===============================================================

Python module to train weighted GMMs using CUDA (via CUDAMat)

If you use this work or draw inspiration from it, kindly cite this using:

@INPROCEEDINGS{8793587,
  author={Bentz, William and Dhanjal, Sahib and Panagou, Dimitra},
  booktitle={2019 International Conference on Robotics and Automation (ICRA)}, 
  title={Unsupervised Learning of Assistive Camera Views by an Aerial Co-robot in Augmented Reality Multitasking Environments}, 
  year={2019},
  volume={},
  number={},
  pages={3003-3009},
  keywords={Visualization;Task analysis;Cameras;Robot vision systems},
  doi={10.1109/ICRA.2019.8793587}}

• Dependencies
• Installation
• Example usage
• Documentation

• Not Windows (only tested on Linux and Mac)
• CUDA 6.0+ (only tested with 6.0)
• numpy
• CUDAMat, avaiable here: https://github.com/cudamat/cudamat.git
• future: http://python-future.org/index.html
• nose (optional, for running tests)

Clone wgmm and CUDAMat in local install path:

cd ${INSTALL_PATH}
git clone https://github.com/sahibdhanjal/Weighted-Expectation-Maximization.git
git clone https://github.com/cudamat/cudamat.git

Compile and install CUDAMat:

cd ${INSTALL_PATH}/cudamat
sudo python setup.py install

Run CUDAMat tests (optional, requires nose):

cd ${INSTALL_PATH}/cudamat
nosetests

Run wgmm tests (optional, requires nose):

cd ${INSTALL_PATH}/wgmm
nosetests

Install wgmm:

cd ${INSTALL_PATH}/wgmm
sudo pip install .

import wgmm.gpu as wgmm

X = some_module.load_training_data()

# N - training examples
# D - data dimension
# K - number of GMM components
N, D = X.shape
K = 128

wgmm.init()
gmm = wgmm.GMM(K,D)

thresh = 1e-3 # convergence threshold
n_iter = 20 # maximum number of EM iterations
init_params = 'wmc' # initialize weights, means, and covariances

# train GMM
gmm.fit(X, thresh, n_iter, init_params=init_params)

# retrieve parameters from trained GMM
weights = gmm.get_weights()
means = gmm.get_means()
covars = gmm.get_covars()

# compute posteriors of data
posteriors = gmm.compute_posteriors(X)

For testing it on randomly generated data, you can run python gmmtest.py from the examples folder. On setting appropriate flags, a data file and a plot (only for 1D/3D case) will also be generated.

Documentation for GGMM by Eric Battenberg available here