now need to add debug for discriminator and add history

Author: carpedm20

config.py

@@ -36,7 +36,7 @@ def add_argument_group(name):
 train_arg.add_argument('--is_train', type=str2bool, default=True, help='')
 train_arg.add_argument('--optimizer', type=str, default='rmsprop', help='')
 train_arg.add_argument('--max_step', type=int, default=200, help='')
-train_arg.add_argument('--reg_scale', type=float, default=0.01, help='')
+train_arg.add_argument('--reg_scale', type=float, default=1, help='')
 train_arg.add_argument('--initial_K_d', type=int, default=200, help='')
 train_arg.add_argument('--initial_K_g', type=int, default=1000, help='')
 train_arg.add_argument('--K_d', type=int, default=1, help='')

layers.py

@@ -43,7 +43,7 @@ def conv2d(inputs, num_outputs, kernel_size, stride,
            layer_dict={}, activation_fn=None,
            weights_initializer=tf.random_normal_initializer(0, 0.001),
            scope=None, name="", reuse=False, **kargv):
-  print tf.random_normal_initializer(0, 0.001)
+  print weights_initializer
   if True:
     outputs = slim.conv2d(
         inputs, num_outputs, kernel_size,

model.py

@@ -78,13 +78,14 @@ def _build_loss(self):
     with tf.variable_scope("refiner"):
       self.realism_loss = tf.reduce_sum(
           SE_loss(self.D_R_x_logits, real_label), [1, 2], name="realism_loss")
-      self.regularization_loss = 0
-      #    self.reg_scale * tf.reduce_sum(
-      #        self.R_x - self.normalized_x, [1, 2, 3],
-      #        name="regularization_loss")
+      #self.regularization_loss = 0
+      self.regularization_loss = \
+          self.reg_scale * tf.reduce_sum(
+              tf.abs(self.R_x - self.normalized_x), [1, 2, 3],
+              name="regularization_loss")
 
       self.refiner_loss = tf.reduce_mean(
-          self.realism_loss, #+ self.regularization_loss,
+          self.realism_loss + self.regularization_loss,
           name="refiner_loss")
 
       if self.debug:
@@ -162,22 +163,24 @@ def test_refiner(sess, inputs, summary_writer=None, with_output=False):
                  self.refiner_summary, summary_writer,
                  output_op=self.R_x if with_output else None)
 
-    def train_discrim(sess, inputs, summary_writer=None):
+    def train_discrim(sess, inputs, summary_writer=None, with_output=False):
       fetch = {
           'loss': self.discrim_loss,
           'optim': self.discrim_optim,
           'step': self.discrim_step,
       }
       return run(sess, inputs, fetch, self.x,
-                 self.discrim_summary, summary_writer)
+                 self.discrim_summary, summary_writer,
+                 output_op=self.D_R_x if with_output else None)
 
-    def test_discrim(sess, inputs, summary_writer=None):
+    def test_discrim(sess, inputs, summary_writer=None, with_output=False):
       fetch = {
           'loss': self.discrim_loss,
           'step': self.discrim_step,
       }
       return run(sess, inputs, fetch, self.x,
-                 self.discrim_summary, summary_writer=summary_writer)
+                 self.discrim_summary, summary_writer,
+                 output_op=self.D_R_x if with_output else None)
 
     self.train_refiner = train_refiner
     self.test_refiner = test_refiner

trainer.py

@@ -57,20 +57,19 @@ def train(self):
     print("[*] Training starts...")
     summary_writer = None
 
+    for k in trange(self.initial_K_d, desc="Train discrim"):
+      res = self.model.train_discrim(sess, self.data_loader.next(),
+                                     summary_writer, with_output=False)
+      summary_writer = self._get_summary_writer(res)
+
     for k in trange(self.initial_K_g, desc="Train refiner"):
       data = self.data_loader.next()
       res = self.model.train_refiner(sess, data,
                                      summary_writer, with_output=True)
       summary_writer = self._get_summary_writer(res)
-      import ipdb; ipdb.set_trace() 
+      # import ipdb; ipdb.set_trace() 
       # self.model.R_x.eval({self.model.x: data)}, session=sess)
       # self.model.layer_dict['refiner/resnet/resnet_5'].eval({self.model.x: data},session=sess).max()
-      x = 123
-
-    for k in trange(self.initial_K_d, desc="Train discrim"):
-      res = self.model.train_discrim(sess, self.data_loader.next(),
-                                     summary_writer, with_output=True)
-      summary_writer = self._get_summary_writer(res)
 
     for step in trange(self.max_step, desc="Train both"):
       for k in xrange(self.K_g):