torch-runstats 0.1.0

Running/online statistics for PyTorch

pytorch_runstats

Running/online statistics for PyTorch.

torch_runstats implements memory-efficient online reductions on tensors.

Notable features:

• Arbitrary sample shapes beyond single scalars
• Reduction over arbitrary dimensions of each sample
• "Batched"/"binned" reduction into multiple running tallies using a per-sample bin index. This can be useful, for example, in accumulating statistics over samples by some kind of "type" index or for accumulating statistics per-graph in a pytorch_geometric-like batching scheme. (This feature uses and is similar to torch_scatter.)

Note: the implementations currently heavily uses in-place operations for peformance and memory efficiency. This probably doesn't play nice with the autograd engine — this is currently likely the wrong library for accumulating running statistics you want to backward through. (See TorchMetrics for a possible alternative.)

For more information, please see the docs.

Install

torch_runstats requires PyTorch and torch_scatter, but neither is specified in install_requires for pip since both require manual installation for correct CUDA compatability.

License

pytorch_runstats is distributed under an MIT license.