def test_generator_run(self):
tf.set_random_seed(1234)
noise = tf.random_normal([100, 64])
image, _ = dcgan.generator(noise)
with self.test_session() as sess:
sess.run(tf.global_variables_initializer())
image.eval()
def _decoder(codes, final_size, is_training, depth=64):
"""Compression decoder."""
decoded_img, _ = dcgan.generator(codes,
depth=depth,
final_size=final_size,
num_outputs=3,
is_training=is_training,
scope='Decoder')
# Map output to [-1, 1].
# Use softsign instead of tanh, as per empirical results of
# http://jmlr.org/proceedings/papers/v9/glorot10a/glorot10a.pdf.
return tf.nn.softsign(decoded_img)
def test_generator_invalid_input(self):
wrong_dim_input = tf.zeros([5, 32, 32])
with self.assertRaises(ValueError):
dcgan.generator(wrong_dim_input)
correct_input = tf.zeros([3, 2])
with self.assertRaisesRegexp(ValueError, 'must be a power of 2'):
dcgan.generator(correct_input, final_size=30)
with self.assertRaisesRegexp(ValueError, 'must be greater than 8'):
dcgan.generator(correct_input, final_size=4)
def generator(noise, is_training=True):
"""Generator to produce CIFAR images.
Args:
noise: A 2D Tensor of shape [batch size, noise dim]. Since this example
does not use conditioning, this Tensor represents a noise vector of some
kind that will be reshaped by the generator into CIFAR examples.
is_training: If `True`, batch norm uses batch statistics. If `False`, batch
norm uses the exponential moving average collected from population
statistics.
Returns:
A single Tensor with a batch of generated CIFAR images.
"""
images, _ = dcgan.generator(noise,
is_training=is_training,
fused_batch_norm=True)
# Make sure output lies between [-1, 1].
return tf.tanh(images)
def conditional_generator(inputs, is_training=True):
"""Generator to produce CIFAR images.
Args:
inputs: A 2-tuple of Tensors (noise, one_hot_labels) and creates a
conditional generator.
is_training: If `True`, batch norm uses batch statistics. If `False`, batch
norm uses the exponential moving average collected from population
statistics.
Returns:
A single Tensor with a batch of generated CIFAR images.
"""
noise, one_hot_labels = inputs
noise = tfgan.features.condition_tensor_from_onehot(noise, one_hot_labels)
images, _ = dcgan.generator(noise,
is_training=is_training,
fused_batch_norm=True)
# Make sure output lies between [-1, 1].
return tf.tanh(images)
def test_generator_graph(self):
tf.set_random_seed(1234)
# Check graph construction for a number of image size/depths and batch
# sizes.
for i, batch_size in zip(xrange(3, 7), xrange(3, 8)):
tf.reset_default_graph()
final_size = 2**i
noise = tf.random_normal([batch_size, 64])
image, end_points = dcgan.generator(noise,
depth=32,
final_size=final_size)
self.assertAllEqual([batch_size, final_size, final_size, 3],
image.shape.as_list())
expected_names = ['deconv%i' % j
for j in xrange(1, i)] + ['logits']
self.assertSetEqual(set(expected_names), set(end_points.keys()))
# Check layer depths.
for j in range(1, i):
layer = end_points['deconv%i' % j]
self.assertEqual(32 * 2**(i - j - 1),
layer.get_shape().as_list()[-1])