README.md

June 4, 2020 · View on GitHub

BERT下游任务finetune列表

finetune基于官方代码改造的模型基于pytorch/tensorflow双版本

*** 2019-10-24: 增加ERNIE1.0, google-bert-base, bert_wwm_ext_base部分结果, xlnet代码和相关结果 ***

*** 2019-10-17: 增加tensorflow多gpu并行 ***

*** 2019-10-16: 增加albert_xlarge结果 ***

*** 2019-10-15: 增加tensorflow(bert/roberta)在cmrc2018上的finetune代码(暂仅支持单卡) ***

*** 2019-10-14: 新增DRCD test结果 ***

*** 2019-10-12: pytorch支持albert ***

*** 2019-12-9: 新增cmrc2019 finetune google版albert, 新增CHID finetune代码***

*** 2019-12-22: 新增c3 finetune代码和CHID, c3的部分结果***

*** 2020-6-4: 新增pytorch转tf，tf转pb，以及pb测试demo***

模型及相关代码来源

  1. 官方Bert (https://github.com/google-research/bert)

  2. transformers (https://github.com/huggingface/transformers)

  3. 哈工大讯飞预训练 (https://github.com/ymcui/Chinese-BERT-wwm)

  4. brightmart预训练 (https://github.com/brightmart/roberta_zh)

  5. 自己瞎折腾的siBert (https://github.com/ewrfcas/SiBert_tensorflow)

关于pytorch的FP16

FP16的训练可以显著降低显存压力(如果有V100等GPU资源还能提高速度)。但是最新版编译的apex-FP16对并行的支持并不友好(https://github.com/NVIDIA/apex/issues/227)
实践下来bert相关任务的finetune任务对fp16的数值压力是比较小的，因此可以更多的以计算精度换取效率，所以我还是倾向于使用老版的FusedAdam+FP16_Optimizer的组合。
由于最新的apex已经舍弃这2个方法了，需要在编译apex的时候额外加入命令--deprecated_fused_adam

pip install -v --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext"  --global-option="--deprecated_fused_adam" ./

关于tensorflow的blocksparse

blocksparse(https://github.com/openai/blocksparse)
可以在tensorflow1.13版本直接pip安装，否则可以自己clone后编译。
其中fast_gelu以及self-attention中的softmax能够极大缓解显存压力。另外部分dropout位置我有所调整，整体显存占用下降大约30%~40%。

modellengthbatchmemory
roberta_base_fp165123216GB
roberta_large_fp165121216GB

参与任务

  1. CMRC 2018：篇章片段抽取型阅读理解（简体中文，只测了dev）

  2. DRCD：篇章片段抽取型阅读理解（繁体中文，转简体, 只测了dev）

  3. CJRC: 法律阅读理解（简体中文, 只有训练集，统一90%训练，10%测试）

  4. CHID: 多选成语阅读理解

  5. C3: 多选中文阅读理解

评测标准

验证集一般会调整learning_rate，warmup_rate，train_epoch等参数，选择最优的参数用五个不同的随机种子测试5次取平均和括号内最大值。测试集会直接用最佳的验证集模型进行验证。

模型介绍

L(transformer layers), H(hidden size), A(attention head numbers), E(embedding size)

特别注意brightmart roberta_large所支持的max_len只有256

modelsconfig
google_bert_baseL=12, H=768, A=12, max_len=512
siBert_baseL=12, H=768, A=12, max_len=512
siALBert_middleL=16, H=1024, E=128, A=16, max_len=512
哈工大讯飞 bert_wwm_ext_baseL=12, H=768, A=12, max_len=512
哈工大讯飞 roberta_wwm_ext_baseL=12, H=768, A=12, max_len=512
哈工大讯飞 roberta_wwm_ext_largeL=24, H=1024, A=16, max_len=512
ERNIE1.0L=12, H=768, A=12, max_len=512
xlnet-midL=24, H=768, A=12, max_len=512
brightmart roberta_middleL=24, H=768, A=12, max_len=512
brightmart roberta_largeL=24, H=1024, A=16, max_len=256
brightmart albert_largeL=24, H=1024, E=128, A=16, max_len=512
brightmart albert_xlargeL=24, H=2048, E=128, A=32, max_len=512
google albert_xxlargeL=12, H=4096, E=128, A=16, max_len=512

结果

参数

未列出均为epoch2, batch=32, lr=3e-5, warmup=0.1

modelscmrc2018DRCDCJRC
哈工大讯飞 roberta_wwm_ext_baseepoch2, batch=32, lr=3e-5, warmup=0.1同左同左
哈工大讯飞 roberta_wwm_ext_largeepoch2, batch=12, lr=2e-5, warmup=0.1epoch2, batch=32, lr=2.5e-5, warmup=0.1-
brightmart roberta_middleepoch2, batch=32, lr=3e-5, warmup=0.1同左同左
brightmart roberta_largeepoch2, batch=32, lr=3e-5, warmup=0.1同左同左
brightmart albert_largeepoch3, batch=32, lr=2e-5, warmup=0.05epoch3, batch=32, lr=2e-5, warmup=0.05epoch2, batch=32, lr=3e-5, warmup=0.1
brightmart albert_xlargeepoch3, batch=32, lr=2e-5, warmup=0.1epoch3, batch=32, lr=2.5e-5, warmup=0.06epoch2, batch=32, lr=2.5e-5, warmup=0.05

cmrc2018(阅读理解)

modelssettingDEV
哈工大讯飞 roberta_wwm_ext_largetf单卡finetune batch=12F1:89.415(89.724) EM:70.593(71.358)
modelsDEV
google_bert_baseF1:85.476(85.682) EM:64.765(65.921)
sibert_baseF1:87.521(88.628) EM:67.381(69.152)
sialbert_middleF1:87.6956(87.878) EM:67.897(68.624)
哈工大讯飞 bert_wwm_ext_baseF1:86.679(87.473) EM:66.959(69.09)
哈工大讯飞 roberta_wwm_ext_baseF1:87.521(88.628) EM:67.381(69.152)
哈工大讯飞 roberta_wwm_ext_largeF1:89.415(89.724) EM:70.593(71.358)
ERNIE1.0F1:87.300(87.733) EM:66.890(68.251)
xlnet-midF1:85.625(86.076) EM:65.312(66.076)
brightmart roberta_middleF1:86.841(87.242) EM:67.195(68.313)
brightmart roberta_largeF1:88.608(89.431) EM:69.935(72.538)
brightmart albert_largeF1:87.860(88.43) EM:67.754(69.028)
brightmart albert_xlargeF1:88.657(89.426) EM:68.897(70.643)

DRCD(阅读理解)

modelsDEVTEST
google_bert_baseF1:92.296(92.565) EM:86.600(87.089)F1:91.464 EM:85.485
siBert_baseF1:93.343(93.524) EM:87.968(88.28)F1:92.818 EM:86.745
siALBert_middleF1:93.865(93.975) EM:88.723(88.961)F1:93.857 EM:88.033
哈工大讯飞 bert_wwm_ext_baseF1:93.265(93.393) EM:88.002(88.28)F1:92.633 EM:87.145
哈工大讯飞 roberta_wwm_ext_baseF1:94.257(94.48) EM:89.291(89.642)F1:93.526 EM:88.119
哈工大讯飞 roberta_wwm_ext_largeF1:95.323(95.54) EM:90.539(90.692)F1:95.060 EM:90.696
ERNIE1.0F1:92.779(93.021) EM:86.845(87.259)F1:92.011 EM:86.029
xlnet-midF1:92.081(92.175) EM:84.404(84.563)F1:91.439 EM:83.281
brightmart roberta_largeF1:94.933(95.057) EM:90.113(90.238)F1:94.254 EM:89.350
brightmart albert_largeF1:93.903(94.034) EM:88.882(89.132)F1:93.057 EM:87.518
brightmart albert_xlargeF1:94.626(95.101) EM:89.682(90.125)F1:94.697 EM:89.780

CJRC(带有yes,no,unkown的阅读理解)

modelsDEV
siBert_baseF1:80.714(81.14) EM:64.44(65.04)
siALBert_middleF1:80.9838(81.299) EM:63.796(64.202)
哈工大讯飞 roberta_wwm_ext_baseF1:81.510(81.684) EM:64.924(65.574)
brightmart roberta_largeF1:80.16(80.475) EM:65.249(66.133)
brightmart albert_largeF1:81.113(81.563) EM:65.346(65.727)
brightmart albert_xlargeF1:81.879(82.328) EM:66.164(66.387)

CHID(多选成语阅读理解)

modelsDEVTESTOUT
google_base82.2082.0477.07
哈工大讯飞 roberta_wwm_ext_base83.7883.62-
哈工大讯飞 roberta_wwm_ext_large85.8185.3781.98
brightmart roberta_large85.3184.50-
brightmart albert_xlarge79.4479.5575.39
google albert_xxlarge83.6183.1579.95

C3(多选中文阅读理解)

modelsDEVTEST
哈工大讯飞 roberta_wwm_ext_base67.0666.50
哈工大讯飞 roberta_wwm_ext_large74.4873.82
google albert_xxlarge73.6673.28
brightmart roberta_large67.7967.55