def testDefaultGeneratorWithConditionalBatchNorm(self):
with tf.Graph().as_default():
# Batch size 8, 32x32x3 images, 10 classes.
z = tf.zeros((8, 128))
y = tf.one_hot(tf.ones((8, ), dtype=tf.int32), 10)
generator = resnet_cifar.Generator(
image_shape=(32, 32, 3),
batch_norm_fn=arch_ops.conditional_batch_norm)
fake_images = generator(z, y=y, is_training=True, reuse=False)
self.assertEqual(fake_images.shape.as_list(), [8, 32, 32, 3])
expected_variables = [
# Name and shape.
("generator/fc_noise/kernel:0", [128, 4096]),
("generator/fc_noise/bias:0", [4096]),
("generator/B1/up_conv_shortcut/kernel:0", [3, 3, 256, 256]),
("generator/B1/up_conv_shortcut/bias:0", [256]),
("generator/B1/bn1/condition/gamma/kernel:0", [10, 256]),
("generator/B1/bn1/condition/beta/kernel:0", [10, 256]),
("generator/B1/up_conv1/kernel:0", [3, 3, 256, 256]),
("generator/B1/up_conv1/bias:0", [256]),
("generator/B1/bn2/condition/gamma/kernel:0", [10, 256]),
("generator/B1/bn2/condition/beta/kernel:0", [10, 256]),
("generator/B1/same_conv2/kernel:0", [3, 3, 256, 256]),
("generator/B1/same_conv2/bias:0", [256]),
("generator/B2/up_conv_shortcut/kernel:0", [3, 3, 256, 256]),
("generator/B2/up_conv_shortcut/bias:0", [256]),
("generator/B2/bn1/condition/gamma/kernel:0", [10, 256]),
("generator/B2/bn1/condition/beta/kernel:0", [10, 256]),
("generator/B2/up_conv1/kernel:0", [3, 3, 256, 256]),
("generator/B2/up_conv1/bias:0", [256]),
("generator/B2/bn2/condition/gamma/kernel:0", [10, 256]),
("generator/B2/bn2/condition/beta/kernel:0", [10, 256]),
("generator/B2/same_conv2/kernel:0", [3, 3, 256, 256]),
("generator/B2/same_conv2/bias:0", [256]),
("generator/B3/up_conv_shortcut/kernel:0", [3, 3, 256, 256]),
("generator/B3/up_conv_shortcut/bias:0", [256]),
("generator/B3/bn1/condition/gamma/kernel:0", [10, 256]),
("generator/B3/bn1/condition/beta/kernel:0", [10, 256]),
("generator/B3/up_conv1/kernel:0", [3, 3, 256, 256]),
("generator/B3/up_conv1/bias:0", [256]),
("generator/B3/bn2/condition/gamma/kernel:0", [10, 256]),
("generator/B3/bn2/condition/beta/kernel:0", [10, 256]),
("generator/B3/same_conv2/kernel:0", [3, 3, 256, 256]),
("generator/B3/same_conv2/bias:0", [256]),
("generator/final_norm/condition/gamma/kernel:0", [10, 256]),
("generator/final_norm/condition/beta/kernel:0", [10, 256]),
("generator/final_conv/kernel:0", [3, 3, 256, 3]),
("generator/final_conv/bias:0", [3]),
]
actual_variables = [(v.name, v.shape.as_list())
for v in tf.trainable_variables()]
for a in actual_variables:
logging.info(a)
for a, e in zip(actual_variables, expected_variables):
logging.info("actual: %s, expected: %s", a, e)
self.assertEqual(a, e)
self.assertEqual(len(actual_variables), len(expected_variables))
def testResNetCifar(self, image_shape):
self.assertArchitectureBuilds(
gen=resnet_cifar.Generator(image_shape=image_shape),
disc=resnet_cifar.Discriminator(),
image_shape=image_shape)