/TextBrewer

A PyTorch-based knowledge distillation toolkit for natural language processing

Primary LanguagePythonApache License 2.0Apache-2.0

English | 中文说明



GitHub Documentation PyPI GitHub release

TextBrewer is a PyTorch-based model distillation toolkit for natural language processing. It includes various distillation techniques from both NLP and CV field and provides an easy-to-use distillation framework, which allows users to quickly experiment with the state-of-the-art distillation methods to compress the model with a relatively small sacrifice in the performance, increasing the inference speed and reducing the memory usage.

Check our paper through ACL Anthology or arXiv pre-print.

Full Documentation

News

Dec 17, 2021

Oct 24, 2021

Jul 8, 2021

  • New examples with Transformers 4
    • The current examples (exmaples/) have been written with old versions of Transformers and they may cause some confusions and bugs. We rewrite the examples with Transformers 4 in Jupyter Notebooks, which are easy to follow and learn.
    • The new examples can be found at examples/notebook_examples. See Examples for details.

Mar 1, 2021

  • BERT-EMD and custom distiller

    • We added an experiment with BERT-EMD in the MNLI exmaple. BERT-EMD allows each intermediate student layer to learn from any intermediate teacher layers adaptively, bassed on optimizing Earth Mover’s Distance. So there is no need to specify the mathcing scheme.
    • We have written a new EMDDistiller to perform BERT-EMD. It demonstrates how to write a custom distiller.
  • updated MNLI example

    • We removed the pretrained_pytorch_bert and used transformers library instead in all the MNLI exmaples.
Click here to see old news

Nov 11, 2020

  • Updated to 0.2.1:

    • More flexible distillation: Supports feeding different batches to the student and teacher. It means the batches for the student and teacher no longer need to be the same. It can be used for distilling models with different vocabularies (e.g., from RoBERTa to BERT).

    • Faster distillation: Users now can pre-compute and cache the teacher outputs, then feed the cache to the distiller to save teacher's forward pass time.

      See Feed Different batches to Student and Teacher, Feed Cached Values for details of the above features.

    • MultiTaskDistiller now supports intermediate feature matching loss.

    • Tensorboard now records more detailed losses (KD loss, hard label loss, matching losses...).

    See details in releases.

August 27, 2020

We are happy to announce that our model is on top of GLUE benchmark, check leaderboard.

Aug 24, 2020

  • Updated to 0.2.0.1:
    • fixed bugs in MultiTaskDistiller and training loops.

Jul 29, 2020

  • Updated to 0.2.0:
    • Added the support for distributed data-parallel training with DistributedDataParallel: TrainingConfig now accpects the local_rank argument. See the documentation of TrainingConfig for detail.
  • Added an example of distillation on the Chinese NER task to demonstrate distributed data-parallel training. See examples/msra_ner_example.

Jul 14, 2020

  • Updated to 0.1.10:
    • Now supports mixed precision training with Apex! Just set fp16 to True in TrainingConfig. See the documentation of TrainingConfig for detail.
    • Added data_parallel option in TrainingConfig to enable data parallel training and mixed precision training work together.

Apr 26, 2020

  • Added Chinese NER task (MSRA NER) results.
  • Added results for distilling to T12-nano model, which has a similar strcuture to Electra-small.
  • Updated some results of CoNLL-2003, CMRC 2018 and DRCD.

Apr 22, 2020

  • Updated to 0.1.9 (added cache option which speeds up distillation; fixed some bugs). See details in releases.
  • Added experimential results for distilling Electra-base to Electra-small on Chinese tasks.
  • TextBrewer has been accepted by ACL 2020 as a demo paper, please use our new bib entry.

Mar 17, 2020

Mar 11, 2020

  • Updated to 0.1.8 (Improvements on TrainingConfig and train method). See details in releases.

Mar 2, 2020

  • Initial public version 0.1.7 has been released. See details in releases.

Table of Contents

Section Contents
Introduction Introduction to TextBrewer
Installation How to install
Workflow Two stages of TextBrewer workflow
Quickstart Example: distilling BERT-base to a 3-layer BERT
Experiments Distillation experiments on typical English and Chinese datasets
Core Concepts Brief explanations of the core concepts in TextBrewer
FAQ Frequently asked questions
Known Issues Known issues
Citation Citation to TextBrewer
Follow Us -

Introduction

Textbrewer is designed for the knowledge distillation of NLP models. It provides various distillation methods and offers a distillation framework for quickly setting up experiments.

The main features of TextBrewer are:

  • Wide-support: it supports various model architectures (especially transformer-based models)
  • Flexibility: design your own distillation scheme by combining different techniques; it also supports user-defined loss functions, modules, etc.
  • Easy-to-use: users don't need to modify the model architectures
  • Built for NLP: it is suitable for a wide variety of NLP tasks: text classification, machine reading comprehension, sequence labeling, ...

TextBrewer currently is shipped with the following distillation techniques:

  • Mixed soft-label and hard-label training
  • Dynamic loss weight adjustment and temperature adjustment
  • Various distillation loss functions: hidden states MSE, attention-matrix-based loss, neuron selectivity transfer, ...
  • Freely adding intermediate features matching losses
  • Multi-teacher distillation
  • ...

TextBrewer includes:

  1. Distillers: the cores of distillation. Different distillers perform different distillation modes. There are GeneralDistiller, MultiTeacherDistiller, BasicTrainer, etc.
  2. Configurations and presets: Configuration classes for training and distillation, and predefined distillation loss functions and strategies.
  3. Utilities: auxiliary tools such as model parameters analysis.

To start distillation, users need to provide

  1. the models (the trained teacher model and the un-trained student model)
  2. datasets and experiment configurations

TextBrewer has achieved impressive results on several typical NLP tasks. See Experiments.

See Full Documentation for detailed usages.

Architecture

Installation

  • Requirements

    • Python >= 3.6
    • PyTorch >= 1.1.0
    • TensorboardX or Tensorboard
    • NumPy
    • tqdm
    • Transformers >= 2.0 (optional, used by some examples)
    • Apex == 0.1.0 (optional, mixed precision training)
  • Install from PyPI

    pip install textbrewer
  • Install from the Github source

    git clone https://github.com/airaria/TextBrewer.git
    pip install ./textbrewer

Workflow

  • Stage 1: Preparation:

    1. Train the teacher model
    2. Define and initialize the student model
    3. Construct a dataloader, an optimizer, and a learning rate scheduler
  • Stage 2: Distillation with TextBrewer:

    1. Construct a TraningConfig and a DistillationConfig, initialize a distiller
    2. Define an adaptor and a callback. The adaptor is used for adaptation of model inputs and outputs. The callback is called by the distiller during training
    3. Call the train method of the distiller

Quickstart

Here we show the usage of TextBrewer by distilling BERT-base to a 3-layer BERT.

Before distillation, we assume users have provided:

  • A trained teacher model teacher_model (BERT-base) and a to-be-trained student model student_model (3-layer BERT).
  • a dataloader of the dataset, an optimizer and a learning rate builder or class scheduler_class and its args dict scheduler_dict.

Distill with TextBrewer:

import textbrewer
from textbrewer import GeneralDistiller
from textbrewer import TrainingConfig, DistillationConfig

# Show the statistics of model parameters
print("\nteacher_model's parametrers:")
result, _ = textbrewer.utils.display_parameters(teacher_model,max_level=3)
print (result)

print("student_model's parametrers:")
result, _ = textbrewer.utils.display_parameters(student_model,max_level=3)
print (result)

# Define an adaptor for interpreting the model inputs and outputs
def simple_adaptor(batch, model_outputs):
      # The second and third elements of model outputs are the logits and hidden states
    return {'logits': model_outputs[1],
            'hidden': model_outputs[2]}

# Training configuration 
train_config = TrainingConfig()
# Distillation configuration
# Matching different layers of the student and the teacher
distill_config = DistillationConfig(
    intermediate_matches=[    
     {'layer_T':0, 'layer_S':0, 'feature':'hidden', 'loss': 'hidden_mse','weight' : 1},
     {'layer_T':8, 'layer_S':2, 'feature':'hidden', 'loss': 'hidden_mse','weight' : 1}])

# Build distiller
distiller = GeneralDistiller(
    train_config=train_config, distill_config = distill_config,
    model_T = teacher_model, model_S = student_model, 
    adaptor_T = simple_adaptor, adaptor_S = simple_adaptor)

# Start!
with distiller:
    distiller.train(optimizer, dataloader, num_epochs=1, scheduler_class=scheduler_class, scheduler_args = scheduler_args, callback=None)

Examples

Experiments

We have performed distillation experiments on several typical English and Chinese NLP datasets. The setups and configurations are listed below.

Models

We have tested different student models. To compare with public results, the student models are built with standard transformer blocks except for BiGRU which is a single-layer bidirectional GRU. The architectures are listed below. Note that the number of parameters includes the embedding layer but does not include the output layer of each specific task.

English models

Model #Layers Hidden size Feed-forward size #Params Relative size
BERT-base-cased (teacher) 12 768 3072 108M 100%
T6 (student) 6 768 3072 65M 60%
T3 (student) 3 768 3072 44M 41%
T3-small (student) 3 384 1536 17M 16%
T4-Tiny (student) 4 312 1200 14M 13%
T12-nano (student) 12 256 1024 17M 16%
BiGRU (student) - 768 - 31M 29%

Chinese models

Model #Layers Hidden size Feed-forward size #Params Relative size
RoBERTa-wwm-ext (teacher) 12 768 3072 102M 100%
Electra-base (teacher) 12 768 3072 102M 100%
T3 (student) 3 768 3072 38M 37%
T3-small (student) 3 384 1536 14M 14%
T4-Tiny (student) 4 312 1200 11M 11%
Electra-small (student) 12 256 1024 12M 12%

Distillation Configurations

distill_config = DistillationConfig(temperature = 8, intermediate_matches = matches)
# Others arguments take the default values

matches are differnt for different models:

Model matches
BiGRU None
T6 L6_hidden_mse + L6_hidden_smmd
T3 L3_hidden_mse + L3_hidden_smmd
T3-small L3n_hidden_mse + L3_hidden_smmd
T4-Tiny L4t_hidden_mse + L4_hidden_smmd
T12-nano small_hidden_mse + small_hidden_smmd
Electra-small small_hidden_mse + small_hidden_smmd

The definitions of matches are at examples/matches/matches.py.

We use GeneralDistiller in all the distillation experiments.

Training Configurations

  • Learning rate is 1e-4 (unless otherwise specified).
  • We train all the models for 30~60 epochs.

Results on English Datasets

We experiment on the following typical Enlgish datasets:

Dataset Task type Metrics #Train #Dev Note
MNLI text classification m/mm Acc 393K 20K sentence-pair 3-class classification
SQuAD 1.1 reading comprehension EM/F1 88K 11K span-extraction machine reading comprehension
CoNLL-2003 sequence labeling F1 23K 6K named entity recognition

We list the public results from DistilBERT, BERT-PKD, BERT-of-Theseus, TinyBERT and our results below for comparison.

Public results:

Model (public) MNLI SQuAD CoNLL-2003
DistilBERT (T6) 81.6 / 81.1 78.1 / 86.2 -
BERT6-PKD (T6) 81.5 / 81.0 77.1 / 85.3 -
BERT-of-Theseus (T6) 82.4/ 82.1 - -
BERT3-PKD (T3) 76.7 / 76.3 - -
TinyBERT (T4-tiny) 82.8 / 82.9 72.7 / 82.1 -

Our results:

Model (ours) MNLI SQuAD CoNLL-2003
BERT-base-cased (teacher) 83.7 / 84.0 81.5 / 88.6 91.1
BiGRU - - 85.3
T6 83.5 / 84.0 80.8 / 88.1 90.7
T3 81.8 / 82.7 76.4 / 84.9 87.5
T3-small 81.3 / 81.7 72.3 / 81.4 78.6
T4-tiny 82.0 / 82.6 75.2 / 84.0 89.1
T12-nano 83.2 / 83.9 79.0 / 86.6 89.6

Note:

  1. The equivalent model structures of public models are shown in the brackets after their names.
  2. When distilling to T4-tiny, NewsQA is used for data augmentation on SQuAD and HotpotQA is used for data augmentation on CoNLL-2003.
  3. When distilling to T12-nano, HotpotQA is used for data augmentation on CoNLL-2003.

Results on Chinese Datasets

We experiment on the following typical Chinese datasets:

Dataset Task type Metrics #Train #Dev Note
XNLI text classification Acc 393K 2.5K Chinese translation version of MNLI
LCQMC text classification Acc 239K 8.8K sentence-pair matching, binary classification
CMRC 2018 reading comprehension EM/F1 10K 3.4K span-extraction machine reading comprehension
DRCD reading comprehension EM/F1 27K 3.5K span-extraction machine reading comprehension (Traditional Chinese)
MSRA NER sequence labeling F1 45K 3.4K (#Test) Chinese named entity recognition

The results are listed below.

Model XNLI LCQMC CMRC 2018 DRCD
RoBERTa-wwm-ext (teacher) 79.9 89.4 68.8 / 86.4 86.5 / 92.5
T3 78.4 89.0 66.4 / 84.2 78.2 / 86.4
T3-small 76.0 88.1 58.0 / 79.3 75.8 / 84.8
T4-tiny 76.2 88.4 61.8 / 81.8 77.3 / 86.1
Model XNLI LCQMC CMRC 2018 DRCD MSRA NER
Electra-base (teacher)) 77.8 89.8 65.6 / 84.7 86.9 / 92.3 95.14
Electra-small 77.7 89.3 66.5 / 84.9 85.5 / 91.3 93.48

Note:

  1. Learning rate decay is not used in distillation on CMRC 2018 and DRCD.
  2. CMRC 2018 and DRCD take each other as the augmentation dataset in the distillation.
  3. The settings of training Electra-base teacher model can be found at Chinese-ELECTRA.
  4. Electra-small student model is intialized with the pretrained weights.

Core Concepts

Configurations

  • TrainingConfig: configuration related to general deep learning model training
  • DistillationConfig: configuration related to distillation methods

Distillers

Distillers are in charge of conducting the actual experiments. The following distillers are available:

  • BasicDistiller: single-teacher single-task distillation, provides basic distillation strategies.
  • GeneralDistiller (Recommended): single-teacher single-task distillation, supports intermediate features matching. Recommended most of the time.
  • MultiTeacherDistiller: multi-teacher distillation, which distills multiple teacher models (of the same task) into a single student model. This class doesn't support Intermediate features matching.
  • MultiTaskDistiller: multi-task distillation, which distills multiple teacher models (of different tasks) into a single student.
  • BasicTrainer: Supervised training a single model on a labeled dataset, not for distillation. It can be used to train a teacher model.

User-Defined Functions

In TextBrewer, there are two functions that should be implemented by users: callback and adaptor.

Callback

At each checkpoint, after saving the student model, the callback function will be called by the distiller. A callback can be used to evaluate the performance of the student model at each checkpoint.

Adaptor

It converts the model inputs and outputs to the specified format so that they could be recognized by the distiller, and distillation losses can be computed. At each training step, batch and model outputs will be passed to the adaptor; the adaptor re-organizes the data and returns a dictionary.

For more details, see the explanations in Full Documentation.

FAQ

Q: How to initialize the student model?

A: The student model could be randomly initialized (i.e., with no prior knowledge) or be initialized by pre-trained weights. For example, when distilling a BERT-base model to a 3-layer BERT, you could initialize the student model with RBT3 (for Chinese tasks) or the first three layers of BERT (for English tasks) to avoid cold start problem. We recommend that users use pre-trained student models whenever possible to fully take advantage of large-scale pre-training.

Q: How to set training hyperparameters for the distillation experiments?

A: Knowledge distillation usually requires more training epochs and larger learning rate than training on the labeled dataset. For example, training SQuAD on BERT-base usually takes 3 epochs with lr=3e-5; however, distillation takes 30~50 epochs with lr=1e-4. The conclusions are based on our experiments, and you are advised to try on your own data.

Q: My teacher model and student model take different inputs (they do not share vocabularies), so how can I distill?

A: You need to feed different batches to the teacher and the student. See the section Feed Different batches to Student and Teacher, Feed Cached Values in the full documentation.

Q: I have stored the logits from my teacher model. Can I use them in the distillation to save the forward pass time?

A: Yes, see the section Feed Different batches to Student and Teacher, Feed Cached Values in the full documentation.

Known Issues

  • Multi-GPU training support is only available through DataParallel currently.
  • Multi-label classification is not supported.

Citation

If you find TextBrewer is helpful, please cite our paper:

@InProceedings{textbrewer-acl2020-demo,
    title = "{T}ext{B}rewer: {A}n {O}pen-{S}ource {K}nowledge {D}istillation {T}oolkit for {N}atural {L}anguage {P}rocessing",
    author = "Yang, Ziqing and Cui, Yiming and Chen, Zhipeng and Che, Wanxiang and Liu, Ting and Wang, Shijin and Hu, Guoping",
    booktitle = "Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics: System Demonstrations",
    year = "2020",
    publisher = "Association for Computational Linguistics",
    url = "https://www.aclweb.org/anthology/2020.acl-demos.2",
    pages = "9--16",
}

Follow Us

Follow our official WeChat account to keep updated with our latest technologies!