DCTorch: Discrete Cosine Transforms for PyTorch

0.1.2 · active · verified Fri Apr 17

DCTorch is a Python library providing fast discrete cosine transform (DCT) and inverse discrete cosine transform (IDCT) implementations optimized for PyTorch tensors. It enables efficient frequency domain analysis and manipulation within deep learning models, supporting 2D and 3D transforms. The current version is 0.1.2, and it appears to be actively maintained with recent commits.

Common errors

```
RuntimeError: Expected object of scalar type Float but got scalar type Long for argument 'input'
```
cause The input tensor provided to a dctorch function has an integer data type (e.g., `torch.long`), but dctorch expects floating-point numbers.

fix Convert the input tensor to a floating-point type: `x = x.to(torch.float32)` or `x = x.float()`.
```
NotImplementedError: The operator 'aten::fft_rfft' is not currently implemented for the MPS device.
```
cause dctorch internally relies on PyTorch's FFT operations, which may not be fully supported on all PyTorch devices (e.g., Apple's MPS backend for certain FFT functions).

fix Run the operation on a supported device (CPU or CUDA). Move your tensor to a different device: `tensor = tensor.to('cpu')` or `tensor = tensor.to('cuda')`.
```
AttributeError: module 'dctorch' has no attribute 'dct'
```
cause The user is trying to call a generic `dct` function, but `dctorch` provides specific functions for 2D (`dct_2d`) and 3D (`dct_3d`) transforms.

fix Use the correct function name: `dctorch.dct_2d()` for 2D transforms or `dctorch.dct_3d()` for 3D transforms.

Warnings

gotcha Performance heavily depends on PyTorch's underlying CUDA/CPU implementation. Ensure PyTorch is correctly installed and configured for your hardware for optimal speed, especially with large tensors.
Fix: Verify your PyTorch installation and device setup. Ensure CUDA drivers are up-to-date if using a GPU. Always move tensors to the correct device (e.g., `tensor.to('cuda')`).
gotcha DCTorch functions assume input tensors are floating-point types (e.g., `torch.float32`, `torch.float64`). Passing integer tensors will result in type errors or unexpected behavior.
Fix: Convert input tensors to a float type using `.to(torch.float32)` or `.to(torch.float64)` before passing them to `dctorch` functions.
gotcha While DCT and IDCT are often used with square inputs in traditional signal processing, `dctorch` functions `dct_2d` and `idct_2d` will work with rectangular inputs. However, ensure the spatial dimensions (height, width) are consistent for corresponding DCT and IDCT operations to correctly reconstruct the original signal.
Fix: No specific fix, but users should be aware of this behavior and ensure input dimensions are managed appropriately for their use case.

Install

pip install dctorch Install dctorch

Imports

dct_2d
```
from dctorch import dct_2d
```
idct_2d
```
from dctorch import idct_2d
```
dct_3d
```
from dctorch import dct_3d
```
idct_3d
```
from dctorch import idct_3d
```

Quickstart

This quickstart demonstrates how to perform 2D Discrete Cosine Transform (DCT) and its inverse (IDCT) on a PyTorch tensor. It includes device selection for GPU acceleration and verifies the reconstruction accuracy.

import torch
from dctorch import dct_2d, idct_2d

# Determine device (CPU or CUDA if available)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print(f"Using device: {device}")

# Create a random tensor (Batch size, Channels, Height, Width)
x = torch.randn(1, 3, 224, 224, device=device)
print(f"Original tensor shape: {x.shape}")

# Perform 2D Discrete Cosine Transform
y = dct_2d(x)
print(f"DCT transformed tensor shape: {y.shape}")

# Perform Inverse 2D Discrete Cosine Transform
x_recon = idct_2d(y)
print(f"Reconstructed tensor shape: {x_recon.shape}")

# Verify reconstruction accuracy
reconstruction_error = torch.norm(x - x_recon).item()
print(f"Reconstruction error (L2 norm): {reconstruction_error:.6f}")

# Note: Error is typically very small due to floating point precision
assert reconstruction_error < 1e-4

view raw JSON →