Using instance norm and addded pad support

Author: hardikbansal

layers.py

@@ -27,10 +27,10 @@ def instance_norm(x):
 def general_conv2d(inputconv, o_d=64, f_h=7, f_w=7, s_h=1, s_w=1, stddev=0.02, padding="VALID", name="conv2d", do_norm=True, do_relu=True, relufactor=0):
     with tf.variable_scope(name):
 
-        conv = tf.contrib.layers.conv2d(inputconv, o_d, f_w, s_w, padding, activation_fn=None, weights_initializer=tf.truncated_normal_initializer(stddev=stddev),biases_initializer=None)
+        conv = tf.contrib.layers.conv2d(inputconv, o_d, f_w, s_w, padding, activation_fn=None, weights_initializer=tf.truncated_normal_initializer(stddev=stddev),biases_initializer=tf.constant_initializer(0.0))
         if do_norm:
-            # conv = instance_norm(conv)
-            conv = tf.contrib.layers.batch_norm(conv, decay=0.9, updates_collections=None, epsilon=1e-5, scale=True, scope="batch_norm")
+            conv = instance_norm(conv)
+            # conv = tf.contrib.layers.batch_norm(conv, decay=0.9, updates_collections=None, epsilon=1e-5, scale=True, scope="batch_norm")
 
         if do_relu:
             if(relufactor == 0):
@@ -45,11 +45,11 @@ def general_conv2d(inputconv, o_d=64, f_h=7, f_w=7, s_h=1, s_w=1, stddev=0.02, p
 def general_deconv2d(inputconv, outshape, o_d=64, f_h=7, f_w=7, s_h=1, s_w=1, stddev=0.02, padding="VALID", name="deconv2d", do_norm=True, do_relu=True, relufactor=0):
     with tf.variable_scope(name):
 
-        conv = tf.contrib.layers.conv2d_transpose(inputconv, o_d, [f_h, f_w], [s_h, s_w], padding, activation_fn=None, weights_initializer=tf.truncated_normal_initializer(stddev=stddev),biases_initializer=None)
+        conv = tf.contrib.layers.conv2d_transpose(inputconv, o_d, [f_h, f_w], [s_h, s_w], padding, activation_fn=None, weights_initializer=tf.truncated_normal_initializer(stddev=stddev),biases_initializer=tf.constant_initializer(0.0))
 
         if do_norm:
-            # conv = instance_norm(conv)
-            conv = tf.contrib.layers.batch_norm(conv, decay=0.9, updates_collections=None, epsilon=1e-5, scale=True, scope="batch_norm")
+            conv = instance_norm(conv)
+            # conv = tf.contrib.layers.batch_norm(conv, decay=0.9, updates_collections=None, epsilon=1e-5, scale=True, scope="batch_norm")
 
         if do_relu:
             if(relufactor == 0):

main.py

@@ -40,7 +40,7 @@
 batch_size = 1
 pool_size = 50
 sample_size = 10
-ngf = 64
+ngf = 32
 ndf = 64
 
 

model.py

@@ -21,7 +21,7 @@
 
 batch_size = 1
 pool_size = 50
-ngf = 64
+ngf = 32
 ndf = 64
 
 
@@ -59,7 +59,8 @@ def build_generator_resnet_6blocks(inputgen, name="generator"):
 
         o_c4 = general_deconv2d(o_r6, [batch_size,64,64,ngf*2], ngf*2, ks, ks, 2, 2, 0.02,"SAME","c4")
         o_c5 = general_deconv2d(o_c4, [batch_size,128,128,ngf], ngf, ks, ks, 2, 2, 0.02,"SAME","c5")
-        o_c6 = general_conv2d(o_c5, img_layer, f, f, 1, 1, 0.02,"SAME","c6",do_relu=False)
+        o_c5_pad = tf.pad(inputgen,[[0, 0], [ks, ks], [ks, ks], [0, 0]], "REFLECT")
+        o_c6 = general_conv2d(o_c5_pad, img_layer, f, f, 1, 1, 0.02,"VALID","c6",do_relu=False)
 
         # Adding the tanh layer