Update defaults for offpolicy algos with features extractor (#935)

Author: araffin

docs/misc/changelog.rst

@@ -4,14 +4,16 @@ Changelog
 ==========
 
 
-Release 1.5.1a8 (WIP)
+Release 1.5.1a9 (WIP)
 ---------------------------
 
 Breaking Changes:
 ^^^^^^^^^^^^^^^^^
 - Changed the way policy "aliases" are handled ("MlpPolicy", "CnnPolicy", ...), removing the former
   ``register_policy`` helper, ``policy_base`` parameter and using ``policy_aliases`` static attributes instead (@Gregwar)
 - SB3 now requires PyTorch >= 1.11
+- Changed the default network architecture when using ``CnnPolicy`` or ``MultiInputPolicy`` with SAC or DDPG/TD3,
+  ``share_features_extractor`` is now set to False by default and the ``net_arch=[256, 256]`` (instead of ``net_arch=[]`` that was before)
 
 New Features:
 ^^^^^^^^^^^^^

stable_baselines3/sac/policies.py

@@ -235,7 +235,7 @@ def __init__(
         optimizer_class: Type[th.optim.Optimizer] = th.optim.Adam,
         optimizer_kwargs: Optional[Dict[str, Any]] = None,
         n_critics: int = 2,
-        share_features_extractor: bool = True,
+        share_features_extractor: bool = False,
     ):
         super().__init__(
             observation_space,
@@ -248,10 +248,7 @@ def __init__(
         )
 
         if net_arch is None:
-            if features_extractor_class == NatureCNN:
-                net_arch = []
-            else:
-                net_arch = [256, 256]
+            net_arch = [256, 256]
 
         actor_arch, critic_arch = get_actor_critic_arch(net_arch)
 
@@ -422,7 +419,7 @@ def __init__(
         optimizer_class: Type[th.optim.Optimizer] = th.optim.Adam,
         optimizer_kwargs: Optional[Dict[str, Any]] = None,
         n_critics: int = 2,
-        share_features_extractor: bool = True,
+        share_features_extractor: bool = False,
     ):
         super().__init__(
             observation_space,
@@ -493,7 +490,7 @@ def __init__(
         optimizer_class: Type[th.optim.Optimizer] = th.optim.Adam,
         optimizer_kwargs: Optional[Dict[str, Any]] = None,
         n_critics: int = 2,
-        share_features_extractor: bool = True,
+        share_features_extractor: bool = False,
     ):
         super().__init__(
             observation_space,

stable_baselines3/td3/policies.py

@@ -119,7 +119,7 @@ def __init__(
         optimizer_class: Type[th.optim.Optimizer] = th.optim.Adam,
         optimizer_kwargs: Optional[Dict[str, Any]] = None,
         n_critics: int = 2,
-        share_features_extractor: bool = True,
+        share_features_extractor: bool = False,
     ):
         super().__init__(
             observation_space,
@@ -134,7 +134,7 @@ def __init__(
         # Default network architecture, from the original paper
         if net_arch is None:
             if features_extractor_class == NatureCNN:
-                net_arch = []
+                net_arch = [256, 256]
             else:
                 net_arch = [400, 300]
 
@@ -281,7 +281,7 @@ def __init__(
         optimizer_class: Type[th.optim.Optimizer] = th.optim.Adam,
         optimizer_kwargs: Optional[Dict[str, Any]] = None,
         n_critics: int = 2,
-        share_features_extractor: bool = True,
+        share_features_extractor: bool = False,
     ):
         super().__init__(
             observation_space,
@@ -335,7 +335,7 @@ def __init__(
         optimizer_class: Type[th.optim.Optimizer] = th.optim.Adam,
         optimizer_kwargs: Optional[Dict[str, Any]] = None,
         n_critics: int = 2,
-        share_features_extractor: bool = True,
+        share_features_extractor: bool = False,
     ):
         super().__init__(
             observation_space,

stable_baselines3/version.txt

@@ -1 +1 @@
-1.5.1a8
+1.5.1a9