[rllib] Improve model config documentation

[pschafhalter]

I've been trying to use the model config as much as possible to avoid writing custom models, but often the documentation is difficult to understand.

For example, to understand the conv_filters option, the documentation says that the model config is documented in the model catalog. In the source code, the model catalog documentation for conv_filters says "conv_filters", # Filter configuration. I had to grep the source code before finding out that conv_filters is a list of [out_channels, kernel, stride] from visionnet.py.

[ericl]

Would you mind submitting any quick fixes as a PR?

Note that in general for customizations it's better to just write a custom model instead, it amounts to copying the model source code and registering it. The amount of stuff we can support in a config is limited.

[drozzy]

Related question on SO:

ValueError: No default configuration for obs shape [6, 94], you must specify `conv_filters` manually as a model option. Default configurations are only available for inputs of shape [42, 42, K] and [84, 84, K]. You may alternatively want to use a custom model or preprocessor.

"For the (6,94) case, is the "dim" supposed to be 6 or 94? Is the conv_filters fine leaving as is, or is a certain network architecture necessary for each shape? If so, where is the link?"

https://stackoverflow.com/questions/57151612/what-is-the-appropriate-conv-filter-specifications-for-custom-observation-space

[dano000]

I'm having the same issue here:

ValueError: No default configuration for obs shape [1, 5], you must specify `conv_filters` manually as a model option. Default configurations are only available for inputs of shape [42, 42, K] and [84, 84, K]. You may alternatively want to use a custom model or preprocessor.

[ericl]

You can force using a fcnet as the custom model, or flatten your observation. It's trying to create a convnet over the inputs and getting confused about the shape.

…

On Sun, Aug 18, 2019, 6:44 PM D Schulz ***@***.***> wrote: I'm having the same issue here: ValueError: No default configuration for obs shape [1, 5], you must specify `conv_filters` manually as a model option. Default configurations are only available for inputs of shape [42, 42, K] and [84, 84, K]. You may alternatively want to use a custom model or preprocessor. — You are receiving this because you modified the open/close state. Reply to this email directly, view it on GitHub <#2538?email_source=notifications&email_token=AAADUSWJKZD7H6HXSCGEVR3QFER2JA5CNFSM4FNNSO7KYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOD4Q5HCI#issuecomment-522310537>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AAADUSQ42YIJ7TFNX5SUHH3QFER2JANCNFSM4FNNSO7A> .

[FruechteBini]

Hey guys, just found this thread. I am having a hard time implementing this aswell. I am a RL beginner.
My observation space is of shape = (4,5,5)
I found this regarding the observation shape:

I guess the "dim" would be 5 in my case but how do I calculate the "conv_filters"?

Thanks in advance, cheers!