将 pytorch 张量转换为 numpy，并重塑

Question

I have a pytorch tensor [100, 1, 32, 32] corresponding to batch size of 100 images, 1 channel, height 32 and width 32. I want to reshape this tensor to have dimension [32*10, 32*10], such that the images are represented as a 10x10 grid, with the first 10 images on row 1, and so on. How to achieve this?

Answer 1

I didn't understand your question completely but trying to address some of the concerns.

You have a tensor of shape [100, 1, 32, 32] that represents 100 images of shape [1, 32, 32] where num_channels = 1 , width = 32 , height = 32 .

First, since the images have only one channel, we can squeeze the channel dimension.

# image_tensor is of shape [100, 1, 32, 32]
image_tensor = image_tensor.squeeze(1) # [100, 32, 32]

We can organize the resulting tensors into 10 rows of 10 images as you described.

image_tensor = image_tensor.reshape(10, 10, 32, 32)

Now, converting the resulting tensor into a tensor of shape [32*10, 32*10] sounds something is wrong. But, let's do that wrong thing and see what we end up with.

image_tensor = image_tensor.permute(2, 0, 3, 1) # [32, 10, 32, 10]

After permutation, we get a tensor of shape [width, num_rows, height, num_img_in_a_row] . Then finally we can reshape to get the desired tensor.

image_tensor = image_tensor.reshape(32*10, 32*10)

So, the final tensor is of shape [width * num_rows, height * num_img_in_a_row] . Do you really want this? I am not sure how to interpret the resulting tensor!!

Answer 2

Update

More efficient and shorter version. To avoid using for-loop, we can permute a first.

import torch
a = torch.arange(9*2*2).view(9,1,2,2)
b = a.permute([0,1,3,2])
torch.cat(torch.split(b, 3),-1).view(6,6).t()
# tensor([[ 0,  1,  4,  5,  8,  9],
#         [ 2,  3,  6,  7, 10, 11],
#         [12, 13, 16, 17, 20, 21],
#         [14, 15, 18, 19, 22, 23],
#         [24, 25, 28, 29, 32, 33],
#         [26, 27, 30, 31, 34, 35]])

Original Answer

You can use torch.split and torch.cat to implement it.

import torch
a = torch.arange(9*2*2).view(9,1,2,2)

Assuming we have a tensor, which is a mini version of your original tensor. And it looks like,

tensor([[[[ 0,  1],
          [ 2,  3]]],
        [[[ 4,  5],
          [ 6,  7]]],
        [[[ 8,  9],
          [10, 11]]],
        [[[12, 13],
          [14, 15]]],
        [[[16, 17],
          [18, 19]]],
        [[[20, 21],
          [22, 23]]],
        [[[24, 25],
          [26, 27]]],
        [[[28, 29],
          [30, 31]]],
        [[[32, 33],
          [34, 35]]]])

Each 2x2 sub-matrix can be seen as one image. What you want to do is stacking the first three images to one row, next three images to the second row, and last three images to the third row. The "row" has actually two dim due to the 2x2 sub-matrix.

three_parts = torch.split(a,3)

torch.cat(torch.split(three_parts[0],1), dim=-1)

#tensor([[[[ 0,  1,  4,  5,  8,  9],
#          [ 2,  3,  6,  7, 10, 11]]]])

Here we only take the first part.

torch.cat([torch.cat(torch.split(three_parts[i],1),-1) for i in range(3)],0).view(6,6)
# tensor([[ 0,  1,  4,  5,  8,  9],
#         [ 2,  3,  6,  7, 10, 11],
#         [12, 13, 16, 17, 20, 21],
#         [14, 15, 18, 19, 22, 23],
#         [24, 25, 28, 29, 32, 33],
#         [26, 27, 30, 31, 34, 35]])

Answer 3

You can use make_grid() :

x = torchvision.utils.make_grid(x, nrow=10, padding=0)