, in load_dataset(path, name, data_dir, data_files, split, cache_dir, features, download_config, download_mode, verification_mode, ignore_verifications, keep_in_memory, save_infos, revision, token, use_auth_token, task, streaming, num_proc, storage_options, **config_kwargs)
verification_mode = VerificationMode(
   (verification_mode or VerificationMode.BASIC_CHECKS) if not save_infos else VerificationMode.ALL_CHECKS
)
# Create a dataset builder
-> 2112 builder_instance = load_dataset_builder(
   path=path,
   name=name,
   data_dir=data_dir,
   data_files=data_files,
   cache_dir=cache_dir,
   features=features,
   download_config=download_config,
   download_mode=download_mode,
   revision=revision,
   token=token,
   storage_options=storage_options,
   **config_kwargs,
)
# Return iterable dataset in case of streaming
if streaming:

File ~/opt/venv/lib/python3.12/site-packages/datasets/load.py:1798, in load_dataset_builder(path, name, data_dir, data_files, cache_dir, features, download_config, download_mode, revision, token, use_auth_token, storage_options, **config_kwargs)
   download_config = download_config.copy() if download_config else DownloadConfig()
   download_config.storage_options.update(storage_options)
-> 1798 dataset_module = dataset_module_factory(
   path,
   revision=revision,
   download_config=download_config,
   download_mode=download_mode,
   data_dir=data_dir,
   data_files=data_files,
)
# Get dataset builder class from the processing script
builder_kwargs = dataset_module.builder_kwargs

File ~/opt/venv/lib/python3.12/site-packages/datasets/load.py:1495, in dataset_module_factory(path, revision, download_config, download_mode, dynamic_modules_path, data_dir, data_files, **download_kwargs)
           if isinstance(e1, FileNotFoundError):
               raise FileNotFoundError(
                   f"Couldn't find a dataset script at {relative_to_absolute_path(combined_path)} or any data file in the same directory. "
                   f"Couldn't find '{path}' on the Hugging Face Hub either: {type(e1).__name__}: {e1}"
               ) from None
-> 1495             raise e1 from None
else:
   raise FileNotFoundError(
       f"Couldn't find a dataset script at {relative_to_absolute_path(combined_path)} or any data file in the same directory."
   )

File ~/opt/venv/lib/python3.12/site-packages/datasets/load.py:1479, in dataset_module_factory(path, revision, download_config, download_mode, dynamic_modules_path, data_dir, data_files, **download_kwargs)
       return HubDatasetModuleFactoryWithScript(
           path,
           revision=revision,
   (...)
           dynamic_modules_path=dynamic_modules_path,
       ).get_module()
   else:
       return HubDatasetModuleFactoryWithoutScript(
           path,
           revision=revision,
           data_dir=data_dir,
           data_files=data_files,
           download_config=download_config,
           download_mode=download_mode,
-> 1479         ).get_module()
except (
   Exception
) as e1:  # noqa all the attempts failed, before raising the error we should check if the module is already cached.
   try:

File ~/opt/venv/lib/python3.12/site-packages/datasets/load.py:1034, in HubDatasetModuleFactoryWithoutScript.get_module(self)
metadata_configs = MetadataConfigs.from_dataset_card_data(dataset_card_data)
dataset_infos = DatasetInfosDict.from_dataset_card_data(dataset_card_data)
patterns = (
   sanitize_patterns(self.data_files)
   if self.data_files is not None
-> 1034     else get_data_patterns(base_path, download_config=self.download_config)
)
data_files = DataFilesDict.from_patterns(
   patterns,
   base_path=base_path,
   allowed_extensions=ALL_ALLOWED_EXTENSIONS,
   download_config=self.download_config,
)
module_name, default_builder_kwargs = infer_module_for_data_files(
   data_files=data_files,
   path=self.name,
   download_config=self.download_config,
)

File ~/opt/venv/lib/python3.12/site-packages/datasets/data_files.py:457, in get_data_patterns(base_path, download_config)
resolver = partial(resolve_pattern, base_path=base_path, download_config=download_config)
try:
--> 457     return _get_data_files_patterns(resolver)
except FileNotFoundError:
   raise EmptyDatasetError(f"The directory at {base_path} doesn't contain any data files") from None

File ~/opt/venv/lib/python3.12/site-packages/datasets/data_files.py:248, in _get_data_files_patterns(pattern_resolver)
for pattern in patterns:
   try:
--> 248         data_files = pattern_resolver(pattern)
   except FileNotFoundError:
       continue

File ~/opt/venv/lib/python3.12/site-packages/datasets/data_files.py:332, in resolve_pattern(pattern, base_path, allowed_extensions, download_config)
   base_path = ""
pattern, storage_options = _prepare_path_and_storage_options(pattern, download_config=download_config)
--> 332 fs, _, _ = get_fs_token_paths(pattern, storage_options=storage_options)
fs_base_path = base_path.split("::")[0].split("://")[-1] or fs.root_marker
fs_pattern = pattern.split("::")[0].split("://")[-1]

File ~/opt/venv/lib/python3.12/site-packages/fsspec/core.py:686, in get_fs_token_paths(urlpath, mode, num, name_function, storage_options, protocol, expand)
   paths = _expand_paths(paths, name_function, num)
elif "*" in paths:
--> 686     paths = [f for f in sorted(fs.glob(paths)) if not fs.isdir(f)]
else:
   paths = [paths]

File ~/opt/venv/lib/python3.12/site-packages/huggingface_hub/hf_file_system.py:521, in HfFileSystem.glob(self, path, **kwargs)
kwargs = {"expand_info": kwargs.get("detail", False), **kwargs}
path = self.resolve_path(path, revision=kwargs.get("revision")).unresolve()
--> 521 return super().glob(path, **kwargs)

File ~/opt/venv/lib/python3.12/site-packages/fsspec/spec.py:639, in AbstractFileSystem.glob(self, path, maxdepth, **kwargs)
       depth = None
allpaths = self.find(root, maxdepth=depth, withdirs=True, detail=True, **kwargs)
--> 639 pattern = glob_translate(path + ("/" if ends_with_sep else ""))
pattern = re.compile(pattern)
out = {
   p: info
   for p, info in sorted(allpaths.items())
   (...)
   )
}

File ~/opt/venv/lib/python3.12/site-packages/fsspec/utils.py:729, in glob_translate(pat)
   continue
elif "**" in part:
--> 729     raise ValueError(
       "Invalid pattern: '**' can only be an entire path component"
   )
if part:
   results.extend(_translate(part, f"{not_sep}*", not_sep))

ValueError: Invalid pattern: '**' can only be an entire path component

dataset['train']

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/")

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",
        "optim.max_epochs": 2,
    }
)

print(student_predictor.evaluate(data=test_df))

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.