facebookresearch / open_lth

A repository in preparation for open-sourcing lottery ticket hypothesis code.
MIT License
624 stars 113 forks source link

Fair comparison- train/val/test #4

Closed rahimentezari closed 3 years ago

rahimentezari commented 4 years ago

Hi As far as I see from the code, For each pruning ration, you report the test accuracy for the last epoch (#160). right? How can I change the code to divide the training into train and validation? The goal is to report the test accuracy for the best validation accuracy.

jfrankle commented 4 years ago

You will need to modify the code to generate train, test, and validation sets (modifying the datasets module) and modify the corresponding test callback to evaluate accordingly.

jfrankle commented 4 years ago

Also, it records test accuracy for every epoch by default, although it only prints the last epoch on the command line.