Supporting Dict observation spaces for ComplexInputNetwork model?

Hi there. I’m attempting to use a modified version of the ComplexInputNetwork (ray/ at aecc4c8d28c6fa1ac73c1d142fe5b9ee355410a4 · ray-project/ray · GitHub), but I’m running into issues when trying to pass in a Dict observation space rather than a Tuple observation space. In particular, the issue is that when passing in a Dict observation space, the model errors out in the forward pass in the file’s get_action_space function, because the action space has somehow been set to be the same as the observation space. When I pass the flattened observation space in instead (as is usually done with Rllib models) and set a breakpoint in the above function, the action space is set correctly to be the action space the model’s initialized with. Any ideas what the issue might be here?

Hi @cwijesundara,

Can you post the modified version you are using?