I met a problem with the model training

[KindredSpirithub]

After training multiple epochs, I found that the accuracy of the model's test set has always been a very low value. After checking the data side, there was no problem, so I don't know where the problem lies?

[linhuixiao]

Hi, there must be something wrong. Could you replicate the results based on the checkpoint I provided? If it's true, environmental concerns can be ruled out. In general, after 1 epoch of training, the accuracy can reach 13.0-25.0.

[KindredSpirithub]

yes, i can replicate the results based on the checkpoint you provided, and I didn't make any changes to the source code

[linhuixiao]

@KindredSpirithub Hi, I have pulled the code from the repository and tested it in a new environment. The model converged quickly and replicated the results. Thus, the issue you are encountering is probably not a bug in the code; it may be a problem with your environment or configuration. Maybe you can double-check the configuration or environment.
（I noticed that your training command is not configured exactly according to the readme instructions; you can try using torch.distributed.launch.）

[KindredSpirithub]

I just change the training command by using torch.distributed.launch, then it converge normally. thank you very much