the trainning batchsize

[Liujingxiu23]

what kind of cuda you used or training？ I use L40s 48G, it can not handle the original config, I must reduce the model size and the batchsize.

[cantabile-kwok]

Hmm, it is strange because the largest GPU in the experiment was 32G V100. And the model is not so large. In the original model size, what is the maximum batch size in your case?

[Liujingxiu23]

Thank you for your replay! I merge the model code to my own trainning framework, maybe there is something mismatch when I transfer the code.