Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

请教下XNLI的复现结果? #4

Closed
zhuchenxi opened this issue Jun 2, 2021 · 1 comment
Closed

请教下XNLI的复现结果? #4

zhuchenxi opened this issue Jun 2, 2021 · 1 comment

Comments

@zhuchenxi
Copy link

hello~ 我这边使用train_xnli.sh的默认参数进行了训练,最终得到平均acc为78.63%,和论文中的79.9%的结果有些出入,且看上去应该是每种语言的acc都低了一点。请问下可能是什么原因导致的呢?非常感谢。
我这边是用的单核GPU,采用Tesla M40 24GB卡,Driver Version: 440.64,CUDA Version: 10.2
我从和论文中的fine-tune参数对比上看。
一个是sh代码里面使用的epoch=2,论文中是用的3/5/10选的
另一个是这里TOTAL_BATCH_SIZE=64,BATCH_SIZE=2,论文中是16/32/64选的
会不会是这两个超参导致的出入呢?
非常感谢~

@hit-computer
Copy link
Collaborator

@zhuchenxi hello! 以下是我们实验的参数设置:

LR=2e-5
EPOCH=2
TOTAL_BATCH_SIZE=64

实验环境配置是:torch Version:1.3.1;transformers Version: 2.3.0;GPU:Tesla V100 16GB;CUDA Version:10.0.130

实验结果在英语的dev集上通过acc选择最优模型(模型会保存到 checkpoint-best 目录下),然后用该模型对测试集进行预测,以下是我们复现的结果:

ar=0.7868263473053893
bg=0.8291417165668663
de=0.8235528942115768
el=0.8191616766467066
en=0.883433133732535
es=0.8471057884231536
fr=0.8297405189620759
hi=0.7606786427145709
ru=0.8011976047904191
sw=0.7351297405189621
th=0.7590818363273453
tr=0.7888223552894211
ur=0.7153692614770459
vi=0.7998003992015968
zh=0.7918163672654691

=====
Avg_lang_acc =0.7980572188955422

有问题欢迎随时交流,十分感谢!

@luofuli luofuli closed this as completed Jun 15, 2021
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

3 participants