/mnist_allreduce

Primary LanguagePython

Tutorial to implement different versions of all-reduce

Create and activate conda environment

  1. Install Anaconda python virtual environment manager (miniconda is recommended)
  2. Create conda environment and install required packages conda env create -n allreduce_env -f environment.yml
  3. Activate the environment conda activate allreduce_env

Install PyCharm

https://www.jetbrains.com/pycharm/download/#section=windows

Modify allreduce implementation

Search for # Modify gradient allreduce here and update code there. Replace star-reduce code with ring-allreduce as per:

  1. https://towardsdatascience.com/distributed-deep-learning-with-horovod-2d1eea004cb2
  2. https://towardsdatascience.com/visual-intuition-on-ring-allreduce-for-distributed-deep-learning-d1f34b4911da

Run tests to make sure that weight all-reduce gives correct results

python test_dataparallel.py

Run neural network training on MNIST datastet

Training is run for the reference and dataparallel models simultaneously

python train_model.py