clean-up '-summary' in stats name. clean-up some use of get_scalar_var

examples/A3C-Gym/train-atari.py

@@ -135,7 +135,7 @@ class Model(ModelDesc):
                                    self.cost, tf.reduce_mean(importance, name='importance'))
 
     def _get_optimizer(self):
-        lr = symbf.get_scalar_var('learning_rate', 0.001, summary=True)
+        lr = tf.get_variable('learning_rate', initializer=0.001, trainable=False)
         opt = tf.train.AdamOptimizer(lr, epsilon=1e-3)
 
         gradprocs = [MapGradient(lambda grad: tf.clip_by_average_norm(grad, 0.1)),

examples/FasterRCNN/train.py

@@ -126,7 +126,8 @@ class Model(ModelDesc):
             decoded_boxes = tf.identity(decoded_boxes, name='fastrcnn_fg_boxes')
 
     def _get_optimizer(self):
-        lr = symbf.get_scalar_var('learning_rate', 0.003, summary=True)
+        lr = tf.get_variable('learning_rate', initializer=0.003, trainable=False)
+        tf.summary.scalar('learning_rate', lr)
 
         factor = get_batch_factor()
         if factor != 1:

examples/GAN/BEGAN.py

@@ -126,7 +126,7 @@ class Model(GANModelDesc):
                 self.g_loss = L_neg
 
         add_moving_summary(L_pos, L_neg, eq, measure, self.d_loss)
-        tf.summary.scalar('kt-summary', kt)
+        tf.summary.scalar('kt', kt)
 
         self.collect_variables()
 

examples/GAN/CycleGAN.py

@@ -157,7 +157,7 @@ class Model(GANModelDesc):
         add_moving_summary(recon_loss_A, recon_loss_B, self.g_loss, self.d_loss)
 
     def _get_optimizer(self):
-        lr = symbolic_functions.get_scalar_var('learning_rate', 2e-4, summary=True)
+        lr = tf.get_variable('learning_rate', initializer=2e-4, trainable=False)
         return tf.train.AdamOptimizer(lr, beta1=0.5, epsilon=1e-3)
 
 

examples/GAN/DCGAN.py

@@ -95,7 +95,7 @@ class Model(GANModelDesc):
         self.collect_variables()
 
     def _get_optimizer(self):
-        lr = symbolic_functions.get_scalar_var('learning_rate', 2e-4, summary=True)
+        lr = tf.get_variable('learning_rate', initializer=2e-4, trainable=False)
         return tf.train.AdamOptimizer(lr, beta1=0.5, epsilon=1e-3)
 
 

examples/HED/hed.py

@@ -96,7 +96,7 @@ class Model(ModelDesc):
             add_moving_summary(costs + [wrong, self.cost])
 
     def _get_optimizer(self):
-        lr = symbf.get_scalar_var('learning_rate', 3e-5, summary=True)
+        lr = tf.get_variable('learning_rate', initializer=3e-5, trainable=False)
         opt = tf.train.AdamOptimizer(lr, epsilon=1e-3)
         return optimizer.apply_grad_processors(
             opt, [gradproc.ScaleGradient(

examples/ResNet/imagenet_utils.py

@@ -207,5 +207,5 @@ class ImageNetModel(ModelDesc):
 
     def _get_optimizer(self):
         lr = tf.get_variable('learning_rate', initializer=0.1, trainable=False)
-        tf.summary.scalar('learning_rate-summary', lr)
+        tf.summary.scalar('learning_rate', lr)
         return tf.train.MomentumOptimizer(lr, 0.9, use_nesterov=True)

examples/boilerplate.py

@@ -30,7 +30,7 @@ class Model(ModelDesc):
         summary.add_moving_summary(self.cost)
 
     def _get_optimizer(self):
-        lr = symbolic_functions.get_scalar_var('learning_rate', 5e-3, summary=True)
+        lr = tf.get_variable('learning_rate', initializer=5e-3, trainable=False)
         return tf.train.AdamOptimizer(lr)
 
 

examples/cifar-convnet.py

@@ -78,7 +78,8 @@ class Model(ModelDesc):
         self.cost = tf.add_n([cost, wd_cost], name='cost')
 
     def _get_optimizer(self):
-        lr = symbf.get_scalar_var('learning_rate', 1e-2, summary=True)
+        lr = tf.get_variable('learning_rate', initializer=1e-2, trainable=False)
+        tf.summary.scalar('lr', lr)
         return tf.train.AdamOptimizer(lr, epsilon=1e-3)
 
 

tensorpack/callbacks/monitor.py

@@ -129,9 +129,12 @@ class Monitors(Callback):
         for val in summary.value:
             if val.WhichOneof('value') == 'simple_value':
                 val.tag = re.sub('tower[0-9]+/', '', val.tag)   # TODO move to subclasses
+
+                # TODO This hack not needed any more, can remove this in the future
                 suffix = '-summary'  # tensorflow#6150, tensorboard#59
                 if val.tag.endswith(suffix):
                     val.tag = val.tag[:-len(suffix)]
+
                 self._dispatch(lambda m: m.process_scalar(val.tag, val.simple_value))
 
         self._dispatch(lambda m: m.process_summary(summary))

tensorpack/tfutils/summary.py

@@ -246,7 +246,7 @@ def add_moving_summary(*args, **kwargs):
             ema_ops.append(ema_op)
         with tf.name_scope(None):
             # cannot add it into colocate group -- will force everything to cpus
-            tf.summary.scalar(name + '-summary', ema_op)    # write the EMA value as a summary
+            tf.summary.scalar(name, ema_op)    # write the EMA value as a summary
     if coll is not None:
         for op in ema_ops:
             # TODO a new collection to summary every step?

tensorpack/tfutils/symbolic_functions.py

@@ -154,9 +154,11 @@ def huber_loss(x, delta=1, name='huber_loss'):
     return tf.where(cond, l2, l1, name=name)
 
 
+# TODO remove this in the future
 def get_scalar_var(name, init_value, summary=False, trainable=False):
     """
-    Get a scalar float variable with certain initial value
+    Get a scalar float variable with certain initial value.
+    You can just call `tf.get_variable(name, initializer=init_value, trainable=False)` instead.
 
     Args:
         name (str): name of the variable.
@@ -170,7 +172,7 @@ def get_scalar_var(name, init_value, summary=False, trainable=False):
                           trainable=trainable)
     if summary:
         # this is recognized in callbacks.StatHolder
-        tf.summary.scalar(name + '-summary', ret)
+        tf.summary.scalar(name, ret)
     return ret