Single GPU Billion-scale Model Training via Parameter-Efficient Finetuning#
As pointed out by a recent paper from Stanford Institute for Human-Centered Artificial Intelligence, AI is undergoing a paradigm shift with the rise of “foundation models”, i.e., giant models that are trained on a diverse collection of datasets generally in a self-supervised way. These foundation models, which are the key of AutoMM, can be easily adapted to down-stream applications. However, as the size of these foundation models grows, finetuning these models becomes increasingly difficult. Following is a figure from the Microsoft research blog that demonstrates the trend:
The goal of AutoMM is to help anyone solve machine learning problems via open source foundation models, including these giant models. To finetune these large-scale models, we adopt the recently popularized parameter-efficient finetuning technique. The idea is to either finetune a small subset of the weights in the foundation model (e.g., BitFit), or adding a tiny tunable structure on top of the fixed backbone (e.g., Prompt Tuning, LoRA, Adapter, MAM Adapter, IA^3). These techniques can effectively reduce the peak memory usage and model training time, while maintaining the performance.
In this tutorial, we introduce how to apply parameter-efficient finetuning in MultiModalPredictor.
We first introduce how to adopt the "ia3_bias" algorithm for parameter-efficient finetuning. Afterwards, we show how you can simply combine "ia3_bias"
and gradient checkpointing to finetune the XL-variant of Google’s FLAN-T5 via a single NVIDIA T4 GPU.
Prepare Dataset#
The Cross-Lingual Amazon Product Review Sentiment dataset contains Amazon product reviews in four languages.
Here, we load the English and German fold of the dataset. In the label column, 0 means negative sentiment and 1 means positive sentiment.
For the purpose of demonstration, we downsampled the training data to 1000 samples. We will train the model on the English dataset and
directly evaluate its performance on the German and Japanese test set.
!wget --quiet https://automl-mm-bench.s3.amazonaws.com/multilingual-datasets/amazon_review_sentiment_cross_lingual.zip -O amazon_review_sentiment_cross_lingual.zip
!unzip -q -o amazon_review_sentiment_cross_lingual.zip -d .
import os
import shutil
os.environ["TRANSFORMERS_CACHE"] = "cache"
def clear_cache():
if os.path.exists("cache"):
shutil.rmtree("cache")
clear_cache()
import pandas as pd
import warnings
warnings.filterwarnings("ignore")
train_en_df = pd.read_csv("amazon_review_sentiment_cross_lingual/en_train.tsv",
sep="\t",
header=None,
names=["label", "text"]) \
.sample(1000, random_state=123).reset_index(drop=True)
test_en_df = pd.read_csv("amazon_review_sentiment_cross_lingual/en_test.tsv",
sep="\t",
header=None,
names=["label", "text"]) \
.sample(200, random_state=123).reset_index(drop=True)
test_de_df = pd.read_csv("amazon_review_sentiment_cross_lingual/de_test.tsv",
sep="\t", header=None, names=["label", "text"]) \
.sample(200, random_state=123).reset_index(drop=True)
test_jp_df = pd.read_csv('amazon_review_sentiment_cross_lingual/jp_test.tsv',
sep='\t', header=None, names=['label', 'text']) \
.sample(200, random_state=123).reset_index(drop=True)
train_en_df.head(5)
| label | text | |
|---|---|---|
| 0 | 0 | This is a film that literally sees little wron... |
| 1 | 0 | This music is pretty intelligent, but not very... |
| 2 | 0 | One of the best pieces of rock ever recorded, ... |
| 3 | 0 | Reading the posted reviews here, is like revis... |
| 4 | 1 | I've just finished page 341, the last page. It... |
test_jp_df.head(5)
| label | text | |
|---|---|---|
| 0 | 1 | 原作はビクトル・ユーゴの長編小説だが、私が子供の頃読んだのは短縮版の「ああ無情」。それでもこ... |
| 1 | 1 | ほかの作品のレビューにみんな書いているのに、何故この作品について書いている人が一人しかいない... |
| 2 | 0 | 一番の問題点は青島が出ていない事でしょう。 ＴＶ番組では『芸人が出ていればバラエティだから... |
| 3 | 0 | 昔、 りんたろう監督によるアニメ「カムイの剣」があった。 「カムイの剣」…を観た人なら本作... |
| 4 | 1 | 以前のアルバムを聴いていないのでなんとも言えないが、クラシックなメタルを聞いてきた耳には、と... |
Finetuning Multilingual Model with IA3 + BitFit#
In AutoMM, to enable efficient finetuning, just specify the optimization.efficient_finetune to be "ia3_bias".
from autogluon.multimodal import MultiModalPredictor
import uuid
model_path = f"./tmp/{uuid.uuid4().hex}-multilingual_ia3"
predictor = MultiModalPredictor(label="label",
path=model_path)
predictor.fit(train_en_df,
presets="multilingual",
hyperparameters={
"optimization.efficient_finetune": "ia3_bias",
"optimization.lr_decay": 0.9,
"optimization.learning_rate": 3e-03,
"optimization.end_lr": 3e-03,
"optimization.max_epochs": 2,
"optimization.warmup_steps": 0,
"env.batch_size": 32,
})
AutoGluon infers your prediction problem is: 'binary' (because only two unique label-values observed).
2 unique label values: [0, 1]
If 'binary' is not the correct problem_type, please manually specify the problem_type parameter during predictor init (You may specify problem_type as one of: ['binary', 'multiclass', 'regression'])
Global seed set to 0
AutoMM starts to create your model. ✨
- AutoGluon version is 0.8.1b20230622.
- Pytorch version is 1.13.1+cu117.
- Model will be saved to "/home/ci/autogluon/docs/tutorials/multimodal/advanced_topics/tmp/9cdb6292a5704d8e8a0bd70f028fdc3d-multilingual_ia3".
- Validation metric is "roc_auc".
- To track the learning progress, you can open a terminal and launch Tensorboard:
```shell
# Assume you have installed tensorboard
tensorboard --logdir /home/ci/autogluon/docs/tutorials/multimodal/advanced_topics/tmp/9cdb6292a5704d8e8a0bd70f028fdc3d-multilingual_ia3
```
Enjoy your coffee, and let AutoMM do the job ☕☕☕ Learn more at https://auto.gluon.ai
╭─────────────────────────────── Traceback (most recent call last) ────────────────────────────────╮ │ in <module>:7 │ │ │ │ 4 model_path = f"./tmp/{uuid.uuid4().hex}-multilingual_ia3" │ │ 5 predictor = MultiModalPredictor(label="label", │ │ 6 │ │ │ │ │ │ │ │ path=model_path) │ │ ❱ 7 predictor.fit(train_en_df, │ │ 8 │ │ │ presets="multilingual", │ │ 9 │ │ │ hyperparameters={ │ │ 10 │ │ │ │ "optimization.efficient_finetune": "ia3_bias", │ │ │ │ /home/ci/autogluon/multimodal/src/autogluon/multimodal/predictor.py:864 in fit │ │ │ │ 861 │ │ │ ) │ │ 862 │ │ │ return predictor │ │ 863 │ │ │ │ ❱ 864 │ │ self._fit(**_fit_args) │ │ 865 │ │ training_end = time.time() │ │ 866 │ │ self._total_train_time = training_end - training_start │ │ 867 │ │ │ │ /home/ci/autogluon/multimodal/src/autogluon/multimodal/predictor.py:1140 in _fit │ │ │ │ 1137 │ │ │ self._output_shape = len(df_preprocessor.label_generator.unique_entity_group │ │ 1138 │ │ │ │ 1139 │ │ if self._model is None: │ │ ❱ 1140 │ │ │ model = create_fusion_model( │ │ 1141 │ │ │ │ config=config, │ │ 1142 │ │ │ │ num_classes=self._output_shape, │ │ 1143 │ │ │ │ classes=self._classes, │ │ │ │ /home/ci/autogluon/multimodal/src/autogluon/multimodal/utils/model.py:442 in create_fusion_model │ │ │ │ 439 │ │ │ 440 │ for model_name in names: │ │ 441 │ │ model_config = getattr(config.model, model_name) │ │ ❱ 442 │ │ model = create_model( │ │ 443 │ │ │ model_name=model_name, │ │ 444 │ │ │ model_config=model_config, │ │ 445 │ │ │ num_classes=num_classes, │ │ │ │ /home/ci/autogluon/multimodal/src/autogluon/multimodal/utils/model.py:209 in create_model │ │ │ │ 206 │ │ │ pretrained=pretrained, │ │ 207 │ │ ) │ │ 208 │ elif model_name.lower().startswith(HF_TEXT): │ │ ❱ 209 │ │ model = HFAutoModelForTextPrediction( │ │ 210 │ │ │ prefix=model_name, │ │ 211 │ │ │ checkpoint_name=model_config.checkpoint_name, │ │ 212 │ │ │ num_classes=num_classes, │ │ │ │ /home/ci/autogluon/multimodal/src/autogluon/multimodal/models/huggingface_text.py:84 in __init__ │ │ │ │ 81 │ │ self.config, self.model = get_hf_config_and_model( │ │ 82 │ │ │ checkpoint_name=checkpoint_name, pretrained=pretrained, low_cpu_mem_usage=lo │ │ 83 │ │ ) │ │ ❱ 84 │ │ self._hf_model_input_names = AutoTokenizer.from_pretrained(checkpoint_name).mode │ │ 85 │ │ │ │ 86 │ │ if isinstance(self.model, T5PreTrainedModel): │ │ 87 │ │ │ self.is_t5 = True │ │ │ │ /home/ci/opt/venv/lib/python3.8/site-packages/transformers/models/auto/tokenization_auto.py:676 │ │ in from_pretrained │ │ │ │ 673 │ │ if model_type is not None: │ │ 674 │ │ │ tokenizer_class_py, tokenizer_class_fast = TOKENIZER_MAPPING[type(config)] │ │ 675 │ │ │ if tokenizer_class_fast and (use_fast or tokenizer_class_py is None): │ │ ❱ 676 │ │ │ │ return tokenizer_class_fast.from_pretrained(pretrained_model_name_or_pat │ │ 677 │ │ │ else: │ │ 678 │ │ │ │ if tokenizer_class_py is not None: │ │ 679 │ │ │ │ │ return tokenizer_class_py.from_pretrained(pretrained_model_name_or_p │ │ │ │ /home/ci/opt/venv/lib/python3.8/site-packages/transformers/tokenization_utils_base.py:1804 in │ │ from_pretrained │ │ │ │ 1801 │ │ │ else: │ │ 1802 │ │ │ │ logger.info(f"loading file {file_path} from cache at {resolved_vocab_fil │ │ 1803 │ │ │ │ ❱ 1804 │ │ return cls._from_pretrained( │ │ 1805 │ │ │ resolved_vocab_files, │ │ 1806 │ │ │ pretrained_model_name_or_path, │ │ 1807 │ │ │ init_configuration, │ │ │ │ /home/ci/opt/venv/lib/python3.8/site-packages/transformers/tokenization_utils_base.py:1959 in │ │ _from_pretrained │ │ │ │ 1956 │ │ │ │ 1957 │ │ # Instantiate tokenizer. │ │ 1958 │ │ try: │ │ ❱ 1959 │ │ │ tokenizer = cls(*init_inputs, **init_kwargs) │ │ 1960 │ │ except OSError: │ │ 1961 │ │ │ raise OSError( │ │ 1962 │ │ │ │ "Unable to load vocabulary from file. " │ │ │ │ /home/ci/opt/venv/lib/python3.8/site-packages/transformers/models/deberta_v2/tokenization_debert │ │ a_v2_fast.py:133 in __init__ │ │ │ │ 130 │ │ mask_token="[MASK]", │ │ 131 │ │ **kwargs │ │ 132 │ ) -> None: │ │ ❱ 133 │ │ super().__init__( │ │ 134 │ │ │ vocab_file, │ │ 135 │ │ │ tokenizer_file=tokenizer_file, │ │ 136 │ │ │ do_lower_case=do_lower_case, │ │ │ │ /home/ci/opt/venv/lib/python3.8/site-packages/transformers/tokenization_utils_fast.py:114 in │ │ __init__ │ │ │ │ 111 │ │ │ fast_tokenizer = TokenizerFast.from_file(fast_tokenizer_file) │ │ 112 │ │ elif slow_tokenizer is not None: │ │ 113 │ │ │ # We need to convert a slow tokenizer to build the backend │ │ ❱ 114 │ │ │ fast_tokenizer = convert_slow_tokenizer(slow_tokenizer) │ │ 115 │ │ elif self.slow_tokenizer_class is not None: │ │ 116 │ │ │ # We need to create and convert a slow tokenizer to build the backend │ │ 117 │ │ │ slow_tokenizer = self.slow_tokenizer_class(*args, **kwargs) │ │ │ │ /home/ci/opt/venv/lib/python3.8/site-packages/transformers/convert_slow_tokenizer.py:1162 in │ │ convert_slow_tokenizer │ │ │ │ 1159 │ │ │ 1160 │ converter_class = SLOW_TO_FAST_CONVERTERS[tokenizer_class_name] │ │ 1161 │ │ │ ❱ 1162 │ return converter_class(transformer_tokenizer).converted() │ │ 1163 │ │ │ │ /home/ci/opt/venv/lib/python3.8/site-packages/transformers/convert_slow_tokenizer.py:438 in │ │ __init__ │ │ │ │ 435 │ │ │ │ 436 │ │ super().__init__(*args) │ │ 437 │ │ │ │ ❱ 438 │ │ from .utils import sentencepiece_model_pb2 as model_pb2 │ │ 439 │ │ │ │ 440 │ │ m = model_pb2.ModelProto() │ │ 441 │ │ with open(self.original_tokenizer.vocab_file, "rb") as f: │ │ │ │ /home/ci/opt/venv/lib/python3.8/site-packages/transformers/utils/sentencepiece_model_pb2.py:92 │ │ in <module> │ │ │ │ 89 │ file=DESCRIPTOR, │ │ 90 │ create_key=_descriptor._internal_create_key, │ │ 91 │ values=[ │ │ ❱ 92 │ │ _descriptor.EnumValueDescriptor( │ │ 93 │ │ │ name="UNIGRAM", │ │ 94 │ │ │ index=0, │ │ 95 │ │ │ number=1, │ │ │ │ /home/ci/opt/venv/lib/python3.8/site-packages/google/protobuf/descriptor.py:796 in __new__ │ │ │ │ 793 │ def __new__(cls, name, index, number, │ │ 794 │ │ │ │ type=None, # pylint: disable=redefined-builtin │ │ 795 │ │ │ │ options=None, serialized_options=None, create_key=None): │ │ ❱ 796 │ _message.Message._CheckCalledFromGeneratedFile() │ │ 797 │ # There is no way we can build a complete EnumValueDescriptor with the │ │ 798 │ # given parameters (the name of the Enum is not known, for example). │ │ 799 │ # Fortunately generated files just pass it to the EnumDescriptor() │ ╰──────────────────────────────────────────────────────────────────────────────────────────────────╯ TypeError: Descriptors cannot not be created directly. If this call came from a _pb2.py file, your generated code is out of date and must be regenerated with protoc >= 3.19.0. If you cannot immediately regenerate your protos, some other possible workarounds are: 1. Downgrade the protobuf package to 3.20.x or lower. 2. Set PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python (but this will use pure-Python parsing and will be much slower). More information: https://developers.google.com/protocol-buffers/docs/news/2022-05-06#python-updates
The fraction of the tunable parameters is around 0.5% of all parameters. Actually, the model trained purely on English data can achieve good performance on the test sets, even on the German / Japanese test set. It obtained comparable results as full-finetuning as in AutoMM for Text - Multilingual Problems.
score_in_en = predictor.evaluate(test_en_df)
score_in_de = predictor.evaluate(test_de_df)
score_in_jp = predictor.evaluate(test_jp_df)
print('Score in the English Testset:', score_in_en)
print('Score in the German Testset:', score_in_de)
print('Score in the Japanese Testset:', score_in_jp)
Training FLAN-T5-XL on Single GPU#
By combining gradient checkpointing and parameter-efficient finetuning, it is feasible to finetune
google/flan-t5-xl that has close to two billion parameterswith a single T4 GPU available in
AWS G4 instances.
To turn on gradient checkpointing, you just need to set "model.hf_text.gradient_checkpointing" to True.
To accelerate the training, we downsample the number of training samples to be 200.
# Just for clean the space
clear_cache()
shutil.rmtree(model_path)
from autogluon.multimodal import MultiModalPredictor
train_en_df_downsample = train_en_df.sample(200, random_state=123)
new_model_path = f"./tmp/{uuid.uuid4().hex}-multilingual_ia3_gradient_checkpoint"
predictor = MultiModalPredictor(label="label",
path=new_model_path)
predictor.fit(train_en_df_downsample,
presets="multilingual",
hyperparameters={
"model.hf_text.checkpoint_name": "google/flan-t5-xl",
"model.hf_text.gradient_checkpointing": True,
"model.hf_text.low_cpu_mem_usage": True,
"optimization.efficient_finetune": "ia3_bias",
"optimization.lr_decay": 0.9,
"optimization.learning_rate": 3e-03,
"optimization.end_lr": 3e-03,
"optimization.max_epochs": 1,
"optimization.warmup_steps": 0,
"env.batch_size": 1,
"env.eval_batch_size_ratio": 1
})
Global seed set to 123
Auto select gpus: [0]
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 | model | HFAutoModelForTextPrediction | 1.2 B
1 | validation_metric | AUROC | 0
2 | loss_func | CrossEntropyLoss | 0
-------------------------------------------------------------------
203 K Trainable params
1.2 B Non-trainable params
1.2 B Total params
4,894.913 Total estimated model params size (MB)
Epoch 0, global step 20: 'val_roc_auc' reached 0.88802 (best 0.88802), saving model to '/home/ubuntu/autogluon/docs/tutorials/multimodal/advanced_topics/multilingual_ia3_gradient_checkpoint/epoch=0-step=20.ckpt' as top 1
Epoch 0, global step 40: 'val_roc_auc' reached 0.94531 (best 0.94531), saving model to '/home/ubuntu/autogluon/docs/tutorials/multimodal/advanced_topics/multilingual_ia3_gradient_checkpoint/epoch=0-step=40.ckpt' as top 1
`Trainer.fit` stopped: `max_epochs=1` reached.
<autogluon.multimodal.predictor.MultiModalPredictor at 0x7fd58c4dbca0>
score_in_en = predictor.evaluate(test_en_df)
print('Score in the English Testset:', score_in_en)
Score in the English Testset: {'roc_auc': 0.931263189629183}
# Just for clean the space
clear_cache()
shutil.rmtree(new_model_path)
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.