Modified CANTO

This builds upon the existing CANTO framework by introducing an alternative optimizer for asynchronous sgd which is ADMM (alternating direction method of multipliers) which is expected to work optimally in a distributed environment. This is implemented as reference to the Training Neural Networks Without Gradients: A Scalable ADMM Approach paper.

Check out the detailed description here and the slides associated here.

Setup

To run a minimalistic locally working CANTO:

• ./deploy.sh Note: Check logs of containers to view the output