search

capstone2019-neuralsearch / AC297r_2019_NAS

Harvard IACS Data Science Capstone: Neural Architecture Search (NAS) with Google
5 stars 1 forks source link

Random search algorithm for comparison #5

Closed dylanrandle closed 4 years ago

dylanrandle commented 4 years ago

Implement random search over the genotype object, such that we can use it for comparison

JiaweiZhuang commented 4 years ago

@jlaasri This discussion seems very useful: quark0/darts#3

JiaweiZhuang commented 4 years ago

Seems like a useful paper: Random Search and Reproducibility for Neural Architecture Search

jlaasri commented 4 years ago

See https://github.com/capstone2019-neuralsearch/darts/commit/ebe158ccd36dedc2ac69ba6153ddfdbbb274c894.

The script is still not totally general but it produces comparable genotypes to what we've produced so far.

jlaasri commented 4 years ago

See branch "random_search" of darts repo.

What we can do now: python random_search.py N --arg_example1 50 -arg_example2 300 ...

This will run "train.py" N times with N different random cells. All the arguments after "N" are directly passed to "train.py". In practice, this code will run the following command N times: python train.py --random --arg_example1 50 -arg_example2 300 ....

Also, when running train.py, the genotype is saved to genotype.arch for easy access.

jlaasri commented 4 years ago

Done with random_search*.py