Skip to content

Latest commit

 

History

History
20 lines (15 loc) · 981 Bytes

README.md

File metadata and controls

20 lines (15 loc) · 981 Bytes

Automix

This repository contains the code to use Fast AutoAugment algorithm to generate and pick automatic augmentations for a given task.

For the set of combinations, we used {4 EDA Augmentations, inter lada, intra lada, adverserial augmentation}

The way FAA Algorithm works is as follows -

  • Create a model (M) on Train, and formulate a method to pick augmentations
  • Repeat
    • Sample augmentations and created augmented val data
    • Find loss of augmented val data on M
    • Save the augmentations with lowest loss (call this ‘aug val loss’)
  • Best augmentations are the one with least loss
  • Using saved augmentations, create the final model M’ on Train.
  • Use M’ to find out ‘final val accuracy’ and test accuracy

Directions to run the code

  • yahoo_with_mixtext/train_*/ - Train a FAA policy using the chosen augmentation on the yahoo dataset.
  • yahoo_with_mixtext/evaluate_any_model.py - Once train generates a policy, run this to evaluate the policy