Is it possible to implement custom rnn without using rllib provided api?

RLlib
1
  • High: It blocks me to complete my task.

Hello, I’m trying to implement a custom RNN with TF. I’m curious if I need to follow the example provided.

In my case, the env provides the observation of shape [timesteps, features]. Thus, the historical data is already provided for the rnn network. Can I just implement forward() without forward_rnn().

2

Definitely.
you can implement your own RNN model by simply extending TFModelV2 instead of RecurrentNetwork (which is just a util class to make normal RNN modeling easier).
then you only have to implement forward().

1 Like