Overwriting convolution weight

autograd
1

I’d like to compute a 2D convolution with kernel weights that I calculate somewhere else. I.e. I want to treat the kernel like an input to the Conv2d module similar to hidden states in a recurrent network. Can I simply overwrite the ‘weight’ attribute of a torch.nn.Conv2d object with a Variable and then calculate the convolution? Or will that cause the gradients to be incorrect?
My concern is that the gradients might not be propagated back to the inputs correctly.

2

You can use torch.nn.functional.conv2d :). The module Conv2d is just a wrapper around the functional interface.

3

Thanks! I’ll do that.