torchpercentile

Percentile computation for pytorch

Presentation

This repository implements a Percentile class that overrides torch.autograd.Function, enabling percentile computation on the GPU for Pytorch.

class Percentile(torch.autograd.Function):
    def forward(ctx, input, percentiles)

This function returns the percentiles of the input, computed along the first dimension.

It works similarly to the numpy.percentile function, except that we don't accept a selection of axis, computations are done on the first axis only. Note however that the shape of the tensor may be arbitrary.

Parameters for Percentile.forward

• input : a Pytorch Tensor: The data for which percentiles will be computed

• percentiles : a Pytorch Tensor, or some type that may be called as a parameter to torch.tensor The different percentiles (between 0 and 100) to compute for the data

Results

a Pytorch tensor with the same shape as input except the first dimension, whose length is that of percentiles

Installation

Type pip install -e .

Usage

Try out python test.py in the examples folder.

calling percentile on a tensor of dimension [10000    10    50]
numpy: 154.368ms,
CPU: 604.269ms, error: : 0.000002%.
GPU: 54.679ms, error: 0.000003%.

This implementation is worth it only on the GPU compared to the numpy version, but note that backward is implemented.