Backward anomaly problem

[ikourbane]

Thank you very much for your work, I found an anomaly when I try to differentiate it and do the backward step, your Net has many in-place operations which makes it impossible to calculate the gradients. If anyone encounters the same problem replace the first few lines in the forward function to:

    def forward(self, img):

        blurred_img_r = self.gaussian_filter_vertical(self.gaussian_filter_horizontal(img[:, 0:1]))
        blurred_img_g = self.gaussian_filter_vertical(self.gaussian_filter_horizontal(img[:, 1:2]))
        blurred_img_b = self.gaussian_filter_vertical(self.gaussian_filter_horizontal(img[:, 2:3]))

        grad_x_r = self.sobel_filter_horizontal(blurred_img_r)
        grad_y_r = self.sobel_filter_vertical(blurred_img_r)
        grad_x_g = self.sobel_filter_horizontal(blurred_img_g)
        grad_y_g = self.sobel_filter_vertical(blurred_img_g)
        grad_x_b = self.sobel_filter_horizontal(blurred_img_b)
        grad_y_b = self.sobel_filter_vertical(blurred_img_b)

[ikourbane]

I just fixed the inplace operations.

[anjaeju]

Thanks you for Issues ikourbane.

I also encounter this issues here.
Even though, i changed the code you write above, the backward gradient doesn't flow.
could you give me some more details?

[anjaeju]

Thanks you for Issues ikourbane.

I also encounter this issues here.
Even though, i changed the code you write above, the backward gradient doesn't flow.
could you give me some more details?

Fixed. I also changed all inplace operations. U should change the all inplace operation in code lines including the above codes.

[XGBoost]

@anjaeju could you please provide the corresponding modified code? Thank you very much!

[jungeun122333]

I also modified grad_mag and other things according to @anjaeju !

before:

grad_mag = torch.sqrt(grad_x_r**2 + grad_y_r**2) 
grad_mag += torch.sqrt(grad_x_g**2 + grad_y_g**2) 
grad_mag += torch.sqrt(grad_x_b**2 + grad_y_b**2)
        

after:

grad_mag = torch.sqrt(grad_x_r**2 + grad_y_r**2) + torch.sqrt(grad_x_g**2 + grad_y_g**2) + torch.sqrt(grad_x_b**2 + grad_y_b**2)