def _conv2d(scope, x, kernel_size, filters, padding='SAME'):
return layers.custom_conv2d(
x=x,
filters=filters,
kernel_size=kernel_size,
padding=padding,
activation=lambda x: layers.pixel_norm(tf.nn.leaky_relu(x)),
he_initializer_slope=0.0,
scope=scope)
def _conv2d(scope, x, kernel_size, filters, padding='SAME'):
return layers.custom_conv2d(
x=x,
filters=filters,
kernel_size=kernel_size,
padding=padding,
activation=lambda x: layers.pixel_norm(tf.nn.leaky_relu(x)),
he_initializer_slope=0.0,
scope=scope)
def generator(z,
progress,
num_filters_fn,
resolution_schedule,
num_blocks=None,
kernel_size=3,
colors=3,
to_rgb_activation=None,
simple_arch=False,
scope='progressive_gan_generator',
reuse=None):
"""Generator network for the progressive GAN model.
Args:
z: A `Tensor` of latent vector. The first dimension must be batch size.
progress: A scalar float `Tensor` of training progress.
num_filters_fn: A function that maps `block_id` to # of filters for the
block.
resolution_schedule: An object of `ResolutionSchedule`.
num_blocks: An integer of number of blocks. None means maximum number of
blocks, i.e. `resolution.schedule.num_resolutions`. Defaults to None.
kernel_size: An integer of convolution kernel size.
colors: Number of output color channels. Defaults to 3.
to_rgb_activation: Activation function applied when output rgb.
simple_arch: Architecture variants for lower memory usage and faster speed
scope: A string or variable scope.
reuse: Whether to reuse `scope`. Defaults to None which means to inherit
the reuse option of the parent scope.
Returns:
A `Tensor` of model output and a dictionary of model end points.
"""
if num_blocks is None:
num_blocks = resolution_schedule.num_resolutions
start_h, start_w = resolution_schedule.start_resolutions
final_h, final_w = resolution_schedule.final_resolutions
def _conv2d(scope, x, kernel_size, filters, padding='SAME'):
return layers.custom_conv2d(
x=x,
filters=filters,
kernel_size=kernel_size,
padding=padding,
activation=lambda x: layers.pixel_norm(tf.nn.leaky_relu(x)),
he_initializer_slope=0.0,
scope=scope)
def _to_rgb(x):
return layers.custom_conv2d(
x=x,
filters=colors,
kernel_size=1,
padding='SAME',
activation=to_rgb_activation,
scope='to_rgb')
he_init = tf_slim.variance_scaling_initializer()
end_points = {}
with tf.variable_scope(scope, reuse=reuse):
with tf.name_scope('input'):
x = tf_slim.flatten(z)
end_points['latent_vector'] = x
with tf.variable_scope(block_name(1)):
if simple_arch:
x_shape = tf.shape(x)
x = tf.layers.dense(x, start_h*start_w*num_filters_fn(1),
kernel_initializer=he_init)
x = tf.nn.relu(x)
x = tf.reshape(x, [x_shape[0], start_h, start_w, num_filters_fn(1)])
else:
x = tf.expand_dims(tf.expand_dims(x, 1), 1)
x = layers.pixel_norm(x)
# Pad the 1 x 1 image to 2 * (start_h - 1) x 2 * (start_w - 1)
# with zeros for the next conv.
x = tf.pad(x, [[0] * 2, [start_h - 1] * 2, [start_w - 1] * 2, [0] * 2])
# The output is start_h x start_w x num_filters_fn(1).
x = _conv2d('conv0', x, (start_h, start_w), num_filters_fn(1), 'VALID')
x = _conv2d('conv1', x, kernel_size, num_filters_fn(1))
lods = [x]
if resolution_schedule.scale_mode == 'H':
strides = (resolution_schedule.scale_base, 1)
else:
strides = (resolution_schedule.scale_base,
resolution_schedule.scale_base)
for block_id in range(2, num_blocks + 1):
with tf.variable_scope(block_name(block_id)):
if simple_arch:
x = tf.layers.conv2d_transpose(
x,
num_filters_fn(block_id),
kernel_size=kernel_size,
strides=strides,
padding='SAME',
kernel_initializer=he_init)
x = tf.nn.relu(x)
else:
x = resolution_schedule.upscale(x, resolution_schedule.scale_base)
x = _conv2d('conv0', x, kernel_size, num_filters_fn(block_id))
x = _conv2d('conv1', x, kernel_size, num_filters_fn(block_id))
lods.append(x)
outputs = []
for block_id in range(1, num_blocks + 1):
with tf.variable_scope(block_name(block_id)):
if simple_arch:
lod = lods[block_id - 1]
lod = tf.layers.conv2d(
lod,
colors,
kernel_size=1,
padding='SAME',
name='to_rgb',
kernel_initializer=he_init)
lod = to_rgb_activation(lod)
else:
lod = _to_rgb(lods[block_id - 1])
scale = resolution_schedule.scale_factor(block_id)
lod = resolution_schedule.upscale(lod, scale)
end_points['upscaled_rgb_{}'.format(block_id)] = lod
# alpha_i is used to replace lod_select. Note sum(alpha_i) is
# garanteed to be 1.
alpha = _generator_alpha(block_id, progress)
end_points['alpha_{}'.format(block_id)] = alpha
outputs.append(lod * alpha)
predictions = tf.add_n(outputs)
batch_size = int(z.shape[0])
predictions.set_shape([batch_size, final_h, final_w, colors])
end_points['predictions'] = predictions
return predictions, end_points
def generator(z,
progress,
num_filters_fn,
resolution_schedule,
num_blocks=None,
kernel_size=3,
colors=3,
to_rgb_activation=None,
simple_arch=False,
scope='progressive_gan_generator',
reuse=None):
"""Generator network for the progressive GAN model.
Args:
z: A `Tensor` of latent vector. The first dimension must be batch size.
progress: A scalar float `Tensor` of training progress.
num_filters_fn: A function that maps `block_id` to # of filters for the
block.
resolution_schedule: An object of `ResolutionSchedule`.
num_blocks: An integer of number of blocks. None means maximum number of
blocks, i.e. `resolution.schedule.num_resolutions`. Defaults to None.
kernel_size: An integer of convolution kernel size.
colors: Number of output color channels. Defaults to 3.
to_rgb_activation: Activation function applied when output rgb.
simple_arch: Architecture variants for lower memory usage and faster speed
scope: A string or variable scope.
reuse: Whether to reuse `scope`. Defaults to None which means to inherit
the reuse option of the parent scope.
Returns:
A `Tensor` of model output and a dictionary of model end points.
"""
if num_blocks is None:
num_blocks = resolution_schedule.num_resolutions
start_h, start_w = resolution_schedule.start_resolutions
final_h, final_w = resolution_schedule.final_resolutions
def _conv2d(scope, x, kernel_size, filters, padding='SAME'):
return layers.custom_conv2d(
x=x,
filters=filters,
kernel_size=kernel_size,
padding=padding,
activation=lambda x: layers.pixel_norm(tf.nn.leaky_relu(x)),
he_initializer_slope=0.0,
scope=scope)
def _to_rgb(x):
return layers.custom_conv2d(
x=x,
filters=colors,
kernel_size=1,
padding='SAME',
activation=to_rgb_activation,
scope='to_rgb')
he_init = tf.contrib.layers.variance_scaling_initializer()
end_points = {}
with tf.variable_scope(scope, reuse=reuse):
with tf.name_scope('input'):
x = tf.contrib.layers.flatten(z)
end_points['latent_vector'] = x
with tf.variable_scope(block_name(1)):
if simple_arch:
x_shape = tf.shape(x)
x = tf.layers.dense(x, start_h*start_w*num_filters_fn(1),
kernel_initializer=he_init)
x = tf.nn.relu(x)
x = tf.reshape(x, [x_shape[0], start_h, start_w, num_filters_fn(1)])
else:
x = tf.expand_dims(tf.expand_dims(x, 1), 1)
x = layers.pixel_norm(x)
# Pad the 1 x 1 image to 2 * (start_h - 1) x 2 * (start_w - 1)
# with zeros for the next conv.
x = tf.pad(x, [[0] * 2, [start_h - 1] * 2, [start_w - 1] * 2, [0] * 2])
# The output is start_h x start_w x num_filters_fn(1).
x = _conv2d('conv0', x, (start_h, start_w), num_filters_fn(1), 'VALID')
x = _conv2d('conv1', x, kernel_size, num_filters_fn(1))
lods = [x]
if resolution_schedule.scale_mode == 'H':
strides = (resolution_schedule.scale_base, 1)
else:
strides = (resolution_schedule.scale_base,
resolution_schedule.scale_base)
for block_id in range(2, num_blocks + 1):
with tf.variable_scope(block_name(block_id)):
if simple_arch:
x = tf.layers.conv2d_transpose(
x,
num_filters_fn(block_id),
kernel_size=kernel_size,
strides=strides,
padding='SAME',
kernel_initializer=he_init)
x = tf.nn.relu(x)
else:
x = resolution_schedule.upscale(x, resolution_schedule.scale_base)
x = _conv2d('conv0', x, kernel_size, num_filters_fn(block_id))
x = _conv2d('conv1', x, kernel_size, num_filters_fn(block_id))
lods.append(x)
outputs = []
for block_id in range(1, num_blocks + 1):
with tf.variable_scope(block_name(block_id)):
if simple_arch:
lod = lods[block_id - 1]
lod = tf.layers.conv2d(
lod,
colors,
kernel_size=1,
padding='SAME',
name='to_rgb',
kernel_initializer=he_init)
lod = to_rgb_activation(lod)
else:
lod = _to_rgb(lods[block_id - 1])
scale = resolution_schedule.scale_factor(block_id)
lod = resolution_schedule.upscale(lod, scale)
end_points['upscaled_rgb_{}'.format(block_id)] = lod
# alpha_i is used to replace lod_select. Note sum(alpha_i) is
# garanteed to be 1.
alpha = _generator_alpha(block_id, progress)
end_points['alpha_{}'.format(block_id)] = alpha
outputs.append(lod * alpha)
predictions = tf.add_n(outputs)
batch_size = z.shape[0].value
predictions.set_shape([batch_size, final_h, final_w, colors])
end_points['predictions'] = predictions
return predictions, end_points
