lingtengqiu / deeperlab-pytorch Goto Github PK

I looked at your code and found that there is only a semantic segmentation part, and there is no instance segmentation part?

Can this code be used for panoptic segmentation?

I found a mistake in the space to dense function:

class space_to_dense(nn.Module):
    def __init__(self,stride):
        super(space_to_dense,self).__init__()
        self.stride = stride
    def forward(self, input):
        assert len(input.shape) == 4,"input tensor must be 4 dimenson"
        stride = self.stride
        B,C,W,H = input.shape
        assert (W %stride == 0 and H %stride == 0),"the W = {} or H = {} must be divided by {}".format(W,H,stride)
        ws = W // stride
        hs = H // stride
        x = input.view(B, C, hs, stride, ws, stride).transpose(3, 4).contiguous()
        x = x.view(B, C, hs*ws, stride * stride).transpose(2, 3).contiguous()
        x = x.view(B, C, stride * stride, hs, ws).transpose(1, 2).contiguous()
        x = x.view(B, stride * stride * C, hs, ws)
        return x

The line "B,C,W,H = input.shape" should be "B,C,H,W = input.shape"

If you load images having different height and width the concatenation between the lower and the higher order feature map complains about the different h and w of the two tensors.

Can't unzip the pretrained model. I have tried both on ubuntu and windows. It said that An error occurred while loading the archive.
Could you please offer another site or kindly send it to my email: [email protected]
Thank you

Thanks for your share, but the kernel size are 7 in the paper.

I've download the pretrained xception model.
I use the command: tar -xvf *.tar to decompress it but it didn't work
How to use the pretrained model?

build.ninja编译，提示torch 中的python_headers.h 文件找不到 头文件#include <Python.h>

best job.
I have two problem:

1. I have a machine and mutil-GPU, can use the distribute training code?
2. the deeperlab use topK and balanced ,the code only use topk.

As written in the paper, Similar to the encoder, the decoder uses two large kernel (7 x 7) depthwise convolutions to further increase the receptive field. There should be SeparableConvBnRelu instead of ConvBnRelu.