Discrete Cosine Transform Using `torch.rfft`

1

torch.rfft lacks of doc and it’s hard to understand how to use it.
Actually, I’d like to use this function to implement a fast discrete cosine transform (DCT).
Please let me know if you have DCT implementations (any differentiable in PyTorch) or concrete example for torch.rfft (especially, 2D case).
Thanks.

1 Like
2

I guess you should look into the following link -
https://pytorch.org/tutorials/advanced/numpy_extensions_tutorial.html
This link explains how to incorporate scipy functions into the computational graph.

3

Thank you for the comment. I saw the page but it is not what I really want because it converts torch tensors to numpy arrays, or vice versa.

4

Here is something which you can look forward to, as you have asked this provides example also. However when I went over the examples I found few typos but that you can easily see. I am suggesting this because it will give you atleast a direction to move forward if not solve your problem completely.

5

I also visited the repository and I found it had some issues. For example, the transforms would be slower when applying nn.DataParallel even if one had more than 1 GPU. Moreover, it seems that this repo hasn’t been updated since PyTorch 0.4.0.

6

I saw that too. and yes there were issues posted regarding nn.DataParallel but I thought you can make modifications to these to fit your need.

7

Hi everyone, I just wrote a library doing exactly this, you might want to check it out:

Would love to hear your feedback!

2 Likes
8

assert (torch.abs(x - y)).sum() < 1e-10 # x == y within numerical tolerance
numerical tolerance error cannot satisfy 1e-10, far beyond this, about 1e-3

1 Like