updated synthesis model to produce 16x16x16 output, see #10

python/mrtoct/model/gan/synthesis.py

@@ -3,9 +3,10 @@
 from mrtoct.model import layers
 
 
-def generator_conv(inputs, kernel_size, num_filters, activation=tf.nn.relu):
+def generator_conv(inputs, kernel_size, num_filters, padding,
+                   activation=tf.nn.relu):
   """Creates a synthesis generator conv layer."""
-  outputs = layers.Conv3D(num_filters, kernel_size)(inputs)
+  outputs = layers.Conv3D(num_filters, kernel_size, padding=padding)(inputs)
   outputs = layers.BatchNorm()(outputs)
   outputs = layers.Activation(activation)(outputs)
 
@@ -18,17 +19,18 @@ def generator_network(params):
 
   for i, ks in enumerate([9, 3, 3, 3, 9, 3, 3, 7, 3]):
     with tf.variable_scope(f'conv{i}'):
-      outputs = generator_conv(outputs, ks, 64 if i in [4, 5] else 32)
+      outputs = generator_conv(outputs, ks, 64 if i in [4, 5] else 32,
+                               'valid' if ks == 9 else 'same')
 
   with tf.variable_scope('final'):
-    outputs = generator_conv(outputs, 3, 1, tf.nn.tanh)
+    outputs = generator_conv(outputs, 3, 1, 'same', activation=tf.nn.tanh)
 
   return layers.Network(inputs, outputs, name='generator')
 
 
 def discriminator_conv(inputs, num_filters):
   """Creates a synthesis discriminator conv layer."""
-  outputs = layers.Conv3D(num_filters, 5, 1)(inputs)
+  outputs = layers.Conv3D(num_filters, 5)(inputs)
   outputs = layers.BatchNorm()(outputs)
   outputs = layers.Activation(tf.nn.relu)(outputs)
   outputs = layers.MaxPool3D(5, 1)(outputs)

python/mrtoct/model/gan/synthesis_test.py

@@ -14,7 +14,7 @@ def test_generator(self):
     x = tf.ones([10, 32, 32, 32, 1])
     y = network(x)
 
-    self.assertAllEqual(x.shape, y.shape)
+    self.assertAllEqual(tf.TensorShape([10, 16, 16, 16, 1]), y.shape)
 
   def test_discriminator(self):
     network = model.synthesis.discriminator_network(self.params)

python/mrtoct/model/layers.py

@@ -13,10 +13,11 @@ def __init__(self, *args, **kwargs):
 class Conv3D(tf.layers.Conv3D):
   """Same as tf.layers.Conv3D but with better defaults."""
 
-  def __init__(self, *args, **kwargs):
+  def __init__(self, num_filters, kernel_size, padding='same'):
     init = tf.contrib.layers.xavier_initializer()
 
-    super().__init__(*args, kernel_initializer=init, padding='same', **kwargs)
+    super().__init__(num_filters, kernel_size,
+                     padding=padding, kernel_initializer=init)
 
 
 class Conv2DTranspose(tf.layers.Conv2DTranspose):

python/patch.py

@@ -84,7 +84,7 @@ def main(args):
   tf.logging.set_verbosity(tf.logging.INFO)
 
   hparams = tf.contrib.training.HParams(
-      sample_num=1000,
+      sample_num=8000,
       seeds=[args.seed,
              args.seed * 2,
              args.seed * 3],