linear v

config/sac_transfer/CrawlerStatic.yaml

       encoder_layers: 3
       policy_layers: 0
       forward_layers: 0
-      value_layers: 2
-      feature_size: 300
+      value_layers: 0
+      feature_size: 512
-      action_feature_size: 128
+      action_feature_size: 256
       separate_policy_train: true
       separate_model_train: true
       # separate_value_net: true

ml-agents/mlagents/trainers/sac_transfer/network.py

             self.value_heads[name] = value
         self.value = tf.reduce_mean(list(self.value_heads.values()), 0)
 
-    def _create_cc_critic(self, hidden_value, scope, create_qs=True):
+    def _create_cc_critic(self, hidden_value, scope, create_qs=True, a_layers=-1, a_features=16):
         """
         Creates just the critic network
         """
             name="external_action_in",
         )
         self.value_vars = self.get_vars(self.join_scopes(scope, "value"))
+
+            # external_action_encoder = self.policy._create_action_encoder(
+            #     self.external_action_in,
+            #     self.h_size,
+            #     a_features,
+            #     a_layers,
+            #     reuse=True
+            # )
+
+            # output_action_encoder = self.policy._create_action_encoder(
+            #     self.policy.output,
+            #     self.h_size,
+            #     a_features,
+            #     a_layers,
+            #     reuse=True
+            # )
+            # hidden_q = tf.concat([hidden_value, external_action_encoder], axis=-1)
+            # hidden_qp = tf.concat([hidden_value, output_action_encoder], axis=-1)
+            
-                self.num_layers,
+                self.num_layers+2,
                 self.h_size,
                 self.join_scopes(scope, "q"),
             )
-                self.num_layers,
+                self.num_layers+2,
                 self.h_size,
                 self.join_scopes(scope, "q"),
                 reuse=True,
         use_recurrent=False,
         encoder_layers=0,
         num_layers=2,
+        action_layers=-1,
+        action_features=16,
         stream_names=None,
         vis_encode_type=EncoderType.SIMPLE,
         separate_train=False,
         use_recurrent=False,
         encoder_layers=0,
         num_layers=2,
+        action_layers=-1,
+        action_features=16,
         stream_names=None,
         vis_encode_type=EncoderType.SIMPLE,
         separate_train=False,
             hidden_critic = tf.stop_gradient(hidden_critic)
 
         if self.policy.use_continuous_act:
-            self._create_cc_critic(hidden_critic, POLICY_SCOPE)
+            self._create_cc_critic(hidden_critic, POLICY_SCOPE, a_layers=action_layers, a_features=action_features)
         else:
             self._create_dc_critic(hidden_critic, POLICY_SCOPE)
             # self._create_dc_critic(hidden, POLICY_SCOPE)

ml-agents/mlagents/trainers/sac_transfer/optimizer.py

                         use_recurrent=self.policy.use_recurrent,
                         encoder_layers=hyperparameters.encoder_layers,
                         num_layers=hyperparameters.value_layers,
+                        action_layers=hyperparameters.action_layers,
+                        action_features=hyperparameters.action_feature_size,
                         stream_names=stream_names,
                         vis_encode_type=vis_encode_type,
                         separate_train=hyperparameters.separate_value_train,
                         use_recurrent=self.policy.use_recurrent,
                         encoder_layers=hyperparameters.encoder_layers,
                         num_layers=hyperparameters.value_layers,
+                        action_layers=hyperparameters.action_layers,
+                        action_features=hyperparameters.action_feature_size,
                         stream_names=stream_names,
                         vis_encode_type=vis_encode_type,
                         separate_train=hyperparameters.separate_value_train,
         
         policy_vars = self.policy.get_trainable_variables(
             train_encoder=not self.separate_policy_train,
-            train_action=self.train_action,
+            train_action=not self.separate_policy_train,
             train_model=False,
             train_policy=self.train_policy
         )