Save memory during backward through shared weight

autograd
1

Hi,

I have two convolutions which have the same shared weight. One is using dilation=1, the other is using dilation=2.

Let x be an input tensor, and w the shared weight.

Is there anything I can do to save memory when calculating:

y = F.conv2d(x, weight=w, dilation=1) + F.conv2d(x, weight=w, dilation=2)

and doing the backward pass?

Right now the memory consumption is doubled which feels unnecessary.

2

Hi,

The memory consumption is doubled most likely because both outputs are saved.
Indeed, to compute the backward of the convolution, we need to know what was the output. So if you do two of them, we need to know both outputs.