DCT (Discrete Cosine Transform) for pytorch

This library implements DCT in terms of the built-in FFT operations in pytorch so that
back propagation works through it, on both CPU and GPU. For more information on
DCT and the algorithms used here, see
Wikipedia and the paper by
J. Makhoul. This
StackExchange article
might also be helpful.

The following are currently implemented:

• 1-D DCT-I and its inverse (which is a scaled DCT-I)
• 1-D DCT-II and its inverse (which is a scaled DCT-III)
• 2-D DCT-II and its inverse (which is a scaled DCT-III)
• 3-D DCT-II and its inverse (which is a scaled DCT-III)

Install

pip install torch-dct

Requires torch>=0.4.1 (lower versions are probably OK but I haven't tested them).

You can run test by getting the source and run pytest. To run the test you also
need scipy installed.

Usage

pythonCopy
import torch
import torch_dct as dct

x = torch.randn(200)
X = dct.dct(x)   # DCT-II done through the last dimension
y = dct.idct(X)  # scaled DCT-III done through the last dimension
assert (torch.abs(x - y)).sum() < 1e-10  # x == y within numerical tolerance

dct.dct1 and dct.idct1 are for DCT-I and its inverse. The usage is the same.

Just replace dct and idct by dct_2d, dct_3d, idct_2d, idct_3d, etc
to get the multidimensional versions.