yeshg / RayNAS

Tool for Neural Architecture Search using the DARTS algorithm and Population-Based-Training from Ray Tune

Ray-NAS

Scaling and Distributing Neural Architecture Search (NAS) Algorithms with the distributed machine learning framework Ray.

Algorithms

• DARTS
  • uses official pytorch implementation
  • CNN only
• ENAS
  • uses unofficial pytorch reimplimentation
  • RNN only
• Random NAS
  • Same search space as ENAS, no RNN controller
  • CNN only

Running experiments

1. Basics

Can search for CNN and RNN architectures from the same entry-point: main.py

First choose algorithm from darts, enas, or random, which correspond to differentiable architecture search, efficient neural architecture search, and a simple random sample based approach which simplifies NAS to hyperparameter tuning.

To search for CNN architecture for cifar10 with DARTS,

python main.py darts cnn --dataset cifar10 --layers 2 --cuda

To search for RNN architecture for ptb with ENAS,

python main.py enas rnn --num_blocks 4 --cuda

To search for CNN architecture for cifar10 with simple random NAS,

python main.py random cnn --dataset cifar10 --cuda

2. Logging details / Monitoring live training progress

Monitor runs in ray dashboard, or by launching tensorboard in the experiment directory created by Ray Tune:

tensorboard --logdir="~/ray_results/exp/"

3. Visualizing Searched Architectures

Find the location of the run which generated the desired architecture and run the entry-point in visualize mode:

python main.py viz --load <path_to_trial> --viz

General Arguments for All NAS algorithms

TBD

To Do

• DARTS
  • RNN implementation
  • Results
• ENAS
  • CNN implementation
  • Results
• RandomNAS
  • Search Space expansion for CNN
  • RNN implementation
  • Results