suppose the shape of x is [2, num_channels, 4, 5], and I compute per-channel summation as follows: s = torch.zeros(num_channels) for i in range(num_channels): s[i] = x[:, i].sum() It turns out this implement is too slow. Is there a faster implement?

Hi, Using built-in functions, it would be like this: num_channels = 10 x = torch.randn(2, num_channels, 4, 5) result = torch.sum(x, dim=[0, 2, 3]) torch.sum accepts a parameter dim which can be a list. If you pass a list, it will reduce the tensor over all the indices in dim. For instance, x = t…

How to compute per-channel sum fast?

vision

Nikronic (Nikan Doosti) May 13, 2020, 9:37am 4

glad it worked,
Actually, there is no built-in function for max or argmax like sum. But you can do the trick something similar to this post: