Convolution in PyTorch with non-trainable pre-defined kernel

Question

I would like to introduce a custom layer to my neural network. The mathematical operation should be a discrete 2D cross correlation (or convolution) with a non-trainable kernel. The values in the kernel depend on three things: kernel shape, strides and padding. I intend to multiply the output element-wise with a weight matrix.

PyTorch already has an implementation of a discrete 2D cross correlation class called 'Conv2d', however it generates a random kernel and trains using the entries of said kernel. If possible I would like a class similar to 'Conv2d' that does what I need, to make sure to use my GPU most effectively. I tried implementing this on my own, but couldn't figure out how to obtain the correct shapes for the input array. 'Conv2d' only uses 'in_channels', if I understood correctly.

Answer 1

If I understand correctly,you want a Conv2d layer with defined kernel and you don't want it to be learnable.

In that case,you can use the conv2d function like this:

import torch.nn.functional as F
output_tensor = F.conv2d(input_tensor, your_kernel, ...)

the parameter your_kernel is your weight matrix,also you need to declare other parameters of the function like padding and stride

Then you need to set the requires_grad attribute to 'False', and exclude it from the optimizer if you don't want it to be learnable.

And about the shape issue,maybe you wanna check this out