Question regarding custom network

[SourabhPrasad]

Hi,
I’m working on a custom network that includes an Actor MLP, a Critic MLP, and an extra MLP (let’s call it MLP-B). The setup involves feeding the output of MLP-B into the Actor MLP. On top of that, I want to train MLP-B using the PPO algorithm while also training it separately with regression. Is it possible to set up an architecture like this? If so, I’d really appreciate any advice on how to make it happen.
Thanks!

[Denys88]

Yes it is possible I think we have nearest example. Please take a look at the aux_loss support implementation here: 7f9cd1e

[SourabhPrasad]

Hey @Denys88.
Thanks for the reply and for pointing me towards aux_loss. I have taken a look at 7f9cd1e and TestNetWithAuxLoss class in test_network.py and I had a follow-up question.
In my implementation, the actor, critic and MLP-B use different architectures and are asymmetric. How can I implement such a model using the network builder, Would you suggest building such a network using NetworkBuilder.BaseNetwork or A2CBuilder

[Denys88]

@SourabhPrasad there are some very simple examples without a2c. a2c is a monster where I supported everything I want to test.