mohamedac29/s2-fpn

Scale-ware Strip Attention Guided Feature Pyramid Network for Real-time Semantic Segmentation

s2-fpn's Issues

Fail downloading ResNet34M pre-trained weights on CamVid

Hi, thanks for your sharing. i am interested in this work. However, i find i cant download the weights CamVid | ResNet34M | trainval | 74.76% | 55.5 | ckpt with this link .

Could you add another link to download ResNet34M pre-trained weights on CamVid?

Thank you a lot!

About Equation 6 in paper 3.2

In paper，Equation 6 is Fscale = A ⊙F2 + A⊙F2, but with the text description, I think it should be Fscale = A⊙F1 + A⊙F2.

About Equation 7 in paper 3.2

In paper, Equation 7 is FSSAM = αFscale + (α-1)F, but in Fig. 4, it shows α and (1-α). And in your code, it's (1 - self.gamma). So I think the Equation 7 should be FSSAM = αFscale + (1-α)F.

the description of SSAM is inconsistent with the code

theoretical code：
att = torch.softmax(torch.mul(mxpool, avgpool), 1)
fusion = attt * avgpool + attt * mxpool

actual code :
att = torch.softmax(torch.mul(mxpool, avgpool), 1)
attt1 = att[:, 0, :, :].unsqueeze(1)
attt2 = att[:, 1, :, :].unsqueeze(1)
fusion = attt1 * avgpool + attt2 * mxpool

why？

SpatialAttention class

In this class, I find mxpool = F.max_pool2d(x, [h, 1]), and after this, maxpool will be (B, C, 1, W), but in paper, it's (B, C, H, 1). I have miss about it. Please explain, thanks.

SpatialAttention class

In the SpatialAttention class, both maxpooling and avgpooling functions use F.avg_pool2d, which is inconsistent with the paper, please explain.Thanks

inform pkl

no cityspaces.pkl

CamVid | ResNet34M的模型

请问CamVid | ResNet34M的训练模型还能下载吗，我发现没有下载权限？

AttributeError: 'NoneType' object has no attribute 'copy' error.

When I train with the cityscapes dataset, there is an AttributeError: 'NoneType' object has no attribute 'copy' error. How should I change the dataset path? Is there another error? Datasets are stored as readme files