This repository is the official implementation of DeBERTa: Decoding-enhanced BERT with Disentangled Attention
- Masked language model task is added
- SuperGLUE tasks is added
- SiFT code is added
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
- 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.
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.
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.
- Add SuperGLUE tasks
- Add SiFT code
- Add Pretraining code
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.
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 |
- 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.
Read our documentation
- 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,
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/
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.
pip install deberta
# 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)
}For glue tasks,
- Get the data
cache_dir=/tmp/DeBERTa/
cd experiments/glue
./download_data.sh $cache_dir/glue_tasks- Run task
task=STS-B
OUTPUT=/tmp/DeBERTa/exps/$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-
- 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.
- By default we will cache the pre-trained model and tokenizer at
-
- 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.
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 |
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 |
- 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.
Pengcheng He(penhe@microsoft.com), Xiaodong Liu(xiaodl@microsoft.com), Jianfeng Gao(jfgao@microsoft.com), Weizhu Chen(wzchen@microsoft.com)
@misc{he2020deberta,
title={DeBERTa: Decoding-enhanced BERT with Disentangled Attention},
author={Pengcheng He and Xiaodong Liu and Jianfeng Gao and Weizhu Chen},
year={2020},
eprint={2006.03654},
archivePrefix={arXiv},
primaryClass={cs.CL}
}
This project welcomes contributions and suggestions. Most contributions require you to agree to a Contributor License Agreement (CLA) declaring that you have the right to, and actually do, grant us the rights to use your contribution. For details, visit https://cla.opensource.microsoft.com.
When you submit a pull request, a CLA bot will automatically determine whether you need to provide a CLA and decorate the PR appropriately (e.g., status check, comment). Simply follow the instructions provided by the bot. You will only need to do this once across all repos using our CLA.
This project has adopted the Microsoft Open Source Code of Conduct. For more information see the Code of Conduct FAQ or contact opencode@microsoft.com with any additional questions or comments.