In the paper， the weight of unsup_loss is set to 1 to comput the whole loss，but in your code, the weight of unsup_loss is set as ratio_unsup=5.0.
For example， https://github.com/ildoonet/unsupervised-data-augmentation/blob/master/train.py#L71 ，the code“loss += C.get()['ratio_unsup'] * torch.mean(loss_kldiv)”，in the wresnet28x2.yaml "ratio_unsup:5.0".
Could you please explain why?
When I set ratio_unsup=1.0， the final test Top-1 error rate is 11.63%.
In the paper， the weight of unsup_loss is set to 1 to comput the whole loss，but in your code, the weight of unsup_loss is set as ratio_unsup=5.0.
For example， https://github.com/ildoonet/unsupervised-data-augmentation/blob/master/train.py#L71 ，the code“loss += C.get()['ratio_unsup'] * torch.mean(loss_kldiv)”，in the wresnet28x2.yaml "ratio_unsup:5.0".
Could you please explain why?
When I set ratio_unsup=1.0， the final test Top-1 error rate is 11.63%.