Alex Hex

Hi @pantlive , you can find details in this paper called **[Non-local Neural Networks](https://arxiv.org/pdf/1711.07971.pdf)**.

@csg1111 请问训练的是demo_MNIST.py吗？ 另外训练的时候loss有下降吗，是不是没有loss.backward()以后，没有更新参数？

如果不设置为0,结果会怎样？ 我测试了一下，demo_MNIST.py训练后不是0. 之前我也有可视化过block里面的矩阵，不会存在全0的情况。 可以把代码放出来吗？会不会是optimizer里面没有把block的参数传进去？

@BryantGary Hi. 我在一些Image Enhancement任务上测试过，Metric有明显提升。另外，tonylins的Repo https://github.com/AlexHex7/Non-local_pytorch/issues/23 也report有提升。

@tonylins No, problem! Thanks for telling me this. I have created the License file and added your repo link into my ReadMe file. By the way, I look through your...

Yes. It is only used in the top of Network. You can find the details in the paper called **Non-local Neural Networks**.

@1483852602 As far as I remember, this paper does not mention it. So the initialization way is defined in nn.Conv2D() of Pytorch. ``` class _ConvNd(Module): ... def reset_parameters(self): n =...

@1483852602 I see and thanks for point it out. As I do not know caffe framework well, I implement my code based on the paper. I am not sure does...

@Mr-doraemon Hi, it seems that **self.W[1].weight** is None.

@wys2929 Hi, 是否因为你的系统上没有检测到GPU？net.pth可能保存为gpu上的形式。可以尝试强制将权重读取到cpu上 `torch.load('weights/net.pth', map_location='cpu')` 抱歉，是代码中检查GPU逻辑的疏忽。 https://github.com/AlexHex7/Non-local_pytorch/blob/ee894ba21bc038ff4892bf7e9be5ac19a896ecee/nl_map_save.py#L16