haralickt
haralickt is a PyTorch-based library for GPU-accelerated computation of Haralick texture features (GLCM). It supports all 13 descriptors, sliding windows, block tiling and batching.
Timeline (last 90 days)
haralickt is a pure PyTorch implementation library that computes all 13 Haralick descriptors on the GPU (also supporting multithreaded CPU), accelerated through vectorized tensor operations and tiled chunking.
Compared with single-threaded mahotas, haralickt's CPU mode achieved an average speedup of 16.7x, ranging from 9.1x to 67.7x, in tests of 27 parameter combinations on 256x256 images with 32 gray levels.
On image sizes of 512x512 and above, haralickt GPU's speedup relative to 10-thread OpenMP mahotas is roughly between 20 and 60 times.
The maximum numerical difference between haralickt and mahotas results is 2.02 × 10⁻⁴, originating from float32 rounding error in Diff Variance.
haralickt is a pure Python + PyTorch (>=2.0) implementation, and installation requires no C compiler toolchain.
haralickt can be integrated with torchvision and MONAI, making it suitable for medical imaging deep learning pipelines.
haralickt provides a sliding-window API `haralick(image, num_levels=32, window_size=32, stride=8, device="cuda")`, returning a tensor of shape (H, W, 13).
haralickt provides `features.haralick()` as a drop-in replacement for mahotas, returning a NumPy array of shape (4, 13).
haralickt's license is MPL-2.0.
haralickt can be installed via `pip install haralickt`.
2 more timeline events
All Facts (12)
haralickt is a pure PyTorch implementation library that computes all 13 Haralick descriptors on the GPU (also supporting multithreaded CPU), accelerated through vectorized tensor operations and tiled chunking.
50%UnverifiedCompared with single-threaded mahotas, haralickt's CPU mode achieved an average speedup of 16.7x, ranging from 9.1x to 67.7x, in tests of 27 parameter combinations on 256x256 images with 32 gray levels.
50%UnverifiedOn image sizes of 512x512 and above, haralickt GPU's speedup relative to 10-thread OpenMP mahotas is roughly between 20 and 60 times.
50%UnverifiedThe maximum numerical difference between haralickt and mahotas results is 2.02 × 10⁻⁴, originating from float32 rounding error in Diff Variance.
50%Unverifiedharalickt is a pure Python + PyTorch (>=2.0) implementation, and installation requires no C compiler toolchain.
50%Unverifiedharalickt can be integrated with torchvision and MONAI, making it suitable for medical imaging deep learning pipelines.
50%Unverifiedharalickt provides a sliding-window API `haralick(image, num_levels=32, window_size=32, stride=8, device="cuda")`, returning a tensor of shape (H, W, 13).
50%Unverifiedharalickt provides `features.haralick()` as a drop-in replacement for mahotas, returning a NumPy array of shape (4, 13).
50%Unverifiedharalickt's license is MPL-2.0.
50%Unverifiedharalickt can be installed via `pip install haralickt`.
50%Unverifiedharalickt's current public benchmarks focus on 256x256 images and a limited set of parameter combinations; performance at larger sizes and more gray levels still needs community verification.
50%UnverifiedFor processing small images, haralickt's speedup relative to CPU libraries is limited; large images and batch-processing scenarios are where the GPU truly pulls ahead.
50%