Hi 师兄 ^_^, I have some questions, some of which may not be related to your code directly. Look forward to your answers.

1. The GCN paper shows that the larger kernel size of GCN is, the better segmentation result is. I see your default kernel size is 7 (not the largest one 15 in the paper). What is your consideration?
2. The paper seems not to do ablation experiment on the kernel size setting of GCN in intermediate layers of resnet. In intermediate layers, the size of feature map is larger than 16*16, do you have any idea about the kernel size setting in intermediate layers?
3. How about combining GCN with dilation?

Thanks for your sharing of the nice code!

For resnet and other architectures in general, the weight init has to be done separately right? The default init for nn modules of pytorch is uniform? From my understanding, there are better inits like glorot uniform, he init etc which we explicitly need to do. Because other libraries like Keras, Lasagne has the best init as the default, unlike PyTorch.

References: [1] - ResNet official: https://github.com/pytorch/vision/blob/master/torchvision/models/resnet.py#L113-L119 [2] - VGG official: https://github.com/pytorch/vision/blob/master/torchvision/models/vgg.py#L46-L59

[3] - Conv2D nn module: https://github.com/pytorch/pytorch/blob/master/torch/nn/modules/conv.py#L40-L47

Hi @ycszen

Thank you for the repo! I was wondering what is the status of the repo - were you able to test the models and see if implementation accuracy matches Arxiv papers?

There is an official implementation of DenseNet now, which can be used to write DenseNet-FCN.
https://github.com/pytorch/vision/tree/master/torchvision/models

Thanks!

When I test picture from VOC2017,I get the wrong message :1234: Ucat Traceback (most recent call last): File "tester.py", line 46, in output = Colorize()(output) File "/home/xiangjie/Documents/master_thesis/segmentation/pytorch-seg/transform.py", line 150, in call color_image[0][mask] = self.cmap[label][0] RuntimeError: inconsistent tensor size, expected tensor [256 x 256] and mask [256] to have the same number of elements, but got 65536 and 256 elements respectively at /pytorch/torch/lib/TH/generic/THTensorMath.c:138

I try to fix it.But did not work ,so ,someone can tell me how to fix this bug???? Thank you!!!!

In the code of upsample.py, you create FCN structure. But, the structure of your implementation seems to be slightly different from original FCN. Do you have any reference paper of your code ?

Thanks !

TypeError Traceback (most recent call last) in () 1 if torch.cuda.is_available(): 2 #model = torch.nn.DataParallel(FCN(22)) ----> 3 model=FCN(22) 4 model.cuda()

in init(self, num_classes) 81 self.refine10 = Refine(self.num_classes) 82 ---> 83 self.out0 = self._classifier(2048) 84 self.out1 = self._classifier(1024) 85 self.out2 = self._classifier(512)

in _classifier(self, inplanes) 95 96 nn.Conv2d(inplanes, inplanes, 3, padding=1, bias=False), ---> 97 nn.BatchNorm2d(inplanes/2), 98 nn.ReLU(inplace=True), 99 nn.Dropout(.1),

/home/omnisky/anaconda3/lib/python3.6/site-packages/torch/nn/modules/batchnorm.py in init(self, num_features, eps, momentum, affine) 16 self.momentum = momentum 17 if self.affine: ---> 18 self.weight = Parameter(torch.Tensor(num_features)) 19 self.bias = Parameter(torch.Tensor(num_features)) 20 else:

TypeError: torch.FloatTensor constructor received an invalid combination of arguments - got (float), but expected one of:

• no arguments
• (int ...) didn't match because some of the arguments have invalid types: (float)
• (torch.FloatTensor viewed_tensor) didn't match because some of the arguments have invalid types: (float)
• (torch.Size size) didn't match because some of the arguments have invalid types: (float)
• (torch.FloatStorage data) didn't match because some of the arguments have invalid types: (float)
• (Sequence data) didn't match because some of the arguments have invalid types: (float)

Hi @ycszen , thanks for sharing the nice codes. I have a question about the performance on cityscape of the models, such as DUC, fcn, etc. Could you please tell me your performance result?