DeBERTa: Decoding-enhanced BERT with Disentangled Attention

This repository is the official implementation of DeBERTa: Decoding-enhanced BERT with Disentangled Attention

News

11/16/2021

• The models of our new work DeBERTa V3: Improving DeBERTa using ELECTRA-Style Pre-Training with Gradient-Disentangled Embedding Sharing are publicly available at huggingface model hub now. The new models are based on DeBERTa-V2 models by replacing MLM with ELECTRA-style objective plus gradient-disentangled embedding sharing which further improves the model efficiency.
• Scripts for DeBERTa V3 model fine-tuning are added

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 the performance.
• Apply bucket to encode relative postions 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	Parameters	Hidden Size	Layers	Note
V2-XXLarge1	1.5B	1536	48	128K new SPM vocab
V2-XLarge	900M	1536	24	128K new SPM vocab
XLarge	750M	1024	48	Same vocab as RoBERTa
Large	400M	1024	24	Same vocab as RoBERTa
Base	140M	768	12	Same vocab as RoBERTa
V2-XXLarge-MNLI	1.5B	1536	48	Fine-turned with MNLI
V2-XLarge-MNLI	900M	1536	24	Fine-turned with MNLI
XLarge-MNLI	750M	1024	48	Fine-turned with MNLI
Large-MNLI	400M	1024	24	Fine-turned with MNLI
Base-MNLI	140M	768	12	Fine-turned with MNLI
DeBERTa-V3-Large2	418M	1024	24	128K new SPM vocab
DeBERTa-V3-Base2	183M	768	12	128K new SPM vocab
DeBERTa-V3-Small2	143M	768	6	128K new SPM vocab
mDeBERTa-V3-Base2	280M	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 the 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 fullfil 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 into your existing code, you need to make two changes on 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 indicate 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 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

2. 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.
• 2. 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 multple 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	88.2/87.9	-	-	-	-	-	-	-

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.

• ¹ Following RoBERTa, for RTE, MRPC, STS-B, we fine-tune the tasks based on DeBERTa-Large-MNLI, DeBERTa-XLarge-MNLI, DeBERTa-V2-XLarge-MNLI, DeBERTa-V2-XXLarge-MNLI. The results of SST-2/QQP/QNLI/SQuADv2 will also be slightly improved when start from MNLI fine-tuned models, however, we only report the numbers fine-tuned from pretrained base models for those 4 tasks.

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}
}