Is there any difference between AdaptiveAvgPool and mean()?

1 
x = torch.rand(2, 3, 128, 128)
F.adaptive_avg_pool2d(x, output_size=1).flatten(start_dim=1)

x = torch.rand(2, 3, 128, 128)
x.mean(dim=(2, 3))

Is there any difference between two above?

I think both are same and the second one is more simple and understandable.
But I assume there is any reason why most people are using first one.

2

There is no difference in the posted code snippets and the former one will also use the latter approach internally for an output_size of 1.
The former allows you to use other output sizes and is thus more flexible. Also, you could use it as an nn.Module without creating a custom module for the mean operation.

1 Like