Unclear sentence about the net_arch argument for on policy algorithms

[EBoguslawski]

📚 Documentation

Hello,
I was reading the documentation about how to customize the networks' architecture here:
https://stable-baselines3.readthedocs.io/en/master/guide/custom_policy.html#on-policy-algorithms

1. And I found this sentence:
   "Otherwise, to have actor and critic that share the same network architecture, you only need to specify net_arch=[128, 128] (here, two hidden layers of 128 units each, this is equivalent to net_arch=dict(pi=[128, 128], vf=[128, 128]))."
   I think it might induce confusion because (unless I miss something) weights are shared by the actor and critic in the 1st case (with net_arch=[128, 128]) but not in the 2nd one (with net_arch=dict(pi=[128, 128], vf=[128, 128]))). Thus, these syntaxes are not really equivalent.

2. The example just below the sentence gives the impression that weights are not shared when using net_arch=[128, 128].
   "Same architecture for actor and critic with two layers of size 128: net_arch=[128, 128]"

        obs
   /            \
 <128>          <128>
  |              |
 <128>          <128>
  |              |
action         value

Maybe you could use:

        obs
         |
        <128>
         |
        <128>
   /            \
action         value

Hoping I did not misunderstand something.
Eva

Checklist

• I have checked that there is no similar issue in the repo
• I have read the documentation

[araffin]

The example just below the sentence gives the impression that weights are not shared when using

It's not an impression, it is the case. We changed the behavior in SB3 v1.8.0+ to match the offpolicy algorithms and simplify the code: #1292 and #1252

[EBoguslawski]

Indeed I use SB3 v1.7.0. Thank you for your answer !