dugzzuli / pytorch-nmfs

A pytorch package for non-negative matrix factorization.

Non-negative Matrix Fatorization in PyTorch

PyTorch is not only a good deep learning framework, but also a fast tool when it comes to matrix operations and convolutions on large data. A great example is PyTorchWavelets.

In this package I implement basic NMF and NMFD module minimizing beta-divergence using multiplicative update rules, and is based on torch.nn.Module so the models can be moved freely among CPU/GPU devices. Part of the multiplier is obtained via torch.autograd so the amount of codes is reduced and easy to maintain (only the denominator is calculated).

The interface is similar to sklearn.decomposition.NMF with some extra options.

Usage

Here is a short example of decompose a spectrogram.

import torch
import librosa
from torchnmf import NMF
from torchnmf.metrics import KL_divergence

y, sr = librosa.load(librosa.util.example_audio_file())
y = torch.from_numpy(y)
windowsize = 2048
S = torch.stft(y, windowsize, window=torch.hann_window(windowsize)).pow(2).sum(2).sqrt().cuda()

R = 8   # number of components

net = NMF(S.shape, n_components=R).cuda()
# run extremely fast on gpu
_, V = net.fit_transform(S)      # fit to target matrix S
print(KL_divergence(V, S))        # KL divergence to S

A more detailed version can be found here, which redo this example with NMFD.

Compare to sklearn

The barchart shows the time cost per iteration with different beta-divergence. It is clear that pytorch-based NMF is faster than scipy-based NMF (sklearn) when beta != 2 (Euclidean distance), and run even faster when computation is done on GPU. The test is conducted on a Acer E5 laptop with i5-7200U CPU and GTX 950M GPU.

Installation

Using pip:

pip install git+http://github.com/yoyololicon/pytorch-NMFs

Or clone this repo and do:

python setup.py install

Requirements

• PyTorch == 0.4.1

Tips

• If you notice significant slow down when operating on CPU, please flush denormal numbers by torch.set_flush_denormal(True).
• Replace zeros in target matrix with small value to avoid overflow.

TODO

• Support sparse matrix.
• Regularization.
• NNDSVD initialization.