Validation data during training

[ghost]

Hi @MSiam

Thanks for sharing the useful code.

I have one question regarding saving the model during training.

For saving the model during training, Do you only consider images of seen classes to check performance (from validation images of pascal ) and save the model or you check performance in all validation images containing seen and unseen classes?

Because in released code of CAnet for fewshot segmentation, I noticed authors check performance on images of unseen classes for saving the best model during training!

[ghost]

I explored your code and I noticed that in the loader for both train_aug and val images all unseen classes pixels set to 250 as ignored pixels. Therefore, you save the model during training based on performance in seen classes in validation images of pascal.

Is it correct?

Thanks