[Fix] Use single scope declaration for models

ml-agents/mlagents/trainers/bc/models.py

@@ -6,10 +6,8 @@
 class BehavioralCloningModel(LearningModel):
     def __init__(self, brain, h_size=128, lr=1e-4, n_layers=2, m_size=128,
                  normalize=False, use_recurrent=False, scope='PPO', seed=0):
-
-        LearningModel.__init__(self, m_size, normalize, use_recurrent, brain, scope, seed)
-
         with tf.variable_scope(scope):
+            LearningModel.__init__(self, m_size, normalize, use_recurrent, brain, seed)
             num_streams = 1
             hidden_streams = self.create_observation_streams(num_streams, h_size, n_layers)
             hidden = hidden_streams[0]

ml-agents/mlagents/trainers/models.py

@@ -8,24 +8,23 @@
 
 
 class LearningModel(object):
-    def __init__(self, m_size, normalize, use_recurrent, brain, scope, seed):
+    def __init__(self, m_size, normalize, use_recurrent, brain, seed):
         tf.set_random_seed(seed)
-        with tf.variable_scope(scope):
-            self.brain = brain
-            self.vector_in = None
-            self.global_step, self.increment_step = self.create_global_steps()
-            self.visual_in = []
-            self.batch_size = tf.placeholder(shape=None, dtype=tf.int32, name='batch_size')
-            self.sequence_length = tf.placeholder(shape=None, dtype=tf.int32, name='sequence_length')
-            self.mask_input = tf.placeholder(shape=[None], dtype=tf.float32, name='masks')
-            self.mask = tf.cast(self.mask_input, tf.int32)
-            self.m_size = m_size
-            self.normalize = normalize
-            self.use_recurrent = use_recurrent
-            self.act_size = brain.vector_action_space_size
-            self.vec_obs_size = brain.vector_observation_space_size * \
-                                brain.num_stacked_vector_observations
-            self.vis_obs_size = brain.number_visual_observations
+        self.brain = brain
+        self.vector_in = None
+        self.global_step, self.increment_step = self.create_global_steps()
+        self.visual_in = []
+        self.batch_size = tf.placeholder(shape=None, dtype=tf.int32, name='batch_size')
+        self.sequence_length = tf.placeholder(shape=None, dtype=tf.int32, name='sequence_length')
+        self.mask_input = tf.placeholder(shape=[None], dtype=tf.float32, name='masks')
+        self.mask = tf.cast(self.mask_input, tf.int32)
+        self.m_size = m_size
+        self.normalize = normalize
+        self.use_recurrent = use_recurrent
+        self.act_size = brain.vector_action_space_size
+        self.vec_obs_size = brain.vector_observation_space_size * \
+                            brain.num_stacked_vector_observations
+        self.vis_obs_size = brain.number_visual_observations
 
     @staticmethod
     def create_global_steps():

ml-agents/mlagents/trainers/policy.py

@@ -29,7 +29,7 @@ def __init__(self, seed, brain, trainer_parameters, sess):
         :param sess: The current TensorFlow session.
         """
         self.m_size = None
-        self.model = LearningModel(0, False, False, brain, scope='Model', seed=0)
+        self.model = None
         self.inference_dict = {}
         self.update_dict = {}
         self.sequence_length = 1

ml-agents/mlagents/trainers/ppo/models.py

@@ -26,8 +26,8 @@ def __init__(self, brain, lr=1e-4, h_size=128, epsilon=0.2, beta=1e-3, max_step=
         :param num_layers Number of hidden layers between encoded input and policy & value layers
         :param m_size: Size of brain memory.
         """
-        LearningModel.__init__(self, m_size, normalize, use_recurrent, brain, scope, seed)
         with tf.variable_scope(scope):
+            LearningModel.__init__(self, m_size, normalize, use_recurrent, brain, seed)
             self.use_curiosity = use_curiosity
             if num_layers < 1:
                 num_layers = 1