gaocegege / darts_asha

Code release to reproduce ASHA experiments from "Random Search and Reproducibility for NAS."

This repo contains the code to run the experiments with ASHA on the two neural architecture search benchmarks presented in "Random Search and Reproducibility for NAS." ASHA is short for the Asynchronous Successive Halving algorithm, which was presented in this paper.

ASHA

Please follow directions in this repo to make sure you have the forked DARTS repo with the correct python packages and also have the data downloaded into the right directories.

You will need the python package mpi4py in order to run ASHA on multiple GPUs.

Make sure the path for darts and the data are correct in the benchmark files:
darts_asha/benchmarks/nas_search/cnn/darts/darts_wrapper.py
darts_asha/benchmarks/nas_search/cnn/darts/darts_trainer.py
darts_asha/benchmarks/nas_search/ptb/darts/darts_wrapper.py
darts_asha/benchmarks/nas_search/ptb/darts/darts_trainer.py
Also, make sure darts_asha is in your PYTHONPATH.

To run ASHA using multiple GPUs on the same machine, issue the following command from darts_asha/searchers:
mpirun -np [# GPUs + 1] -output-filename [mpi filename root] python mpi_hyperband.py -a asha -m [darts_cnn/dart_ptb] -d [cifar10/ptb] -o [output dir] -s [seed] -r 1 -R 300 -t [# hours]

You can also run this on a cluster of GPUs by submitting a job through qsub.

The random seeds used for ASHA in the paper are available in this spreadsheet. Note that this repo contains the fix for non-deterministic CNN training, while our experiments were conducted prior to this fix using the original random seeding in DARTS.