d-li14 / psconv

[ECCV 2020] PSConv: Squeezing Feature Pyramid into One Compact Poly-Scale Convolutional Layer

psconv's Issues

Can you share the code for training imagenet?

Thanks for your great work. Can you share the code for training imagenet? I use the code (https://github.com/rwightman/pytorch-image-models) and train ResNeXt-50+PSConv with the same setting in your paper (SGD, 120 epochs, half cosine function shaped schedule, cross entropy, standard augmentation, weight decay 1e-4). The top1 performance is only 77.57%, which is lower than the results reported in your paper (79.622%).

I wonder why the code psconv.py are not consistent with the description in the paper?

x1, x2 = x.chunk(2, dim=1)
x_shift = self.gwconv_shift(torch.cat((x2, x1), dim=1))
return self.gwconv(x) + self.conv(x) + x_shift

I think the code is simply add with different dilation.

Optimized version of PSConv

Hello, thank you for nice work!

It seems like that you made optimized version of PSConv as mentioned in the supplementary material of the paper. As it shows a great speedup, it could be very useful implementations for applying PSConv to various models.

Do you have a plan to release this optimized version of PSConv?

Paper link

Hi! We meet again. Congratulations! It is an interesting work. Where is the paper ?

How does the code in the psconv.py file achieve the different dilation rates written in the paper?

How does the code in the psconv.py file achieve the differentdilation rates written in the paper?

I had a problem trying to use PSconv

I tested PSconv through a small code in the psconv.py file.

if __name__ == '__main__':
    net = PSConv2d(3, 3, 1, 1, 1, 1, 1, False)
    print(net)
    net = net.cuda()

    var = torch.FloatTensor(1, 3, 127, 127).cuda()

RuntimeError: The size of tensor a (129) must match the size of tensor b (131) at non-singleton dimension 3

how to i use psconv in my own model

This is very interesting work. I am tempted to give it a try on my own model for segmentation. Can you provide some instructions on how to replace a standard conv with psconv in my model?

Can't simply replace nn.Conv2d with PSConv2d

Thanks to the compact characteristic of PSConv, just replace nn.Conv2d with PSConv2d. Note that there exists another hyperparameter named parts you may set in our PSConv operator.

Originally posted by @d-li14 in #3 (comment)

Hi, I'm reading your paper and there comes some problems, hope you could help me figer it out!
Origin Conv2d was
Conv2d(3, 64, kernel_size=(7, 7), stride=(2, 2), padding=(3, 3), bias=False)
PSConv2D are
PSGConv2d( (gwconv): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=4, bias=False) (gwconv_shift): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(2, 2), dilation=(2, 2), groups=4, bias=False) (conv): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False) )
And I add
self.weight = self.conv.weight self.bias = self.conv.bias
To solve the problem
" 'PSGConv2d' object has no attribute 'weight' " during building Resnet,
I don't know I'm right or wrong but it works temporarily.
After that ,durning training part,in PSConv2d.forward
self.gwconv(x).shape:
torch.Size([6, 256, 202, 274])
self.conv(x).shape
torch.Size([6, 256, 202, 274])
but x_shift.shape:
torch.Size([6, 256, 204, 276])

So....The size of tensor a (274) must match the size of tensor b (276) at non-singleton dimension 3

I'm using FCOS original code from tianzhi0549,thanks for the contribution, and add "psconv.py | conv_module.py | conv_ws.py " norm.py" from this respository.

Thank you for your time!