Paper reproduction

[xuliuwei]

I recently tried to reproduce the results of your paper, but I was unable to achieve the results of the paper. I trained with two 3090 students and batchsize of 12, and the learning rate is the same as that of the paper. May I ask whether the learning rate is the reason for the poor result? Does the learning rate have a big impact on your result? Can you provide your pre-training model?

[lhaippp]

Did you leverage the python environments in this repo?

pip install -r requirements.txt

[xuliuwei]

I'll try to change the environment. Another problem is that I found that the training parameter Settings in the paper are different from the parameter Settings in the code, which one should I use?

[lhaippp]

I recommend you to use the experiment setting in this repo

[xuliuwei]

Did you run it with four graphics cards?

[lhaippp]

I think the number of GPU cards is not important. As long as you can run the code with the setting

[xuliuwei]

How long does it take you to train an epoch? I used two 3090s and it took 3.5 hours to train an epoch.

[six-finger]

@xuliuwei May I chat with you about this issue? Thank you~ Wechat: margin333