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
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%。
model
length
batch
memory
roberta_base_fp16
512
32
16GB
roberta_large_fp16
512
12
16GB
参与任务
CMRC 2018:篇章片段抽取型阅读理解(简体中文,只测了dev)
DRCD:篇章片段抽取型阅读理解(繁体中文,转简体, 只测了dev)
CJRC: 法律阅读理解(简体中文, 只有训练集,统一90%训练,10%测试)
CHID: 多选成语阅读理解
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
models
config
google_bert_base
L=12, H=768, A=12, max_len=512
siBert_base
L=12, H=768, A=12, max_len=512
siALBert_middle
L=16, H=1024, E=128, A=16, max_len=512
哈工大讯飞 bert_wwm_ext_base
L=12, H=768, A=12, max_len=512
哈工大讯飞 roberta_wwm_ext_base
L=12, H=768, A=12, max_len=512
哈工大讯飞 roberta_wwm_ext_large
L=24, H=1024, A=16, max_len=512
ERNIE1.0
L=12, H=768, A=12, max_len=512
xlnet-mid
L=24, H=768, A=12, max_len=512
brightmart roberta_middle
L=24, H=768, A=12, max_len=512
brightmart roberta_large
L=24, H=1024, A=16, max_len=256
brightmart albert_large
L=24, H=1024, E=128, A=16, max_len=512
brightmart albert_xlarge
L=24, H=2048, E=128, A=32, max_len=512
google albert_xxlarge
L=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
models
cmrc2018
DRCD
CJRC
哈工大讯飞 roberta_wwm_ext_base
epoch2, batch=32, lr=3e-5, warmup=0.1
同左
同左
哈工大讯飞 roberta_wwm_ext_large
epoch2, batch=12, lr=2e-5, warmup=0.1
epoch2, batch=32, lr=2.5e-5, warmup=0.1
-
brightmart roberta_middle
epoch2, batch=32, lr=3e-5, warmup=0.1
同左
同左
brightmart roberta_large
epoch2, batch=32, lr=3e-5, warmup=0.1
同左
同左
brightmart albert_large
epoch3, batch=32, lr=2e-5, warmup=0.05
epoch3, batch=32, lr=2e-5, warmup=0.05
epoch2, batch=32, lr=3e-5, warmup=0.1
brightmart albert_xlarge
epoch3, batch=32, lr=2e-5, warmup=0.1
epoch3, batch=32, lr=2.5e-5, warmup=0.06
epoch2, batch=32, lr=2.5e-5, warmup=0.05
cmrc2018(阅读理解)
models
setting
DEV
哈工大讯飞 roberta_wwm_ext_large
tf单卡finetune batch=12
F1:89.415(89.724) EM:70.593(71.358)
models
DEV
google_bert_base
F1:85.476(85.682) EM:64.765(65.921)
sibert_base
F1:87.521(88.628) EM:67.381(69.152)
sialbert_middle
F1:87.6956(87.878) EM:67.897(68.624)
哈工大讯飞 bert_wwm_ext_base
F1:86.679(87.473) EM:66.959(69.09)
哈工大讯飞 roberta_wwm_ext_base
F1:87.521(88.628) EM:67.381(69.152)
哈工大讯飞 roberta_wwm_ext_large
F1:89.415(89.724) EM:70.593(71.358)
ERNIE1.0
F1:87.300(87.733) EM:66.890(68.251)
xlnet-mid
F1:85.625(86.076) EM:65.312(66.076)
brightmart roberta_middle
F1:86.841(87.242) EM:67.195(68.313)
brightmart roberta_large
F1:88.608(89.431) EM:69.935(72.538)
brightmart albert_large
F1:87.860(88.43) EM:67.754(69.028)
brightmart albert_xlarge
F1:88.657(89.426) EM:68.897(70.643)
DRCD(阅读理解)
models
DEV
TEST
google_bert_base
F1:92.296(92.565) EM:86.600(87.089)
F1:91.464 EM:85.485
siBert_base
F1:93.343(93.524) EM:87.968(88.28)
F1:92.818 EM:86.745
siALBert_middle
F1:93.865(93.975) EM:88.723(88.961)
F1:93.857 EM:88.033
哈工大讯飞 bert_wwm_ext_base
F1:93.265(93.393) EM:88.002(88.28)
F1:92.633 EM:87.145
哈工大讯飞 roberta_wwm_ext_base
F1:94.257(94.48) EM:89.291(89.642)
F1:93.526 EM:88.119
哈工大讯飞 roberta_wwm_ext_large
F1:95.323(95.54) EM:90.539(90.692)
F1:95.060 EM:90.696
ERNIE1.0
F1:92.779(93.021) EM:86.845(87.259)
F1:92.011 EM:86.029
xlnet-mid
F1:92.081(92.175) EM:84.404(84.563)
F1:91.439 EM:83.281
brightmart roberta_large
F1:94.933(95.057) EM:90.113(90.238)
F1:94.254 EM:89.350
brightmart albert_large
F1:93.903(94.034) EM:88.882(89.132)
F1:93.057 EM:87.518
brightmart albert_xlarge
F1:94.626(95.101) EM:89.682(90.125)
F1:94.697 EM:89.780
CJRC(带有yes,no,unkown的阅读理解)
models
DEV
siBert_base
F1:80.714(81.14) EM:64.44(65.04)
siALBert_middle
F1:80.9838(81.299) EM:63.796(64.202)
哈工大讯飞 roberta_wwm_ext_base
F1:81.510(81.684) EM:64.924(65.574)
brightmart roberta_large
F1:80.16(80.475) EM:65.249(66.133)
brightmart albert_large
F1:81.113(81.563) EM:65.346(65.727)
brightmart albert_xlarge
F1:81.879(82.328) EM:66.164(66.387)
CHID(多选成语阅读理解)
models
DEV
TEST
OUT
google_base
82.20
82.04
77.07
哈工大讯飞 roberta_wwm_ext_base
83.78
83.62
-
哈工大讯飞 roberta_wwm_ext_large
85.81
85.37
81.98
brightmart roberta_large
85.31
84.50
-
brightmart albert_xlarge
79.44
79.55
75.39
google albert_xxlarge
83.61
83.15
79.95
C3(多选中文阅读理解)
models
DEV
TEST
哈工大讯飞 roberta_wwm_ext_base
67.06
66.50
哈工大讯飞 roberta_wwm_ext_large
74.48
73.82
google albert_xxlarge
73.66
73.28
brightmart roberta_large
67.79
67.55