torch-volpy

PyTorch tools for voltage imaging movie processing and signal extraction.

The package currently provides:

• HDF5-backed movie I/O through Movie
• Motion template building, translation estimation, and motion correction
• Gaussian high-pass filtering
• Summary image generation
• Cellpose-based segmentation support
• ALI and SpikePursuit signal extraction

Performance

The performance test shown below was run on a 512x400 movie acquired at 1000 fps. The test machine used an AMD Ryzen 7 5800X CPU and an NVIDIA GeForce RTX 2080 Ti GPU. Runtime will vary with movie length, selected processing steps, extraction method, storage speed, and CUDA availability.

Installation

pip install torch-volpy

Install the GUI extra for interactive movie viewing and ROI trace extraction:

pip install "torch-volpy[gui]"
torch-volpy-gui

For local development from this repository:

python -m pip install -e ".[gui,dev]"

Basic Imports

from torch_volpy.movie import Movie
from torch_volpy.motion import MotionCorrect, Template, Translation
from torch_volpy.filter import Filter
from torch_volpy.model import Summary
from torch_volpy.extraction import ALI, Spikepursuit
from torch_volpy.model import Cellpose

GUI

The PyQt GUI opens HDF5 movies (.h5/.hdf5, dataset defaults to movie) and TIFF stacks (.tif/.tiff). It provides frame playback, Cellpose ROI generation from a summary image, loading existing ROI masks (.tif, .npy, .npz, .h5/.hdf5), click-to-select ROI picking, and trace extraction with:

• Spikepursuit as the default extraction method
• ALI for cropped ROI activity localization
• a simple mean-ROI trace for quick inspection

The Cellpose ROI button follows the test workflow in _test/test_cellpose.py: build Summary(movie), stack [mean, mean, corr], and pass that image to Cellpose.build(...). The resulting labeled mask is shown as an overlay; click an ROI in Select mode to choose which label is used for trace extraction.

When opening a TIFF stack, the GUI first converts it to a sibling HDF5 file, runs motion correction into corrected_<name>.h5, and then displays the corrected HDF5 movie. A progress bar in the Movie panel reports conversion and motion-correction phases.

How To Build

python -m build

The build artifacts are written to dist/.