Features to save and reuse the trained models are now integrated

sanjaythakur commented 6 years ago

The 'train.py' script now saves the model once the training has been completed which can be reused again at a later point in time.

pat-coady commented 6 years ago

Sanjay - very sorry for the delay. Can you describe any testing you did to check your improvements didn't break anything?

Thanks a lot, Pat

sanjaythakur commented 6 years ago

Hi Pat, the only changes I made in the code are

saving the policy network after the prescribed training has completed,
an additional script by the name of reload.py to show how to use those saved policy networks later.

Other than these two changes, I didn't change/modify any part of the existing code. So I don't think it should break anything. For validating the robustness, I have run python train.py from src/ on all the MUJOCO environments except Ant, Humanoid, and HumanoidStandup (because they take a lot of time) over the last three days and then verified the saved models by running reload.py. Everything worked out well.

I hope this answers your question. Let me know if you think it needs more testing.

pat-coady commented 6 years ago

Looks good, merged. Nice addition, thanks a lot.

pat-coady commented 6 years ago

Sanjay -

Sorry, I had to revert this. I got errors when running locally. Also, new code doesn't have docstrings, seems to be indented with tabs and other PEP8 violations.

Some errors I saw on reload. Also, code created some unexpected files src directory.

from: can't read /var/mail/gym ./reload.py: line 8: syntax error near unexpected token (' ./reload.py: line 8:def getScaleAndOffset(env_name):'

pat-coady / trpo

Features to save and reuse the trained models are now integrated #17