The implementation of DeBERTa
Перейти к файлу
Pengcheng He 4d7fe0bd4f
Update README.md
2023-03-25 03:22:59 -07:00
DeBERTa 1. Add code for DeBERTaV3 pre-training; 2. Fix error in torch 1.11; 3. Add code for ONNX export 2023-03-18 23:14:19 -07:00
docker 1. DeBERTa v2 2021-02-09 10:50:16 -05:00
docs Add pytorch back to required package 2020-07-21 01:21:45 -04:00
experiments Update README.md 2023-03-25 03:22:59 -07:00
.gitignore Initial commit 2020-06-08 15:57:16 +00:00
CODE_OF_CONDUCT.md Initial CODE_OF_CONDUCT.md commit 2020-06-08 08:57:19 -07:00
LICENSE Initial LICENSE commit 2020-06-08 08:57:22 -07:00
README.md Update README.md 2023-03-25 03:19:10 -07:00
SECURITY.md Initial SECURITY.md commit 2020-06-08 08:57:21 -07:00
VERSION Update README to point to the new model hub page of DeBERTa-V2 2021-02-16 13:25:52 -05:00
requirements.txt 1. Add document for DeBERTa pre-training 2021-12-03 10:56:01 -08:00
run_docker.sh First release of DeBERTa 2020-06-15 18:42:48 -04:00
setup.cfg First release of DeBERTa 2020-06-15 18:42:48 -04:00
setup.py Update setup.py 2021-11-23 13:39:27 -08:00

README.md

DeBERTa: Decoding-enhanced BERT with Disentangled Attention

This repository is the official implementation of DeBERTa: Decoding-enhanced BERT with Disentangled Attention and DeBERTa V3: Improving DeBERTa using ELECTRA-Style Pre-Training with Gradient-Disentangled Embedding Sharing

News

03/18/2023

  • DeBERTaV3 paper is accepted by ICLR 2023.
  • The code for DeBERTaV3 pre-training and continous training is added. Please check Language Model for details.

12/8/2021

  • DeBERTa-V3-XSmall is added. With only 22M backbone parameters which is only 1/4 of RoBERTa-Base and XLNet-Base, DeBERTa-V3-XSmall significantly outperforms the later on MNLI and SQuAD v2.0 tasks (i.e. 1.2% on MNLI-m, 1.5% EM score on SQuAD v2.0). This further demonstrates the efficiency of DeBERTaV3 models.

11/16/2021

3/31/2021

  • Masked language model task is added
  • SuperGLUE tasks is added
  • SiFT code is added

2/03/2021

DeBERTa v2 code and the 900M, 1.5B model are here now. This includes the 1.5B model used for our SuperGLUE single-model submission and achieving 89.9, versus human baseline 89.8. You can find more details about this submission in our blog

What's new in v2

  • Vocabulary In v2 we use a new vocabulary of size 128K built from the training data. Instead of GPT2 tokenizer, we use sentencepiece tokenizer.
  • nGiE(nGram Induced Input Encoding) In v2 we use an additional convolution layer aside with the first transformer layer to better learn the local dependency of input tokens. We will add more ablation studies on this feature.
  • Sharing position projection matrix with content projection matrix in attention layer Based on our previous experiment, we found this can save parameters without affecting performance.
  • Apply bucket to encode relative positions In v2 we use log bucket to encode relative positions similar to T5.
  • 900M model & 1.5B model In v2 we scale our model size to 900M and 1.5B which significantly improves the performance of downstream tasks.

12/29/2020

With DeBERTa 1.5B model, we surpass T5 11B model and human performance on SuperGLUE leaderboard. Code and model will be released soon. Please check out our paper for more details.

06/13/2020

We released the pre-trained models, source code, and fine-tuning scripts to reproduce some of the experimental results in the paper. You can follow similar scripts to apply DeBERTa to your own experiments or applications. Pre-training scripts will be released in the next step.

Introduction to DeBERTa

DeBERTa (Decoding-enhanced BERT with disentangled attention) improves the BERT and RoBERTa models using two novel techniques. The first is the disentangled attention mechanism, where each word is represented using two vectors that encode its content and position, respectively, and the attention weights among words are computed using disentangled matrices on their contents and relative positions. Second, an enhanced mask decoder is used to replace the output softmax layer to predict the masked tokens for model pretraining. We show that these two techniques significantly improve the efficiency of model pre-training and performance of downstream tasks.

Pre-trained Models

Our pre-trained models are packaged into zipped files. You can download them from our releases, or download an individual model via the links below:

Model Vocabulary(K) Backbone Parameters(M) Hidden Size Layers Note
V2-XXLarge1 128 1320 1536 48 128K new SPM vocab
V2-XLarge 128 710 1536 24 128K new SPM vocab
XLarge 50 700 1024 48 Same vocab as RoBERTa
Large 50 350 1024 24 Same vocab as RoBERTa
Base 50 100 768 12 Same vocab as RoBERTa
V2-XXLarge-MNLI 128 1320 1536 48 Fine-turned with MNLI
V2-XLarge-MNLI 128 710 1536 24 Fine-turned with MNLI
XLarge-MNLI 50 700 1024 48 Fine-turned with MNLI
Large-MNLI 50 350 1024 24 Fine-turned with MNLI
Base-MNLI 50 86 768 12 Fine-turned with MNLI
DeBERTa-V3-Large2 128 304 1024 24 128K new SPM vocab
DeBERTa-V3-Base2 128 86 768 12 128K new SPM vocab
DeBERTa-V3-Small2 128 44 768 6 128K new SPM vocab
DeBERTa-V3-XSmall2 128 22 384 12 128K new SPM vocab
mDeBERTa-V3-Base2 250 86 768 12 250K new SPM vocab, multi-lingual model with 102 languages

Note

  • 1 This is the model(89.9) that surpassed T5 11B(89.3) and human performance(89.8) on SuperGLUE for the first time. 128K new SPM vocab.
  • 2 These V3 DeBERTa models are deberta models pre-trained with ELECTRA-style objective plus gradient-disentangled embedding sharing which significantly improves the model efficiency.

Try the model

Read our documentation

Requirements

  • Linux system, e.g. Ubuntu 18.04LTS
  • CUDA 10.0
  • pytorch 1.3.0
  • python 3.6
  • bash shell 4.0
  • curl
  • docker (optional)
  • nvidia-docker2 (optional)

There are several ways to try our code,

Use docker

Docker is the recommended way to run the code as we already built every dependency into our docker bagai/deberta and you can follow the docker official site to install docker on your machine.

To run with docker, make sure your system fulfills the requirements in the above list. Here are the steps to try the GLUE experiments: Pull the code, run ./run_docker.sh , and then you can run the bash commands under /DeBERTa/experiments/glue/

Use pip

Pull the code and run pip3 install -r requirements.txt in the root directory of the code, then enter experiments/glue/ folder of the code and try the bash commands under that folder for glue experiments.

Install as a pip package

pip install deberta

Use DeBERTa in existing code


# To apply DeBERTa to your existing code, you need to make two changes to your code,
# 1. change your model to consume DeBERTa as the encoder
from DeBERTa import deberta
import torch
class MyModel(torch.nn.Module):
  def __init__(self):
    super().__init__()
    # Your existing model code
    self.deberta = deberta.DeBERTa(pre_trained='base') # Or 'large' 'base-mnli' 'large-mnli' 'xlarge' 'xlarge-mnli' 'xlarge-v2' 'xxlarge-v2'
    # Your existing model code
    # do inilization as before
    # 
    self.deberta.apply_state() # Apply the pre-trained model of DeBERTa at the end of the constructor
    #
  def forward(self, input_ids):
    # The inputs to DeBERTa forward are
    # `input_ids`: a torch.LongTensor of shape [batch_size, sequence_length] with the word token indices in the vocabulary
    # `token_type_ids`: an optional torch.LongTensor of shape [batch_size, sequence_length] with the token types indices selected in [0, 1]. 
    #    Type 0 corresponds to a `sentence A` and type 1 corresponds to a `sentence B` token (see BERT paper for more details).
    # `attention_mask`: an optional parameter for input mask or attention mask. 
    #   - If it's an input mask, then it will be torch.LongTensor of shape [batch_size, sequence_length] with indices selected in [0, 1]. 
    #      It's a mask to be used if the input sequence length is smaller than the max input sequence length in the current batch. 
    #      It's the mask that we typically use for attention when a batch has varying length sentences.
    #   - If it's an attention mask then if will be torch.LongTensor of shape [batch_size, sequence_length, sequence_length]. 
    #      In this case, it's a mask indicating which tokens in the sequence should be attended by other tokens in the sequence. 
    # `output_all_encoded_layers`: whether to output results of all encoder layers, default, True
    encoding = deberta.bert(input_ids)[-1]

# 2. Change your tokenizer with the tokenizer built-in DeBERta
from DeBERTa import deberta
vocab_path, vocab_type = deberta.load_vocab(pretrained_id='base')
tokenizer = deberta.tokenizers[vocab_type](vocab_path)
# We apply the same schema of special tokens as BERT, e.g. [CLS], [SEP], [MASK]
max_seq_len = 512
tokens = tokenizer.tokenize('Examples input text of DeBERTa')
# Truncate long sequence
tokens = tokens[:max_seq_len -2]
# Add special tokens to the `tokens`
tokens = ['[CLS]'] + tokens + ['[SEP]']
input_ids = tokenizer.convert_tokens_to_ids(tokens)
input_mask = [1]*len(input_ids)
# padding
paddings = max_seq_len-len(input_ids)
input_ids = input_ids + [0]*paddings
input_mask = input_mask + [0]*paddings
features = {
'input_ids': torch.tensor(input_ids, dtype=torch.int),
'input_mask': torch.tensor(input_mask, dtype=torch.int)
}

Run DeBERTa experiments from command line

For glue tasks,

  1. Get the data
cache_dir=/tmp/DeBERTa/
cd experiments/glue
./download_data.sh  $cache_dir/glue_tasks
  1. Run task
task=STS-B 
OUTPUT=/tmp/DeBERTa/exps/$task
export OMP_NUM_THREADS=1
python3 -m DeBERTa.apps.run --task_name $task --do_train  \
  --data_dir $cache_dir/glue_tasks/$task \
  --eval_batch_size 128 \
  --predict_batch_size 128 \
  --output_dir $OUTPUT \
  --scale_steps 250 \
  --loss_scale 16384 \
  --accumulative_update 1 \  
  --num_train_epochs 6 \
  --warmup 100 \
  --learning_rate 2e-5 \
  --train_batch_size 32 \
  --max_seq_len 128

Notes

    1. By default we will cache the pre-trained model and tokenizer at $HOME/.~DeBERTa, you may need to clean it if the downloading failed unexpectedly.
    1. You can also try our models with HF Transformers. But when you try XXLarge model you need to specify --sharded_ddp argument. Please check our XXLarge model card for more details.

Experiments

Our fine-tuning experiments are carried on half a DGX-2 node with 8x32 V100 GPU cards, the results may vary due to different GPU models, drivers, CUDA SDK versions, using FP16 or FP32, and random seeds. We report our numbers based on multiple runs with different random seeds here. Here are the results from the Large model:

Task Command Results Running Time(8x32G V100 GPUs)
MNLI xxlarge v2 experiments/glue/mnli.sh xxlarge-v2 91.7/91.9 +/-0.1 4h
MNLI xlarge v2 experiments/glue/mnli.sh xlarge-v2 91.7/91.6 +/-0.1 2.5h
MNLI xlarge experiments/glue/mnli.sh xlarge 91.5/91.2 +/-0.1 2.5h
MNLI large experiments/glue/mnli.sh large 91.3/91.1 +/-0.1 2.5h
QQP large experiments/glue/qqp.sh large 92.3 +/-0.1 6h
QNLI large experiments/glue/qnli.sh large 95.3 +/-0.2 2h
MRPC large experiments/glue/mrpc.sh large 91.9 +/-0.5 0.5h
RTE large experiments/glue/rte.sh large 86.6 +/-1.0 0.5h
SST-2 large experiments/glue/sst2.sh large 96.7 +/-0.3 1h
STS-b large experiments/glue/Stsb.sh large 92.5 +/-0.3 0.5h
CoLA large experiments/glue/cola.sh 70.5 +/-1.0 0.5h

And here are the results from the Base model

Task Command Results Running Time(8x32G V100 GPUs)
MNLI base experiments/glue/mnli.sh base 88.8/88.5 +/-0.2 1.5h

Fine-tuning on NLU tasks

We present the dev results on SQuAD 1.1/2.0 and several GLUE benchmark tasks.

Model SQuAD 1.1 SQuAD 2.0 MNLI-m/mm SST-2 QNLI CoLA RTE MRPC QQP STS-B
F1/EM F1/EM Acc Acc Acc MCC Acc Acc/F1 Acc/F1 P/S
BERT-Large 90.9/84.1 81.8/79.0 86.6/- 93.2 92.3 60.6 70.4 88.0/- 91.3/- 90.0/-
RoBERTa-Large 94.6/88.9 89.4/86.5 90.2/- 96.4 93.9 68.0 86.6 90.9/- 92.2/- 92.4/-
XLNet-Large 95.1/89.7 90.6/87.9 90.8/- 97.0 94.9 69.0 85.9 90.8/- 92.3/- 92.5/-
DeBERTa-Large1 95.5/90.1 90.7/88.0 91.3/91.1 96.5 95.3 69.5 91.0 92.6/94.6 92.3/- 92.8/92.5
DeBERTa-XLarge1 -/- -/- 91.5/91.2 97.0 - - 93.1 92.1/94.3 - 92.9/92.7
DeBERTa-V2-XLarge1 95.8/90.8 91.4/88.9 91.7/91.6 97.5 95.8 71.1 93.9 92.0/94.2 92.3/89.8 92.9/92.9
DeBERTa-V2-XXLarge1,2 96.1/91.4 92.2/89.7 91.7/91.9 97.2 96.0 72.0 93.5 93.1/94.9 92.7/90.3 93.2/93.1
DeBERTa-V3-Large -/- 91.5/89.0 91.8/91.9 96.9 96.0 75.3 92.7 92.2/- 93.0/- 93.0/-
DeBERTa-V3-Base -/- 88.4/85.4 90.6/90.7 - - - - - - -
DeBERTa-V3-Small -/- 82.9/80.4 88.3/87.7 - - - - - - -
DeBERTa-V3-XSmall -/- 84.8/82.0 88.1/88.3 - - - - - - -

Fine-tuning on XNLI

We present the dev results on XNLI with zero-shot crosslingual transfer setting, i.e. training with english data only, test on other languages.

Model avg en fr es de el bg ru tr ar vi th zh hi sw ur
XLM-R-base 76.2 85.8 79.7 80.7 78.7 77.5 79.6 78.1 74.2 73.8 76.5 74.6 76.7 72.4 66.5 68.3
mDeBERTa-V3-Base 79.8+/-0.2 88.2 82.6 84.4 82.7 82.3 82.4 80.8 79.5 78.5 78.1 76.4 79.5 75.9 73.9 72.4

Notes.

Pre-training with MLM and RTD objectives

To pre-train DeBERTa with MLM and RTD objectives, please check experiments/language_models

Contacts

Pengcheng He(penhe@microsoft.com), Xiaodong Liu(xiaodl@microsoft.com), Jianfeng Gao(jfgao@microsoft.com), Weizhu Chen(wzchen@microsoft.com)

Citation

@misc{he2021debertav3,
      title={DeBERTaV3: Improving DeBERTa using ELECTRA-Style Pre-Training with Gradient-Disentangled Embedding Sharing}, 
      author={Pengcheng He and Jianfeng Gao and Weizhu Chen},
      year={2021},
      eprint={2111.09543},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}
@inproceedings{
he2021deberta,
title={DEBERTA: DECODING-ENHANCED BERT WITH DISENTANGLED ATTENTION},
author={Pengcheng He and Xiaodong Liu and Jianfeng Gao and Weizhu Chen},
booktitle={International Conference on Learning Representations},
year={2021},
url={https://openreview.net/forum?id=XPZIaotutsD}
}