Question about awesome-semantic-segmentation-pytorch

[Tramac]

Hi, in the alternative estimation, there is no data augmentation (rescale, random crop..) for the single test image.
In the previous test, We forgot to change mode to testval, which controls the data augmentation operation in dataloader.

[82018303kazuma]

Hi Thank you for your reply

[Tramac]

Thank you for your participation.
Is there any further testing? Has the mIoU improved?

[82018303kazuma]

Performance is improved by the optimization algorithm.
Improvements were also seen with Augmentation(random Erase).
I feel that the performance was bad even with overfitting last time. Since it has been improved this time, I will try to improve the performance from now on.

[Tramac]

Thank you for your job!
Now the code already supports saving the best model, you can change fcn32s_%s_%s % (backbone, acronyms[dataset]) to fcn32s_%s_%s_best_model % (backbone, acronyms[dataset]) in there( for FCN32s). It This may improve performance.

[82018303kazuma]

I feel that the training itself was not done well last time.
In the framework Pytorch was just too low, but the code has been improved and is almost equivalent.

Thank you for the meaningful code.