Add test PBA search script, remove dataset assumptions

arcelien / pba

Efficient Learning of Augmentation Policy Schedules

https://bair.berkeley.edu/blog/2019/06/07/data_aug/

Apache License 2.0

504 stars 86 forks source link

Add test PBA search script, remove dataset assumptions #2

Closed arcelien closed 5 years ago

arcelien commented 5 years ago

Add scripts/test_search.sh to test with random noise data and 224x224 image size. Removes many hardcoded assumptions about SVHN or CIFAR. Adding a new dataset should now only require loading the data arrays in the dataloader object of pba/data_utils.py

(needs lint)

arcelien commented 5 years ago

Tested on R-SVHN search and retraining with R-SVHN on WRN-28-10.