According to the each data of the batch ,normalize the output x

autograd
1

I want to normalize the output,for example,my batch size is 64 ,I will normalize for each line

def x_normalization(x):
       num_pic,_=x.data.shape
       for i in range(num_pic):
            x[i,:]=x[i,:]/torch.max(x[i,:])    #normalize for each line
       return x

I call this function in
image
I get a error
one of the variables needed for gradient computation has been modified by an inplace operation

2

Hi,

The problem is that you modify x in place while it’s value is needed to compute gradients.
You can replace your function with:

def x_normalization(x):
       x = x / x.max(0, keepdim=True)[0]
       return x
3

excuse me. I also encounter this problem when I run the code below in a forward() function(code is in the image). If I comment the guess_img_dist_arr[n]=(torch.norm(guess_img[y][x]-guess_img[y_co][x_co]))
It will not raise the error, but get a wrong result.
could you tell me how to fix this?

27
271480×754 88.2 KB

4

Hi,

You should not modify a single Variable inplace multiple times.
In your case, one wait to solve this would be:

for i in range(k_clusters):
  # some code

  guess_img_dist_list = []
  for n in range(length):
    # some code
    guess_img_dist_list.append(torch.norm(...))

  guess_img_dist_arr = torch.cat(guess_img_dist_list, 0)
5

Thanks a lot!
You saved my day!

7

thanks, that is feasible