A PyTorch wrapper for masked convolutions
A PyTorch implementation of a thin wrapper for masked convolutions.
What are masked convolutions?
Similarly to partial convolutions, masked convolutions mask a part of the kernel, essentially ignoring data at specific locations. For an example, consider
a = [1, 2, 3, 4, 5]
assuming we have a convolution kernel
kernel = [1, 1, 1]
a would give us
a_conv = [6, 9, 12]
However, if we were to mask the convolution kernel with a mask
mask = [1, 0, 1]
masked convolving over
a would return
a_masked_conv = [4, 6, 8]
One use of masked convolutions is emulating skip-grams.
First, make sure you have PyTorch installed. This was tested on Python 3.8 and PyTorch 1.7.1. Further testing is needed to determine whether it works on a different setup - chances are it does. The recommended way to install this is through PyPi by running:
pip install masked-convolution
Other than that, you can clone this repository, and in its root directory (where
setup.py is located) run
pip install .
Every build, automatic benchmarks are run in order to determine how much overhead the implementation brings. The ordinary convolutions are used as a baseline, while the the performance of masked convolutions is described as a percentage of throughput of their respective baselines.
Keep in mind that these benchmarks are in no way professional, they only serve to give users a general idea. Their results greatly differ, so they should be taken with a grain of salt.
- Masked Convolution 1D: 85.29 % Convolution 1D throughput
- Masked Convolution 2D: 85.64 % Convolution 2D throughput
- Masked Convolution 3D: 97.79 % Convolution 3D throughput
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
|Filename, size||File type||Python version||Upload date||Hashes|
|Filename, size masked-convolution-0.3.1.tar.gz (6.4 kB)||File type Source||Python version None||Upload date||Hashes View|