Problem about the function torch.index_select(input, dim, index, out=None)

[pcshih]

Is this function diff. ?

[Tony-Y]

It is not diff. wrt index.

[pcshih]

"By the way,

summary = x_select+h_select
summary = self.relu_summary(summary)

summary should resemble x_select, but when summary equals to x_select, h_select is vanished."

I just follow FCSN to do the above implementation.

Thank you very much, Tony-Y.

[pcshih]

How can I implement "If Score(key) > Score(not) then it is a key frame."

I can only think of "max" implementation to get the index but now it's non-differentiable.

[Tony-Y]

If you really want back propagation though index_mask, you could use sigmoid function as the following:

index_mask = sigmoid( Score(key) -Score(not) )