AutoHAS

A complete, working PyTorch implementation of AutoHAS

What it does

At a high level, it works as follows:

• instantiate a categorical-policy based RL controller controller
• while controller has not converged:
  • sample candidate_model(s), candidate_hyperparameter(s) from controller
  • for each candidate_model, training on some training set using each candidate_hyperparameter as hyperparameters
  • create copy candidate_model_copy of each candidate_model and calculate each copy's reward_signal (probably as validation accuracy)
  • update controller policies using each candidate_model_copy's reward_signal for policy gradient

As explained in the paper, we evaluate the validation accuracy of copies and not directly on the candidate models themselves to decouple the process of optimizing the shared network weights of the candidate models and the controller optimization step.

How to use it

The main entry point is train_controller.py which contains all the logic for training an RL controller as specified in the AutoHAS paper. Specific requirements are included in requirements.txt, but you also need to do pip install -e . to install the AutoHAS module (called module) itself.

If you just want to see it working, you can just do python train_controller.py. By default, this trains a controller to create child models for MNIST. You can see training progress using Tensorboard (logs are saved to the runs/ directory by default) with tensorboard --logdir=runs.

Identifying muon track centroids

To use this technique to identify centroids in muon tracks from light and charge signal, you have to first install SparseConvNet.