TorchGMM

TorchGMM allows to run Gaussian Mixture Models on single or multiple CPUs/GPUs. The repository is a fork from PyCave and LightKit, two amazing packages developed by Oliver Borchert that are not being maintained anymore. While PyCave implements additional models such as Markov Chains, TorchGMM focuses only on Gaussian Mixture Models.

The models are implemented in PyTorch and PyTorch Lightning, and provide an Estimator API that is fully compatible with scikit-learn.

For Gaussian mixture model, TorchGMM allows for 100x speed ups when using a GPU and enables to train on markedly larger datasets via mini-batch training. The full suite of benchmarks run to compare TorchGMM models against scikit-learn models is available on the documentation website.

Features

• Support for GPU and multi-node training by implementing models in PyTorch and relying on PyTorch Lightning

• Mini-batch training for all models such that they can be used on huge datasets

• Well-structured implementation of models

  • High-level Estimator API allows for easy usage such that models feel and behave like in scikit-learn
  • Medium-level LightingModule implements the training algorithm
  • Low-level PyTorch Module manages the model parameters

Getting started

Please refer to the documentation. In particular, the API documentation

Requirements

TorchGMM requires PyTorch to be installed. The installation instructions can be found on the PyTorch website.

TorchGMM is available via pip:

pip install torchgmm

Usage

If you've ever used scikit-learn, you'll feel right at home when using TorchGMM. First, let's create some artificial data to work with:

import torch

X = torch.cat([
    torch.randn(10000, 8) - 5,
    torch.randn(10000, 8),
    torch.randn(10000, 8) + 5,
])

This dataset consists of three clusters with 8-dimensional datapoints. If you want to fit a K-Means model, to find the clusters' centroids, it's as easy as:

from torchgmm.clustering import KMeans

estimator = KMeans(3)
estimator.fit(X)

# Once the estimator is fitted, it provides various properties. One of them is
# the `model_` property which yields the PyTorch module with the fitted parameters.
print("Centroids are:")
print(estimator.model_.centroids)

Due to the high-level estimator API, the usage for all machine learning models is similar. The API documentation provides more detailed information about parameters that can be passed to estimators and which methods are available.

GPU and Multi-Node training

For GPU- and multi-node training, TorchGMM leverages PyTorch Lightning. The hardware that training runs on is determined by the Trainer class. It's init method provides various configuration options.

If you want to run K-Means with a GPU, you can pass the options accelerator='gpu' and devices=1 to the estimator's initializer:

estimator = KMeans(3, trainer_params=dict(accelerator='gpu', devices=1))

Similarly, if you want to train on 4 nodes simultaneously where each node has one GPU available, you can specify this as follows:

estimator = KMeans(3, trainer_params=dict(num_nodes=4, accelerator='gpu', devices=1))

In fact, you do not need to change anything else in your code.

Implemented Models

Currently, TorchGMM implements two different models:

• GaussianMixture
• K-Means

Contribution

If you found a bug or you want to propose a new feature, please use the issue tracker.

License

TorchGMM is licensed under the MIT License.