aclegg3
commented
1 year ago Hey @man469,
The NN and training code for Habitat resides in habitat-baselines package/directory. Outside of the baselines module, there is no implicit model structure requirement (e.g. the tasks don't dictate your model). If you intend to use an external training library you can do whatever you want, if you prefer to use habitat-baselines DDPPO trainer, there are some parameters built in (See here.). We typically load an existing pre-trainined network architecture to avoid cold-starting the CNN layers.
❓ Questions and Help
I know that habitat-lab can implement reinforcement learning using PPO algorithm. My question is can I configure the neural network structure in habitat-lab agent? I cannot find reference that explains what network structure is used in habitat-lab. Could anyone help? Thanks a lot !!