accuracy mismatch and logger is empty

[Urheen]

Hi authors:

after running cifar100_resnet18_k200.sh, I found that the results is 76.09%, but what reported in the paper is higher than it.

Furthermore, I found that the logger file "./outputs/cifar100/train_p_diff.log" is empy.
Could you please help with this issue?

Best

[1zeryu]

The performance of generated models is related to your training data that is training using SGD.
Would you provide more details like original model performance? The SGD-trained model performance may be affected by many factors, i.e., random seed, GPU type, so the 76.09% maybe reasonable results.
The train_p_diff.log contains terminal output about experiment initialization and torch.distributed.nn. This does not affect the conduct of the experiment.
If still have any questions or concerns, feel free to contact us.

[Urheen]

Hi,

You mentioned, "The train_p_diff.log contains terminal output about experiment initialization and torch.distributed.nn. This does not affect the conduct of the experiment."

However, after running your code, I found that the log file is empty. Everything is shown in the terminal but is not saved to the local log file.