BayesWatch / pytorch-experiments-template

A pytorch based classification experiments template
GNU General Public License v3.0
5 stars 1 forks source link

Automating experiment config generation for hyper-parameter searching #15

Closed AntreasAntoniou closed 3 years ago

AntreasAntoniou commented 3 years ago

This is something I have been working on for a few years, and it has reached a good level. I'll adapt it and contribute it here. Open to your thoughts.

The basic idea is:

  1. You have your arg parser with your default arguments, their keys and values.
  2. You want a way to select a subset and search over some space you want to define.
  3. You use a script where all you need to define is a nice tuple of the variable names you want to search overs as well as the values you want to search over. The script finds all possible combinations, and generates a json named automatically using the default experiment name + the arguments that vary.
  4. You then use a script that picks up any templates in a folder with 'experiment_script_templates', e.g. one for CDT cluster, one for GPU boxes like edwards etc and generates experiment scripts by combining them with all the available experiment json files from 3.

With a measly 2 clicks you get every single combo of the stuff you want to search over in a script format ready to run on the machine of your choice.

AntreasAntoniou commented 3 years ago

Done.