model weights and large transfer learning weight

Project/Assets/ML-Agents/Examples/Reacher/Prefabs/NewAgent.prefab

     VectorActionSize: 04000000
     VectorActionDescriptions: []
     VectorActionSpaceType: 1
-  m_Model: {fileID: 11400000, guid: 79166d8db14004ba0aa1ffdce3f62667, type: 3}
+  m_Model: {fileID: 11400000, guid: fccf649e8d1dc4ad1b7d28f21b572cbe, type: 3}
   m_InferenceDevice: 0
   m_BehaviorType: 0
   m_BehaviorName: Reacher

config/sac_transfer/Reacher.yaml

       learning_rate_schedule: linear
       model_schedule: constant
       batch_size: 256
-      buffer_size: 10000000
+      buffer_size: 6000000
       buffer_init_steps: 0
       tau: 0.005
       steps_per_update: 20.0
       encoder_layers: 1
       policy_layers: 2
-      forward_layers: 1
+      forward_layers: 2
-      feature_size: 32
+      feature_size: 64
       action_feature_size: 16
       separate_policy_train: true
       separate_policy_net: true
         gamma: 0.99
         strength: 1.0
     keep_checkpoints: 5
-    max_steps: 10000000
+    max_steps: 6000000
     time_horizon: 1000
     summary_freq: 60000
     threaded: true

config/sac_transfer/ReacherTransfer.yaml

   Reacher:
     trainer_type: sac_transfer
     hyperparameters:
-      learning_rate: 0.0003
+      learning_rate: 0.003
+      model_weight: 0.05
       batch_size: 256
       buffer_size: 6000000
       buffer_init_steps: 0
       reward_signal_steps_per_update: 20.0
       encoder_layers: 1
       policy_layers: 2
-      forward_layers: 1
+      forward_layers: 2
-      feature_size: 32
+      feature_size: 64
       action_feature_size: 16
       separate_policy_train: true
       separate_policy_net: true
       train_model: false
       load_action: true
       train_action: false
-      transfer_path: "results/reacher-f32/Reacher"
+      transfer_path: "results/reacher-lr/Reacher"
     network_settings:
       normalize: true
       hidden_units: 128

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

         self.train_model = hyperparameters.train_model
         self.train_policy = hyperparameters.train_policy
         self.train_value = hyperparameters.train_value
+        self.model_weight = hyperparameters.model_weight
 
         # Transfer
         self.use_transfer = hyperparameters.use_transfer
                 )
                 self.model_learning_rate = ModelUtils.create_schedule(
                     hyperparameters.model_schedule,
-                    lr,
+                    hyperparameters.model_learning_rate,
                     self.policy.global_step,
                     int(max_step),
                     min_value=1e-10,
 
         # Make sure policy is updated first, then value, then entropy.
         if self.use_transfer:
-            value_loss = self.total_value_loss + self.model_loss
+            value_loss = self.total_value_loss + self.model_weight * self.model_loss
         else:
             value_loss = self.total_value_loss
         with tf.control_dependencies([self.update_batch_policy]):

ml-agents/mlagents/trainers/settings.py

 @attr.s(auto_attribs=True)
 class SACTransferSettings(SACSettings):
     model_schedule: ScheduleType = ScheduleType.LINEAR
+    model_learning_rate: float = 3.0e-4
+    model_weight: float = 0.5
 
     separate_value_train: bool = False
     separate_policy_train: bool = False

run.sh

+#!/bin/bash
+
+# mlagents-learn config/sac/3DBallHard.yaml --run-id=hardball_sac --env=envs/3dballhard_qudrew --num-envs=4 --no-graphics
+# mlagents-learn config/sac_transfer/3DBall.yaml --run-id=ball_2f --env=envs/3dball_qudrew --num-envs=4 --no-graphics
+# mlagents-learn config/sac_transfer/3DBall.yaml --run-id=ball_1f --env=envs/3dball1f_qudrew --num-envs=4 --no-graphics
+mlagents-learn config/sac_transfer/3DBall1fTransfer.yaml --run-id=ball_transfer_2f_noload-model --env=envs/3dball_qudrew --num-envs=4 --no-graphics
+# mlagents-learn config/sac_transfer/3DBallHardTransfer.yaml --run-id=transfer_action-enc_linear --env=envs/3dballhard_qudrew --num-envs=4 --no-graphics
+# mlagents-learn config/sac_transfer/3DBallHard.yaml --run-id=hardball_action-enc_linear --env=envs/3dballhard_qudrew --num-envs=4 --no-graphics
+# mlagents-learn config/sac_transfer/3DBallHardTransfer1.yaml --run-id=sac_transfer_hardball_fixpol_ov --env=envs/3dballhard --num-envs=4 --no-graphics
+
+# mlagents-learn config/sac_transfer/CrawlerStatic.yaml --run-id=oldcs --env=envs/old_crawler_static --num-envs=4 --no-graphics
+# mlagents-learn config/sac_transfer/CrawlerStaticTransfer.yaml --run-id=transfer_newcs --env=envs/new_crawler_static --num-envs=4 --no-graphics
+# mlagents-learn config/sac_transfer/CrawlerStatic.yaml --run-id=newcs--env=envs/new_crawler_static --num-envs=4 --no-graphics