update: image scaling function at 'res_block'

BigGAN/biggan_model.py

@@ -98,32 +98,35 @@ def __init__(self, s, batch_size=64, height=128, width=128, channel=3, n_classes
  # Placeholders
  self.x = tf.placeholder(tf.float32,
  shape=[None, self.height, self.width, self.channel],
- name="x-image") # (64, 64, 64, 3)
+ name="x-image") # (bs, 128, 128, 3)
  self.y = tf.placeholder(tf.float32,
  shape=[None, self.n_classes],
- name="y-label") # (64, n_classes)
+ name="y-label") # (bs, n_classes)
  self.z = tf.placeholder(tf.float32, shape=[None, self.z_dim], name="z-noise") # (-1, 128)
 
- self.build_sagan() # build SAGAN model
+ self.build_sagan() # build BigGAN model
 
  @staticmethod
- def res_block(x, c, z, f, scale_type, name):
+ def res_block(x, f, scale_type, name):
  with tf.variable_scope("res_block_up-%s" % name):
  assert scale_type in ["up", "down"]
  scale_up = False if scale_type == "down" else True
 
  ssc = x
- cz = tf.concat([c, z], axis=-1)
 
- x = t.batch_norm(tf.concat([x, cz], axis=-1), name="bn-1")
+ x = t.batch_norm(x, name="bn-1")
  x = tf.nn.relu(x)
- x = t.conv2d_alt(x, f, s=2, sn=True, name="conv2d-1") if not scale_up \
- else t.deconv2d_alt(x, f, s=2, sn=True, name="deconv2d-1")
+ x = t.conv2d_alt(x, f, sn=True, name="conv2d-1")
 
- x = t.batch_norm(tf.concat([x, cz], axis=-1), name="bn-2")
+ x = t.batch_norm(x, name="bn-2")
  x = tf.nn.relu(x)
- x = t.conv2d_alt(x, f, s=2, sn=True, name="conv2d-2") if not scale_up \
- else t.deconv2d_alt(x, f, s=2, sn=True, name="deconv2d-2")
+
+ if not scale_up:
+ x = t.conv2d_alt(x, f, sn=True, name="conv2d-2")
+ x = tf.layers.average_pooling2d(x, pool_size=(2, 2))
+ else:
+ x = t.deconv2d_alt(x, f, sn=True, name="up-sampling")
+
  return x + ssc
 
  @staticmethod
@@ -229,7 +232,6 @@ def generator(self, z, c=None, reuse=None):
  res = x
  for i in range(4):
  res = self.res_block(res,
- c=None, z=z[i],
  f=(16 // (2 ** i)) * self.channel,
  scale_type="up",
  name="res%d" % (i + 1))