search

chainer / chainerrl

ChainerRL is a deep reinforcement learning library built on top of Chainer.
MIT License
1.16k stars 226 forks source link

Saving models at timestamps #586

Closed Ujwal2910 closed 4 years ago

Ujwal2910 commented 4 years ago

I am using the train_ppo_gym script and the problem with this script is that it does not save model when the training crashes or is closed for some reason. Earlier with a3c i used to get 10_except as saved model in case i wanted to stop the training. I want to save mdoels for evey 100th step or episode. Could you please provide an update on this behavior?

prabhatnagarajan commented 4 years ago

I am using the train_ppo_gym script and the problem with this script is that it does not save model when the training crashes or is closed for some reason. Could you clarify this?

We have functionality that can checkpoint models periodically, though it has not been implemented for train_ppo_gym. Checkout the train_ppo_ale.py example: https://github.com/chainer/chainerrl/blob/master/examples/atari/train_ppo_ale.py#L87

prabhatnagarajan commented 4 years ago

Does this answer your question?

Ujwal2910 commented 4 years ago

Yeah this works. Just one more question is that whether any of the train_ppo files have used LSTM networks?

prabhatnagarajan commented 4 years ago

Checkout https://github.com/chainer/chainerrl/blob/master/examples/atari/train_ppo_ale.py#L131. Not LSTMS, but GRUs.