Skip to content

nosaydomore/MobileBert_paddle

Repository files navigation

MobileBert_paddle

这是一个mobileBert的基于paddle深度学习框架的复现

MobileBERT: a Compact Task-Agnostic BERT for Resource-Limited Devices. [https://arxiv.org/pdf/2004.02984.pdf]

原论文效果 glue squad

快速开始

环境要求:

python==3.7

paddlepaddle==2.1.2
pytorch==1.7.1
numpy>=1.7
paddlenlp==2.0.7
visualdl==2.2.0
transformers==4.9.2

#克隆本仓库
git clone https://github.com/nosaydomore/MobileBert_paddle.git
#进入项目目录
cd MobileBert_paddle

复现环境使用的是baidu的AIStudio至尊版(V100 32G)

AIStudio项目地址:https://aistudio.baidu.com/aistudio/projectdetail/2367246 , AiStudio运行更方便. 强烈建议在AIStudio的终端上进行结果复现(版本选择:MobileBert_PaddleV3)。

1. 模型转换

提供下载好的pytorch权重以及转换完成的paddle权重

将链接中weight里的文件放在项目中weight的目录下即可

链接: https://pan.baidu.com/s/1o35WuWIUPNWQYyc43dj_pQ 提取码: 92fk

python convert.py
'''
transpose paramter of key:mobilebert.embeddings.embedding_transformation.weight
transpose paramter of key:mobilebert.encoder.layer.0.attention.self.query.weight
transpose paramter of key:mobilebert.encoder.layer.0.attention.self.key.weight
...
...
...
...
transpose paramter of key:mobilebert.encoder.layer.23.ffn.2.output.dense.weight
transpose parameter 361
finnish convert pytorch to paddle.
model_saved: ./weight/paddle/model_state.pdparams
'''

2. 模型精度对齐

python compare.py
'''
sent_feat_pd: [1, 11, 512]
sent_feat_pt: torch.Size([1, 11, 512])
cls_feat_pd: [512]
cls_feat_pt: torch.Size([512])

cls token feat:
diff_mean: tensor(1.7400)
diff_max: tensor(12.)

other token feat:
diff_mean: tensor(2.8562e-06)
diff_max: tensor(3.0518e-05)
'''

复现到这一步,发现cls token的feat误差较大,主要原因是数值较大导致,只比较数值的话差异还可以。 此外比较了其他 token的误差,在可接受的范围内。 分析原因: 可能是模型为了减少延迟,使用no_norm替代了layer_norm,使得各个token之间在数值量级上少了约束,导致cls与其他差异较大,这里认为模型前向对齐了

此外在复现过程中paddlenlp的tokenizer与hugginface没有对齐,这里也做了对齐,运行python compare_tokenizer.py

模型训练与评估

模型在SQuADV1,SQuADV2,MNLI上进行训练微调 与论文中report的结果对比如下

数据集 原论文 复现结果
SQuADV1 82.9/90.0(em/f1) 83.0842 / 90.1204
SQuADV2 76.2/79.2(em/f1) 76.6192/79.7484
MNLI 83.3/82.6(m/mm) 0.83423/0.83563

评估

提供已经训练好的模型权重,以供评估(模型训练过程中的log也保存在里面)

链接: https://pan.baidu.com/s/1Uga9Wwx9cN8CdqD5G4-bFQ 提取码: jftn

将链接中task文件夹下的文件放在项目的task目录下(直接覆盖)

在SQuADV1上评估:bash scripts/eval_squadv1.sh

{
    "exact": 83.08420056764427,
    "f1": 90.12042552144304,
    "total": 10570
    "HasAns_exact": 83.08420056764427
    "HasAns_f1": 90.12042552144304
    "HasAns_total": 10570
}

在SQuADV2上评估:bash scripts/eval_squadv2.sh

10320/12354
{
  "exact": 76.61922007917123,
  "f1": 79.74841781219972,
  "total": 11873,
  "HasAns_exact": 71.54183535762483,
  "HasAns_f1": 77.80920456886743,
  "HasAns_total": 5928,
  "NoAns_exact": 81.68208578637511,
  "NoAns_f1": 81.68208578637511,
  "NoAns_total": 5945,
  "best_exact": 76.83820432915017,
  "best_exact_thresh": -2.503580152988434,
  "best_f1": 79.85988742758532,
  "best_f1_thresh": -1.783524513244629
}

在MNLI上评估:bash scripts/eval_mnli.sh

m_acc:
 eval loss: 0.316141, acc: 0.8342333163525216 
mm_acc:
 eval loss: 0.475563, acc: 0.8356387306753458

训练

模型 边训练边在dev集上测试

除了SQuADV1以外,其他数据集在训练总进度70%左右时会慢慢达到论文中的精度,具体收敛趋势可参考task目录下的各数据集的模型训练log run.log

SQuADV1: bash scripts/train_squadv1.sh

模型与log保存在output_squadv1 目录下

模型在SQuADV1上有极少概率会出现最好结果差 0.1(em:82.8), 若出现这一情况建议重新再跑1次 \ :p

SQuADV2: bash scripts/train_squadv2.sh

模型与log保存在output_squadv2 目录下

MNLI: bash scripts/train_mnli.sh

模型与log保存在output_mnli 目录下

Reference

项目在部分代码参考:JunnYu/paddle-mpnet

@inproceedings{2020MobileBERT,
  title={MobileBERT: a Compact Task-Agnostic BERT for Resource-Limited Devices},
  author={ Sun, Z.  and  Yu, H.  and  Song, X.  and  Liu, R.  and  Yang, Y.  and  Zhou, D. },
  booktitle={Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics},
  year={2020},
}

https://aistudio.baidu.com/aistudio/projectdetail/2367246

About

No description, website, or topics provided.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published