search

GuHongyang / AM-LFS-PyTorch

The reproduce for "AM-LFS: AutoML for Loss Function Search"
13 stars 6 forks source link

Where is the reward calculation #1

Open RizhaoCai opened 3 years ago

RizhaoCai commented 3 years ago

Hello, thanks for your reproduction.

In the AM-LFS paper, the accuracy from the validation is used for loss calculation to update the parameter based on the REINFORCE rule. image

However, in https://github.com/GuHongyang/AM-LFS-PyTorch/blob/bb4e4f08c0c88abe7919d3ea2db46db7ef345cf9/main.py#L44, why accuracy is not returned from validation to calculate reward?

GuHongyang commented 3 years ago

Hello! In the trainer.py, I have calculated the reward from validation data!

This version of the code is more difficult to understand, and finally I re-written it based on the Ray framework. But the result is still not very good.