reconstruction

[LeeHaoRanRan]

Hello, after I reconstructed it with your network, why is the reconstructed image getting closer and closer to the blank image ------------->

[ci-ber]

It seems like an unexpected behavior, can you give me more context? what are you trying to do? Re-train the network, run inference? on what data? It seems like the input brain scan has a different orientation (rotated 90 degrees) but I still would not expect this behavior

[LeeHaoRanRan]

这似乎是一个意想不到的行为，你能给我更多的背景吗？你想做什么？重新训练网络，运行推理？在什么数据上？似乎输入脑部扫描的方向不同（旋转 90 度），但我仍然不希望出现这种行为

I used my own dataset for training, and the image of the input model and the image after adding noise are normal, but when the noise is predicted during the training process, the predicted noise becomes whiter and whiter, but the training loss is indeed decreasing

[LeeHaoRanRan]

这似乎是一个意想不到的行为，你能给我更多的背景吗？你想做什么？重新训练网络，运行推理？在什么数据上？似乎输入脑部扫描的方向不同（旋转 90 度），但我仍然不希望出现这种行为

And during the training process, I also process my own data into the range of [-1--1].

[ci-ber]

You could try normalizing to [0,1] and also could you let me know what output are you visualizing for the second 'white' image?

[LeeHaoRanRan]

您可以尝试归一化为 [0,1]，也可以让我知道您为第二个“白色”图像可视化的输出是什么？

Ok, I'll go and implement that later, right now I don't have the GPU resources

[LeeHaoRanRan]

您可以尝试归一化为 [0,1]，也可以让我知道您为第二个“白色”图像可视化的输出是什么？

好的，我稍后会去实现它，现在我没有 GPU 资源

After changing the image to the 0-1 range, there is still an error in predicting the change in noise, and the following column shows the change in the predicted noise, and finally there is still white -----------> -----------> ----------->

[ci-ber]

Could you share which outputs you are plotting and maybe the input and noised input image?

[LeeHaoRanRan]

In the training process, the input image, the input noise image and the output noise are respectively Then the input noise and the predicted output noise gradually become Training losses are falling accordingly In the verification phase of a particular epoch in the same environment, the initial input image and the reconstructed image are

After training for some time, the original image and the reconstructed image become respectively I conducted training and verification with a very small amount of data, but this situation also occurred after 60 rounds of training on a large amount of data. I don't understand what caused it

[ci-ber]

The input noise image seems like pure noise, the input image is not recognizable anymore, are you sure this is right? I added now the mid-axial slices from IXI as png (CAPS_IXI/png). You can use these as a reference to check the model implementation. You can use the (ixi_normal_train.csv) file

[Babaala]

In the training process, the input image, the input noise image and the output noise are respectively Then the input noise and the predicted output noise gradually become Training losses are falling accordingly In the verification phase of a particular epoch in the same environment, the initial input image and the reconstructed image are

After training for some time, the original image and the reconstructed image become respectively I conducted training and verification with a very small amount of data, but this situation also occurred after 60 rounds of training on a large amount of data. I don't understand what caused it

hello, I met the same problem as you do. And I fixes it by changing this to image_0 = torch.clamp(image_0, -1, 1).

You can try it. Hope you fix your problem soon.