From channels to tiles

1

Hi,

I’m new to Pytorch and I was wondering how to write a code that efficiently transforms channels into tiles. So from shape [c, n/(c/2), n/(c/2)] to shape [1, n, n], but respecting the tiles ordering, which .reshape does not do. I’ll explain better with an example:

Let’s assume I have a tensor with shape input.shape=torch.Size([4, 64, 64]), and I would like an output tensor with shape output.shape=torch.Size([1, 128, 128]), but where each channel of the first input tensor is a tile in a quarter of the output tensor, e.g. output[64:, :64]==input[1, 64, 64].

But I don’t want to use assignation or for loops. The code would have to be general enough to take also e.g. an input tensor of shape [16, 32, 32], and output a tensor with size [1, 128, 128], where the 16 channels are ordered as tiles.

Thank you!

2

To answer this question, it might be helpful if you wrote a naive for-loop version for testing purposes and to clarify how you want to rearrange the underlying data.

3

Hi @eqy, here’s an example.

Let’s say I have an input of 4 channels all of shape [128, 128]. In this simple example, the channels all look the same. I would like to rearrange each channel like a tile in a mosaic with shape [128x2, 128x2].

output = torch.zeros(128*2, 128*2)
output[:128, :128] = input[0]
output[:128, 128:] = input[1]
output[128:, :128] = input[2]
output[128:, 128:] = input[3]

Since I am a new user I can only embed a single image. So here’s the output I would like. And each of the 4 images was a channel in input.
2b8454a0-bd3f-4a83-b2b3-d9950a9cfcf5

I would like the code to be flexible enough to choose based on input channels, how to rearrange the tiles in the output. Ideally, given input with shape [C, H, W] (with H==W), I’d like the output to have shape [1, HxC/2. WxC/2]. But .reshape or .view do not rearrange the pixels in the way that I’d like them to be.

Thanks for your help!

4

Here’s an example that more or less achieves what you want I think

import torch
input = torch.randn(4, 128, 128)
output = torch.zeros(128*2, 128*2)
output[:128, :128] = input[0]
output[:128, 128:] = input[1]
output[128:, :128] = input[2]
output[128:, 128:] = input[3]

# steps
# break input channels into 2x2
# tile means outer-channel split is H, inner channel split is W
# contiguous is not necessary unless you want the underlying data order to be changed
output2 = input.reshape(2, 2, 128, 128).permute(0, 2, 1, 3).reshape(128*2, 128*2).contiguous()
print(torch.allclose(output, output2))