FADA
FADA copied to clipboard
why is the test size different from the train size in the self-distill phase ?
I trained the self-distill phase and found that the train size is (1024, 512) but the test size is (2048, 1024). Fortunately, the mIoU increased. But I felt so confused because of this. Can you explain me this issue ? Thanks !
This is because different data augmentation is performed. As you can see here, when INPUT_SCALES_TRAIN
is provided, the SOURCE_INPUT_SIZE_TRAIN
will be explained as crop size on the scaled image. This naming might be a little bit confusing, but I cannot come up with a better way to name these operations. Do you have any suggestions?
But why is the test input size (1024, 512) in source training phase and adversarial phase ? (default) Why didn't you use the same test size with the distillation phase. That made change the performance a lot.
Okay, I found the difference. It's so amazing. But I feel like how you augmented the data is quite reluctant with the scale ratio and the crop ratio and when they were used (source training phase, adversarial phase and distillation phase). Can you show me the connection and the idea what lead you to augmentation ? Thanks !