Question about Retraining the WAE model

joffery / M-ADA

The Pytorch implementation for "Learning to Learn Single Domain Generalization" (CVPR 2020)

https://arxiv.org/abs/2003.13216

143 stars 25 forks source link

Question about Retraining the WAE model #5

Open tkong1998 opened 3 years ago

tkong1998 commented 3 years ago

Hi Fengchun,

Could you explain the intuition behind retraining a wae model on the latest domain augmentation instead of keeping the wae mode frozen throughout the pipeline? The augmented domain is different from the source domain, by retaining the wae on the augmented data, wouldn't it affect the performance?

Thank you so much!

Best, Monique

joffery commented 3 years ago

Hi @tkong1998,

This training strategy enables the model to create domain augmentations gradually away from the source. The idea is similar to the curriculum learning. And these domain augmentations are bounded by the worst-case constraint, making themselves not too far away from the source.

Best, Fengchun