ChengShen

The image size is 256, batchsize is 32, and 480k steps, which does not look good. 

Please provide more information, such as which dataset for testing and which version of MegEngine. In my case, this test script worked well and the result was 39.765.

After installing MegEngine=1.5, your results still cannot be reproduced. Please check whether the image pair to be calculated is correct, and check whether nan appears in each image pair or...

Thanks for pointing out this error, I have deleted these external packages. But I don't think this is the reason why the output is **nan** and please clone this project...

@xiongchuanchao After DnCNN, recent CNN denoisers prefer to output the residual image and then add it to the input to get the final image. @FunkyKoki you can find it in...

You can find the validation dataset on the official website of SIDD.

Sorry for the missing code to dump images. Please refer to line 31 in test.py, you can take out the "pred" item and convert it into a numpy array, so...

So, what exactly is all you need？ patches, specific network structure, or just numerous parameters...

Interesting! I replaced the attention in Uformer with pooling and achieved the same effect on low-level tasks.

I also meet the problem. Did you solve it?