nn

PyTorch implementation of iterative winners-take-all

This module is an experimental PyTorch implementation of iWTA. It includes additional metrics that help you to build intuition upon our work. It's also faster, if you have CUDA support.

Here is a snapshot of running the clustering experiment.

Note, to follow PyTorch batch logic, we swap the axis in weights, compared to the Numpy implementation.

Quick start

Python 3.6+ is required.

1. Install PyTorch

   conda install pytorch torchvision cpuonly -c pytorch
   

2. Install the requirements

   pip install git+https://github.com/dizcza/pytorch-mighty.git#egg=pytorch-mighty
   pip install matplotlib
   pip install networkx
   

3. Start a Visdom server by running the following command in a new terminal window

   python -m visdom.server
   

4. Run the experiment on your choice from the project root directory

   python nn/clustering.py
   

5. Navigate to http://localhost:8097 to see the training progress as in the picture above.