TimeSformer-pytorch

Why we included this project

TimeSformer-pytorch is a compact, faithful implementation of the Divided Space-Time Attention variant from Facebook AI's video classification paper. The core idea is to run attention along the time axis first and then the spatial axis, which is what sets the architecture apart from earlier video transformers. The package installs with pip and gives you a single model class with knobs for depth, heads, patch size, and frame count, plus a frame mask for batches that mix videos of different lengths. It is a research codebase rather than a full training pipeline, so you will need to bring your own data loading and training loop. That makes it most useful for prototyping, reproducing the paper, or adapting the architecture into a larger video project.