Merge pull request vanhuyz#20 from dereklh4/support_all_image_sizes

Support all image sizes

generator.py

@@ -38,7 +38,7 @@ def __call__(self, input):
       u64 = ops.uk(res_output, 2*self.ngf, is_training=self.is_training, norm=self.norm,
           reuse=self.reuse, name='u64')                                 # (?, w/2, h/2, 64)
       u32 = ops.uk(u64, self.ngf, is_training=self.is_training, norm=self.norm,
-          reuse=self.reuse, name='u32')                                 # (?, w, h, 32)
+          reuse=self.reuse, name='u32', output_size=self.image_size)         # (?, w, h, 32)
 
       # conv layer
       # Note: the paper said that ReLU and _norm were used

model.py

@@ -51,7 +51,7 @@ def __init__(self,
     self.G = Generator('G', self.is_training, ngf=ngf, norm=norm, image_size=image_size)
     self.D_Y = Discriminator('D_Y',
         self.is_training, norm=norm, use_sigmoid=use_sigmoid)
-    self.F = Generator('F', self.is_training, norm=norm)
+    self.F = Generator('F', self.is_training, norm=norm, image_size=image_size)
     self.D_X = Discriminator('D_X',
         self.is_training, norm=norm, use_sigmoid=use_sigmoid)
 

ops.py

@@ -94,7 +94,7 @@ def n_res_blocks(input, reuse, norm='instance', is_training=True, n=6):
     input = output
   return output
 
-def uk(input, k, reuse=False, norm='instance', is_training=True, name=None):
+def uk(input, k, reuse=False, norm='instance', is_training=True, name=None, output_size=None):
   """ A 3x3 fractional-strided-Convolution-BatchNorm-ReLU layer
       with k filters, stride 1/2
   Args:
@@ -104,6 +104,7 @@ def uk(input, k, reuse=False, norm='instance', is_training=True, name=None):
     is_training: boolean or BoolTensor
     reuse: boolean
     name: string, e.g. 'c7sk-32'
+    output_size: integer, desired output size of layer
   Returns:
     4D tensor
   """
@@ -113,7 +114,8 @@ def uk(input, k, reuse=False, norm='instance', is_training=True, name=None):
     weights = _weights("weights",
       shape=[3, 3, k, input_shape[3]])
 
-    output_size = input_shape[1]*2
+    if not output_size:
+      output_size = input_shape[1]*2
     output_shape = [input_shape[0], output_size, output_size, k]
     fsconv = tf.nn.conv2d_transpose(input, weights,
         output_shape=output_shape,