guoshuxuan / expandnets Goto Github PK

Hello, your work is great.
I noticed that the padding operation is used in 1x1 convolution, but it's usually used in 3x3 convolution and I would like to know why.

Very interesting work! I'm curious why not add BN after each extended CONV layer. I don't know if it can bring improvement.

Hi,
I find your work really interesting but I have a question.
How can I expand a network with existing weights?
I want to train a network using this scheme, but I need to start from existing weights, such as imagenet pretraining.

Adding 1x1 convolution without changing the model is simple.
So Expand-CL can be implemented using existing weights.

I'm not sure how to do Expand-CK. If I have an existing k X k kernel for k > 3, how do I split it into a 3x3 kernels?

Hi there, Great Jobs!
Still, when I tried the compute_ck, compute_cl and compute_cl_2 functions, it seems two of them have some errors compared to the origin network output.

The code link is below, pretrained network layers from official torchvision used for calculation.
https://github.com/hiterjoshua/BasicSR/blob/master/tests/expand_net.py

Directly run the upper expand_net.py and you would get the output of error:
error 3*3 and 3*3: -80360.1015625
error 3*3 and 1*1: 0.20960576832294464
error 1*1 and 3*3: -2098.904541015625

It seems the functions error 3*3 and 3*3 and 1*1 and 3*3 have some errors here, would you take a look at this problem? Hoping your answer! Thanks!