better EMA scope under existing name_scope.

examples/GAN/CycleGAN.py

@@ -86,15 +86,17 @@ class Model(GANModelDesc):
 
     def _build_graph(self, inputs):
         A, B = inputs
+        with tf.name_scope('preprocess'):
             A = tf.transpose(A / 128.0 - 1.0, [0, 3, 1, 2])
             B = tf.transpose(B / 128.0 - 1.0, [0, 3, 1, 2])
 
         def viz3(name, a, b, c):
+            with tf.name_scope(name):
                 im = tf.concat([a, b, c], axis=3)
                 im = tf.transpose(im, [0, 2, 3, 1])
                 im = (im + 1.0) * 128
                 im = tf.clip_by_value(im, 0, 255)
-            im = tf.cast(im, tf.uint8, name='viz_' + name)
+                im = tf.cast(im, tf.uint8, name='viz')
             tf.summary.image(name, im, max_outputs=50)
 
         # use the initializers from torch
@@ -124,7 +126,6 @@ class Model(GANModelDesc):
                     B_dis_fake = self.discriminator(AB)
 
         def LSGAN_losses(real, fake):
-            with tf.name_scope('LSGAN_losses'):
             d_real = tf.reduce_mean(tf.squared_difference(real, 0.9), name='d_real')
             d_fake = tf.reduce_mean(tf.square(fake), name='d_fake')
             d_loss = tf.multiply(d_real + d_fake, 0.5, name='d_loss')
@@ -133,6 +134,7 @@ class Model(GANModelDesc):
             add_moving_summary(g_loss, d_loss)
             return g_loss, d_loss
 
+        with tf.name_scope('losses'):
             with tf.name_scope('LossA'):
                 # reconstruction loss
                 recon_loss_A = tf.reduce_mean(tf.abs(A - ABA), name='recon_loss')
@@ -183,7 +185,7 @@ def get_data(datadir, isTrain=True):
 class VisualizeTestSet(Callback):
     def _setup_graph(self):
         self.pred = self.trainer.get_predictor(
-            ['inputA', 'inputB'], ['viz_A_recon', 'viz_B_recon'])
+            ['inputA', 'inputB'], ['A_recon/viz', 'B_recon/viz'])
 
     def _before_train(self):
         global args

tensorpack/tfutils/summary.py

@@ -179,10 +179,12 @@ def add_moving_summary(v, *args, **kwargs):
                 ema_var = tf.get_variable(name, shape=c.shape, dtype=c.dtype,
                                           initializer=tf.constant_initializer(), trainable=False)
                 ns = vs.original_name_scope
-            with tf.name_scope(ns):
+            # first clear NS to avoid duplicated name in variables
+            with tf.name_scope(None), tf.name_scope(ns):
                 ema_op = moving_averages.assign_moving_average(
                     ema_var, c, decay,
                     zero_debias=True, name=name + '_EMA_apply')
+            with tf.name_scope(None):
                 tf.summary.scalar(name + '-summary', ema_op)
             tf.add_to_collection(coll, ema_op)
             # TODO a new collection to summary every step?