migrate to new model interface without is_training

examples/Atari2600/DQN.py

@@ -70,7 +70,7 @@ class Model(ModelDesc):
                 InputVar(tf.float32, (None,) + IMAGE_SHAPE3, 'next_state'),
                 InputVar(tf.bool, (None,), 'isOver') ]
 
-    def _get_DQN_prediction(self, image, is_training):
+    def _get_DQN_prediction(self, image):
         """ image: [0,255]"""
         image = image / 255.0
         with argscope(Conv2D, nl=PReLU.f, use_bias=True):
@@ -91,9 +91,9 @@ class Model(ModelDesc):
                 .FullyConnected('fc0', 512, nl=lambda x, name: LeakyReLU.f(x, 0.01, name))
                 .FullyConnected('fct', NUM_ACTIONS, nl=tf.identity)())
 
-    def _build_graph(self, inputs, is_training):
+    def _build_graph(self, inputs):
         state, action, reward, next_state, isOver = inputs
-        self.predict_value = self._get_DQN_prediction(state, is_training)
+        self.predict_value = self._get_DQN_prediction(state)
         action_onehot = tf.one_hot(action, NUM_ACTIONS, 1.0, 0.0)
         pred_action_value = tf.reduce_sum(self.predict_value * action_onehot, 1)    #N,
         max_pred_reward = tf.reduce_mean(tf.reduce_max(
@@ -101,14 +101,14 @@ class Model(ModelDesc):
         add_moving_summary(max_pred_reward)
 
         with tf.variable_scope('target'):
-            targetQ_predict_value = self._get_DQN_prediction(next_state, False)    # NxA
+            targetQ_predict_value = self._get_DQN_prediction(next_state)    # NxA
 
         # DQN
         #best_v = tf.reduce_max(targetQ_predict_value, 1)    # N,
 
         # Double-DQN
         tf.get_variable_scope().reuse_variables()
-        next_predict_value = self._get_DQN_prediction(next_state, is_training)
+        next_predict_value = self._get_DQN_prediction(next_state)
         self.greedy_choice = tf.argmax(next_predict_value, 1)   # N,
         predict_onehot = tf.one_hot(self.greedy_choice, NUM_ACTIONS, 1.0, 0.0)
         best_v = tf.reduce_sum(targetQ_predict_value * predict_onehot, 1)

examples/Inception/inception-bn.py

@@ -30,7 +30,7 @@ class Model(ModelDesc):
         return [InputVar(tf.float32, [None, INPUT_SHAPE, INPUT_SHAPE, 3], 'input'),
                 InputVar(tf.int32, [None], 'label') ]
 
-    def _build_graph(self, input_vars, is_training):
+    def _build_graph(self, input_vars):
         image, label = input_vars
         image = image / 128.0
 
@@ -57,7 +57,7 @@ class Model(ModelDesc):
                 outs.append(x4)
                 return tf.concat(3, outs, name='concat')
 
-        with argscope(Conv2D, nl=BNReLU(is_training), use_bias=False):
+        with argscope(Conv2D, nl=BNReLU(), use_bias=False):
             l = Conv2D('conv0', image, 64, 7, stride=2)
             l = MaxPooling('pool0', l, 3, 2, padding='SAME')
             l = Conv2D('conv1', l, 64, 1)

examples/Inception/inceptionv3.py

@@ -33,7 +33,7 @@ class Model(ModelDesc):
         return [InputVar(tf.float32, [None, INPUT_SHAPE, INPUT_SHAPE, 3], 'input'),
                 InputVar(tf.int32, [None], 'label') ]
 
-    def _build_graph(self, input_vars, is_training):
+    def _build_graph(self, input_vars):
         image, label = input_vars
         image = image / 128.0 - 1   # ?
 
@@ -69,7 +69,7 @@ class Model(ModelDesc):
                 .Conv2D('conv277ba', ch_r, [7,1])
                 .Conv2D('conv277bb', ch, [1,7])())
 
-        nl = BNReLU(is_training, decay=0.9997, epsilon=1e-3)
+        nl = BNReLU(decay=0.9997, epsilon=1e-3)
         with argscope(Conv2D, nl=nl, use_bias=False):
             l = (LinearWrap(image)
                 .Conv2D('conv0', 32, 3, stride=2, padding='VALID') #299
@@ -168,7 +168,7 @@ class Model(ModelDesc):
 
             l = GlobalAvgPooling('gap', l)
             # 1x1x2048
-            l = tf.nn.dropout(l, keep_prob=0.8 if is_training else 1)
+            l = Dropout('drop', l, 0.8)
             logits = FullyConnected('linear', l, out_dim=1000, nl=tf.identity)
 
         loss1 = tf.nn.sparse_softmax_cross_entropy_with_logits(br1, label)

examples/OpenAIGym/run-atari.py

@@ -39,7 +39,7 @@ class Model(ModelDesc):
                 InputVar(tf.int32, (None,), 'action'),
                 InputVar(tf.float32, (None,), 'futurereward') ]
 
-    def _get_NN_prediction(self, image, is_training):
+    def _get_NN_prediction(self, image):
         """ image: [0,255]"""
         image = image / 255.0
         with argscope(Conv2D, nl=tf.nn.relu):
@@ -56,9 +56,9 @@ class Model(ModelDesc):
         policy = FullyConnected('fc-pi', l, out_dim=NUM_ACTIONS, nl=tf.identity)
         return policy
 
-    def _build_graph(self, inputs, is_training):
+    def _build_graph(self, inputs):
         state, action, futurereward = inputs
-        policy = self._get_NN_prediction(state, is_training)
+        policy = self._get_NN_prediction(state)
         self.logits = tf.nn.softmax(policy, name='logits')
 
 def play_one_episode(player, func, verbose=False):

examples/ResNet/cifar10-resnet.py

@@ -37,7 +37,7 @@ class Model(ModelDesc):
                 InputVar(tf.int32, [None], 'label')
                ]
 
-    def _build_graph(self, input_vars, is_training):
+    def _build_graph(self, input_vars):
         image, label = input_vars
         image = image / 128.0 - 1
 
@@ -59,12 +59,12 @@ class Model(ModelDesc):
 
             with tf.variable_scope(name) as scope:
                 if not first:
-                    b1 = BatchNorm('bn1', l, is_training)
+                    b1 = BatchNorm('bn1', l)
                     b1 = tf.nn.relu(b1)
                 else:
                     b1 = l
                 c1 = conv('conv1', b1, out_channel, stride1)
-                b2 = BatchNorm('bn2', c1, is_training)
+                b2 = BatchNorm('bn2', c1)
                 b2 = tf.nn.relu(b2)
                 c2 = conv('conv2', b2, out_channel, 1)
 
@@ -76,7 +76,7 @@ class Model(ModelDesc):
                 return l
 
         l = conv('conv0', image, 16, 1)
-        l = BatchNorm('bn0', l, is_training)
+        l = BatchNorm('bn0', l)
         l = tf.nn.relu(l)
         l = residual('res1.0', l, first=True)
         for k in range(1, self.n):
@@ -91,7 +91,7 @@ class Model(ModelDesc):
         l = residual('res3.0', l, increase_dim=True)
         for k in range(1, self.n):
             l = residual('res3.' + str(k), l)
-        l = BatchNorm('bnlast', l, is_training)
+        l = BatchNorm('bnlast', l)
         l = tf.nn.relu(l)
         # 8,c=64
         l = GlobalAvgPooling('gap', l)

examples/svhn-digit-convnet.py

@@ -24,9 +24,8 @@ class Model(ModelDesc):
         return [InputVar(tf.float32, [None, 40, 40, 3], 'input'),
                 InputVar(tf.int32, [None], 'label') ]
 
-    def _build_graph(self, input_vars, is_training):
+    def _build_graph(self, input_vars):
         image, label = input_vars
-        keep_prob = tf.constant(0.5 if is_training else 1.0)
 
         image = image / 128.0 - 1
 
@@ -37,7 +36,7 @@ class Model(ModelDesc):
                 .Conv2D('conv3', 32, 3, padding='VALID')
                 .MaxPooling('pool2', 2, padding='SAME')
                 .Conv2D('conv4', 64, 3, padding='VALID')
-                .tf.nn.dropout(keep_prob)
+                .Dropout('drop', 0.5)
                 .FullyConnected('fc0', 512,
                         b_init=tf.constant_initializer(0.1))
                 .FullyConnected('linear', out_dim=10, nl=tf.identity)())