Please check if the code is the same as the code used in the experiment.

[yeonghwansong]

Hello! I enjoyed studying your work.
This code causes runtime errors when executed with serescam50 because the order of transformations between training code and model code is different. While odd transformation locates last in senet code, it locates first in training code. Furthermore, even if the network is trained after fixing this problem, can't achieve performance argued by paper. Maybe, I think the training code and VGG code were revised after experiments using Se-ResNet.
I'd like to ask you a question about PST, too. According to your code, PST is applied only when the random number is 1 or 3. If the random number is 0 or 4, just rotate 180 degrees. I checked it experimentally. I think it is not identical to paper. I'd appreciate it if you could answer.

[FriedRonaldo]

Hello, thanks for pointing out the issue.

We checked the source code and we found out that the code that you pointed out was missing while refactoring.

For reproduction, we uploaded the raw source codes for VGG and SE-ResNet in the different branches.

We also checked the experimental results and found that we used PST=1.0 in the paper for consistency, but the model showed better performance with PST=3.0.

In the case of SE-ResNet, using PST only outperforms the model trained with PST and rotation usually.

The exact training configuration and the checkpoint for SE-ResNet of the paper are missing because of the issue on our server computers.

We share the pretrained models that we reffered to and it shows better performance.
(VGG: 83.xx --> 85.xx (PST=1.0 --> 3.0) / SE-ResNet: 85.xx --> 87.xx (using PST only))

We uploaded the checkpoint files at Google drive.