/torchfold

Tools for PyTorch

Primary LanguagePythonApache License 2.0Apache-2.0

PyPi version DOI

TorchFold

Blog post: http://near.ai/articles/2017-09-06-PyTorch-Dynamic-Batching/

Analogous to TensorFlow Fold, implements dynamic batching with super simple interface. Replace every direct call in your computation to nn module with f.add('function name', arguments). It will construct an optimized version of computation and on f.apply will dynamically batch and execute the computation on given nn module.

Installation

We recommend using pip package manager:

pip install torchfold

Example

    f = torchfold.Fold()
   
    def dfs(node):
        if is_leaf(node):
            return f.add('leaf', node)
        else:
            prev = f.add('init')
            for child in children(node):
                prev = f.add('child', prev, child)
            return prev

    class Model(nn.Module):
        def __init__(self, ...):
            ...

        def leaf(self, leaf):
            ...

        def child(self, prev, child):
            ...

    res = dfs(my_tree)
    model = Model(...)
    f.apply(model, [[res]])

To cite this repository in publications:

@misc{illia_polosukhin_2018_1299387,
  author       = {Illia Polosukhin and
                  Maksym Zavershynskyi},
  title        = {nearai/torchfold: v0.1.0},
  month        = jun,
  year         = 2018,
  doi          = {10.5281/zenodo.1299387},
  url          = {https://doi.org/10.5281/zenodo.1299387}
}