Nice Work! I'm interested in ECS. I want to realize it.

[yuheyuan]

I want to know.
If
for each img x
you predict x using model and do softmax for the output. You calculate every class's probabiliy in all pixels.

For example.
for a img shape [batch, channel, H, W], for 16 classification.
You can get a ECS map shaple like [batch, channel=16 ,H ,W], each channel represent one class ECS map?
Then you use

I want to know topk(e, k) if represents choose k classes which have the high CS do mix img.
Here is not something that can be considered similar. For example, I calculate the entropy of each class, and then I sort and select the seven classes with the lowest entropy to do image mixing.