pkumivision / ffc Goto Github PK

Hello, the forward method of BasicBlock says:

x = self.conv1(x)
x_l, x_g = self.conv2(x)
x_l, x_g = self.se_block(x)

Is this correct? x_l and x_g are overwritten by the se_block. Shouldn't it be like this?

x = self.conv1(x)
x = self.conv2(x)
x_l, x_g = self.se_block(x)

Thanks!

Hi! Greatly appreciate your work! It seems that pretrained weights that you provided in issue #1 doesn't match the code:

• layers are named differently
• there is size mismatch in some layers, f.e. size mismatch for layer3.3.conv2.ffc.convl2l.weight: copying a param with shape torch.Size([192, 192, 3, 3]) from checkpoint, the shape in current model is torch.Size([128, 128, 3, 3]).

Can you please check it out?

Hi, I am using Pytorch 1.9 and I can not find this function torch.rfft.

It seems like Pytorch 1.9 has moved this function into sth else: https://pytorch.org/docs/1.9.0/generated/torch.fft.rfft.html

But I do not see their input arguments similar anymore, can u update this block to be compatible with the new Pytorch version ?

can any one told that what is ratio_gin and out

Thanks for the amazing work.
Could you provide the code for keypoint detection? If I have two images from the same scene, the first one is RGB image, the others is infraed images,whether it extract the same keypoints or not?

Please share implementation on anyone reasonably deep EfficienNet or EfficientNet_v2. I tried to follow method of ResNet but getting channel mismatching error.

Hi，Could you provide the weights trained on imagenet?Thank you !

Using ffc() to directly replace nn.Conv2d() will report an error:
RuntimeError: Given groups=1, weight of size [5, 3, 3, 3], expected input[3, 6, 64, 64] to have 3 channels, but got 6 channels instead

Hi. First of all, thanks a lot for contributing your work to the community.
I noticed that this repo does't include a license file, so it's difficult to know the conditions for rightful use of it. Could you confirm under what conditions can your code be used and preferrably add a license file?

Hello! May I ask where can I get the fcc_resnet parameter pre-trained on ImageNet in the experimental results?

For anyone who might need TensorFlow 2.x implementation of Fast Fourier Convolution (FFC) there is an unofficial port of this project: https://github.com/uladzislau-varabei/ffc-tf2.

In case you have ideas on how to improve it or found a bug just let me know.