Suppose I have three consistent convolutional layers like (256, 144, 1)-> (144, 144, 3) -> (144, 256, 1). In fact, the first and the last one are matrices. Is there any way to ‘merge’ these three layers in a one preserving the trained information? Since there is no non-linearity between them. It’d be similar to a linear layers, for example (we can replace 2 of them with a one).
I found smth similar here https://stackoverflow.com/questions/58357815/how-do-i-merge-2d-convolutions-in-pytorch