Hey, how much time it approximately takes to train the SONY model using the specified configuration?

gt_full = np.expand_dims(np.float32(im/65535.0),axis = 0)，
I think it should be 65532.

How to understand origin_full = scale_full？Why are the results very different?
How to understand scale_full = scale_full*np.mean(gt_full)/np.mean(scale_full)？
What is the role of scale_full?

In the train_sony code,how to modify the batch_size?

Hi

CPU works, but cuda shows error. I cant figure out how to fix it.
TypeError: can't convert cuda:0 device type tensor to numpy. Use Tensor.cpu() to copy the tensor to host memory first.

Thanks

Hi ninetf135246, I am not able to understand why you are permuting the input tensor in line #142 of train_Sony.py. Please help me to understand your code.
https://github.com/ninetf135246/pytorch-Learning-to-See-in-the-Dark/blob/8af3e309d6947b1addf52765869190ec1e6e2ded/train_Sony.py#L142

Here in the second if, I think the axis should be 2. axis=0 means flip along the batch dimension

    if np.random.randint(2,size=1)[0] == 1:  # random flip 
        input_patch = np.flip(input_patch, axis=1)
        gt_patch = np.flip(gt_patch, axis=1)
    if np.random.randint(2,size=1)[0] == 1: 
        input_patch = np.flip(input_patch, **axis=0**)
        gt_patch = np.flip(gt_patch, **axis=0**)
    if np.random.randint(2,size=1)[0] == 1:  # random transpose 
        input_patch = np.transpose(input_patch, (0,2,1,3))
        gt_patch = np.transpose(gt_patch, (0,2,1,3))

How to understand im = raw.postprocess(use_camera_wb=True, half_size=False, no_auto_bright=True, output_bps=16)

Thanks for releasing the code. I want to know if you have got the results of the paper when using the pytorch codes.

Hello,
You mentioned that you slice the original input images into small pieces(1024*1024) in readme, however, I cannot find the corresponding operations in the test code. Would you please tell me where is the slice operation? thank you~

when you test,you can use torch.nograd().

I'm pretty sure you can just pass the whole image into the network for testing, so you wouldn't need to slice it up.

I noticed that the initializing operation was just defined as a function, and was not used in the init() part. Does it still work in this way?

You have added a _initialize_weights() function, does it is necessary to pytorch?

def _initialize_weights(self): for m in self.modules(): if isinstance(m, nn.Conv2d): m.weight.data.normal_(0.0, 0.02) if m.bias is not None: m.bias.data.normal_(0.0, 0.02) if isinstance(m, nn.ConvTranspose2d): m.weight.data.normal_(0.0, 0.02)

Hi, thanks for your work!
I am trying test.py on gpu, but got error "cuda out of memory".(rtx 2080ti)
Any solution could solve this?
thanks!

I want to use pytorch to implement a new loss function. And almost all the online information say that I should make a new class that inherit the nn.Module. But I found you just create a function ‘reduce_mean()’. Does it means that it is just so easy in pytorch to make an own funtion.

虽然时间过去很久了，但是可以的话还是想问下，颜色信息这个模型是学不出来吗？我刚练到40多个epoch