search

google-deepmind / distribution_shift_framework

This repository contains the code of the distribution shift framework presented in A Fine-Grained Analysis on Distribution Shift (Wiles et al., 2022).
Apache License 2.0
80 stars 8 forks source link

Reproducing Learned Augmentation (CycleGAN) #4

Closed ysharma1126 closed 1 year ago

ysharma1126 commented 1 year ago
  1. Was CycleGAN trained on the true, uncorrelated distribution, implying the same model was used for each dataset averaged over in Fig. 3?
  2. Any plans for release?
oawiles commented 1 year ago
  1. No, CycleGAN was trained on the train set, so it inherited the biases of the training set. We trained a different model for each biased training set so a different model was used for each result (and seed).
  2. At the moment we are not planning on this.