search

cesare-spinoso / HopperMujoco

1 stars 0 forks source link

Adding an evaluation script #17

Closed cesare-spinoso closed 2 years ago

cesare-spinoso commented 2 years ago

Because when they run our model on the server they will use 50 runs for 5 different seeds, I suggest that we re-evaluate the best trained models for each hyperparameters in this way. I also suggest that for each trained model we re-train only for 100_000 time steps for 5 different seeds and calculate the AUC so that we have a better idea of which model is the most sample efficient. I've already added some code but it needs more work here: https://github.com/cesare-spinoso/GROUP_013/blob/vpg_train_cesare/evaluate_agent.py

Copy over https://github.com/cesare-spinoso/GROUP_013/projects/1#card-80068217 and make sure randomness is ensured

sabina-elkins commented 2 years ago

@cesare-spinoso checkout my PR, I already have a method to calculate sample efficiency (i.e. "100_000 time steps for 5 different seeds and calculate the AUC ")

though mine is for training sample efficiency. I assume we want both?