sair-lab / kervolution Goto Github PK

What is the difference between the Kerv2D polynomial kernel with a degree of 2, balance of 1 and Conv2D kernel which is followed by an activation layer of x^2?

This is an issue concerning the architectural arrangement.

In deeper layers, for example, in Resnet18, changing all convolutional layers to kervolutional layers (with degree 3), the network gives a NaN loss which suggests explosion. How do you arrange the architecture to give good results

Hi, I find the nn.unfold is much slow in pytorch. And your implementation is based on nn.unfold. I want to know if you have any tricks to speed up it ?

Hi,
I got much insight from your amazing work.
But it cost too much memory because the operator F.unflod.
It will almost consume 9 times memory if the kernel size is (3, 3).
Is there any way that implements KERV with little memory consumption?

Hi. I am working on 3D images. I want to implement kerv3d similar to 3D CNN. Can you please provide the kerv3d layer in your repository. How can we extend your method for three-dimensional data?

Hi. I am working on 3D images. I want to implement kerv3d similar to 3D CNN. Can you please provide the kerv3d layer in your repository. How can we extend your method for three-dimensional data?

Hi,

If you try to use kervolution without bias, it does not work if you set bias=False.

The reason is in line 85 in kervolution.py in function Kerv2d:

if self.bias is not None:
output += self.bias.view(self.out_channels, -1)

So you need to set bias=None to not have bias

The same is true for the function Kerv1d