Masking tensor of same shape in PyTorch

user_6210807 0
Viewed 14177

Given an array and mask of same shapes, I want the masked output of the same shape and containing 0 where mask is False.

For example,

# input array
img = torch.randn(2, 2)
print(img)
# tensor([[0.4684, 0.8316],
#        [0.8635, 0.4228]])
print(img.shape)
# torch.Size([2, 2])

# mask
mask = torch.BoolTensor(2, 2)
print(mask)
# tensor([[False,  True],
#        [ True,  True]])
print(mask.shape)
# torch.Size([2, 2])

# expected masked output of shape 2x2
# tensor([[0, 0.8316],
#        [0.8635, 0.4228]])

Issue: The masking changes the shape of the output as follows:

#1: shape changed
img[mask]
# tensor([0.8316, 0.8635, 0.4228])
pytorch python
3 Answers

Simply type-cast your boolean mask to an integer mask, followed by float to bring the mask to the same type as in img. Perform element-wise multiplication afterwards.

masked_output = img * mask.int().float()

user_122669230

One of the ways I found to solve it was:

img[mask==False] = 0

or using

img[~mask] = 0

It'll change the img itself.

user_62108070

The most straight forward way would be creating another tensor to handle it.

import torch

def generate_masked_tensor(input, mask, fill=0):
    masked_tensor = torch.zeros(input.size()) + fill
    masked_tensor[mask] = input[mask]
    return masked_tensor

if __name__ == "__main__":
    img = torch.randn(2, 2)
    mask = torch.tensor([False, True, True, False]).bool().view(2, 2)
    masked_img = generate_masked_tensor(img, mask)
    print (masked_img)

The output:

tensor([[0.0000, 0.8028],
        [1.5411, 0.0000]])
user_66241010
Related