kkahatapitiya / linearconv Goto Github PK

First, thank you for putting such clear and concise work at our disposal. I have a question concerning the FLOPs' calculation of the LinearConv layer

Looking at the linear operations' calculation, I see to get the total addition, you subtracted the total multiplications by one. Why did you do that? If each output filters were just a linear combination of alpha*(number of filters), is the number of multiplication not supposed to be the same as the number of additions?

Hi,

I have tried to run your code and train a model of mine with it. But I stumble upon the increase in loss when I added the regularization correction on the xCNN weights.
Applying that to the low-rank version of your layer I get an huge increase. Have you ever found something like that as well? Or is this correction only meant for the vanilla implementation?

Greetings,
Seb

Hi @kkahatapitiya ，can we use LinearConv in image reconstruction tasks?
When I use LinearConv in image denoising, the color of the result image is not right. The loss of the tasks contains L1 loss of the (label-out) and regularization loss of weight. Can you please find what's the problem of my job?