search

aqeelanwar / PEDRA

Programmable Engine for Drone Reinforcement Learning Applications
MIT License
266 stars 59 forks source link

Resuming training using learned weights #22

Closed SumukhaNadig closed 4 years ago

SumukhaNadig commented 4 years ago

Hi, while training my system crashes at times due to unknown reasons so I was wondering if there is any way to reload the weights from previous training sessions which were saved at some point instead of starting all over.

aqeelanwar commented 4 years ago

Hi, You can set custom_load parameter to True and provide the path of the saved weights to custom_load_path in the DeepQLearning.cfg file. This will initialize the NN weights with saved weights.

SumukhaNadig commented 4 years ago

Thank you, just wanted to know which file the weights are getting stored to? I couldn't see any .npy files and all of them seem to have a different encoding.

aqeelanwar commented 4 years ago

The weights are not stored as .npy file, rather they are stored as checkpoints.

This might help: https://stackoverflow.com/questions/41265035/tensorflow-why-there-are-3-files-after-saving-the-model

SumukhaNadig commented 4 years ago

Thank you. Just want to confirm, its enough to have ../../drone0/drone0_user in the load path right? In the given config file, it was given as ../../drone0/drone0. Also I don't need to explicitly load index and meta files, I assume and just drone0/drone0_user should be able to load all 3 required files.

Also just wanted to know what the DeepReinforce.cfg does, couldn't find any documentation on it.

aqeelanwar commented 4 years ago

Yes it is enough. If you are using saver.restore() then you don't need to explicitly load other files.

That is work in progress. Apparently I mistakenly pushed the file to git. It will be the Policy Gradient RL algorithm.

SumukhaNadig commented 4 years ago

Thank you so much! :)