def _get_conv(self,
inputs,
in_channels,
out_channels,
scale,
suffix,
kernel_size=(3, 3),
strides=(1, 1)):
"""Performs a convolution in the ResNet block."""
if inputs.get_shape().as_list()[-1] != in_channels:
raise ValueError("Unexpected number of input channels.")
if scale not in ["up", "down", "none"]:
raise ValueError(
"Scale: got {}, expected 'up', 'down', or 'none'.".format(
scale))
outputs = inputs
if scale == "up":
outputs = unpool(outputs)
outputs = ops.conv2d(outputs,
output_dim=out_channels,
k_h=kernel_size[0],
k_w=kernel_size[1],
d_h=strides[0],
d_w=strides[1],
use_sn=self._spectral_norm,
name="{}_{}".format(
"same" if scale == "none" else scale, suffix))
if scale == "down":
outputs = tf.nn.pool(outputs, [2, 2],
"AVG",
"SAME",
strides=[2, 2],
name="pool_%s" % suffix)
return outputs
def apply(self, x, is_training):
num_units = self.nef
num_layers = self.nel
for i in range(num_layers):
scale = 2**(num_layers - i - 1)
x = ops.conv2d(x,
num_units // scale,
k_w=5,
k_h=5,
d_h=2,
d_w=2,
stddev=0.0099999,
name='h%d_conv' % i)
x = ops.batch_norm(x, is_training, name='h%d_bn' % i)
x = tf.nn.relu(x)
x = tf.reshape(x, [x.shape[0], -1])
mu = ops.linear(x, self.dim_z, scope='mu')
log_sigma = ops.linear(x, self.dim_z, scope='log_sigma')
return mu, log_sigma, reparametric(mu, log_sigma)
def apply(self, z, y, is_training):
"""Build the generator network for the given inputs.
Args:
z: `Tensor` of shape [batch_size, z_dim] with latent code.
y: `Tensor` of shape [batch_size, num_classes] with one hot encoded
labels.
is_training: boolean, are we in train or eval model.
Returns:
A tensor of size [batch_size] + self._image_shape with values in [0, 1].
"""
shape_or_none = lambda t: None if t is None else t.shape
logging.info("[Generator] inputs are z=%s, y=%s", z.shape,
shape_or_none(y))
# Each block upscales by a factor of 2.
seed_size = 4
z_dim = z.shape[1].value
in_channels, out_channels = self._get_in_out_channels()
num_blocks = len(in_channels)
if self._embed_z:
z = ops.linear(z,
z_dim,
scope="embed_z",
use_sn=False,
use_bias=self._embed_bias)
if self._embed_y:
y = ops.linear(y,
self._embed_y_dim,
scope="embed_y",
use_sn=False,
use_bias=self._embed_bias)
y_per_block = num_blocks * [y]
if self._hierarchical_z:
z_per_block = tf.split(z, num_blocks + 1, axis=1)
z0, z_per_block = z_per_block[0], z_per_block[1:]
if y is not None:
y_per_block = [tf.concat([zi, y], 1) for zi in z_per_block]
else:
z0 = z
z_per_block = num_blocks * [z]
logging.info("[Generator] z0=%s, z_per_block=%s, y_per_block=%s",
z0.shape, [str(shape_or_none(t)) for t in z_per_block],
[str(shape_or_none(t)) for t in y_per_block])
# Map noise to the actual seed.
net = ops.linear(z0,
in_channels[0] * seed_size * seed_size,
scope="fc_noise",
use_sn=self._spectral_norm)
# Reshape the seed to be a rank-4 Tensor.
net = tf.reshape(net, [-1, seed_size, seed_size, in_channels[0]],
name="fc_reshaped")
for block_idx in range(num_blocks):
name = "B{}".format(block_idx + 1)
block = self._resnet_block(name=name,
in_channels=in_channels[block_idx],
out_channels=out_channels[block_idx],
scale="up")
net = block(net,
z=z_per_block[block_idx],
y=y_per_block[block_idx],
is_training=is_training)
if name in self._blocks_with_attention:
logging.info("[Generator] Applying non-local block to %s",
net.shape)
net = ops.non_local_block(net,
"non_local_block",
use_sn=self._spectral_norm)
# Final processing of the net.
# Use unconditional batch norm.
logging.info("[Generator] before final processing: %s", net.shape)
net = ops.batch_norm(net, is_training=is_training, name="final_norm")
net = tf.nn.relu(net)
net = ops.conv2d(net,
output_dim=self._image_shape[2],
k_h=3,
k_w=3,
d_h=1,
d_w=1,
name="final_conv",
use_sn=self._spectral_norm)
logging.info("[Generator] after final processing: %s", net.shape)
net = (tf.nn.tanh(net) + 1.0) / 2.0
return net
def apply(self, x, y, is_training):
"""Apply the discriminator on a input.
Args:
x: `Tensor` of shape [batch_size, ?, ?, ?] with real or fake images.
y: `Tensor` of shape [batch_size, num_classes] with one hot encoded
labels.
is_training: Boolean, whether the architecture should be constructed for
training or inference.
Returns:
Tuple of 3 Tensors, the final prediction of the discriminator, the logits
before the final output activation function and logits form the second
last layer.
"""
logging.info("[Discriminator] inputs are x=%s, y=%s", x.shape,
None if y is None else y.shape)
resnet_ops.validate_image_inputs(x)
in_channels, out_channels = self._get_in_out_channels(
colors=x.shape[-1].value, resolution=x.shape[1].value)
num_blocks = len(in_channels)
net = ops.conv2d(x,
output_dim=in_channels[0],
k_h=3,
k_w=3,
d_h=1,
d_w=1,
name="initial_conv",
use_sn=self._spectral_norm)
for block_idx in range(num_blocks):
scale = "down" if block_idx % 2 == 0 else "none"
block = self._resnet_block(name="B{}".format(block_idx + 1),
in_channels=in_channels[block_idx],
out_channels=out_channels[block_idx],
scale=scale)
net = block(net, z=None, y=y, is_training=is_training)
# At resolution 64x64 there is a self-attention block.
if scale == "none" and net.shape[1].value == 64:
logging.info("[Discriminator] Applying non-local block to %s",
net.shape)
net = ops.non_local_block(net,
"non_local_block",
use_sn=self._spectral_norm)
# Final part
logging.info("[Discriminator] before final processing: %s", net.shape)
net = tf.nn.relu(net)
h = tf.math.reduce_sum(net, axis=[1, 2])
out_logit = ops.linear(h,
1,
scope="final_fc",
use_sn=self._spectral_norm)
logging.info("[Discriminator] after final processing: %s", net.shape)
if self._project_y:
if y is None:
raise ValueError(
"You must provide class information y to project.")
with tf.variable_scope("embedding_fc"):
y_embedding_dim = out_channels[-1]
# We do not use ops.linear() below since it does not have an option to
# override the initializer.
kernel = tf.get_variable(
"kernel", [y.shape[1], y_embedding_dim],
tf.float32,
initializer=tf.initializers.glorot_normal())
if self._spectral_norm:
kernel = ops.spectral_norm(kernel)
embedded_y = tf.matmul(y, kernel)
logging.info("[Discriminator] embedded_y for projection: %s",
embedded_y.shape)
out_logit += tf.reduce_sum(embedded_y * h,
axis=1,
keepdims=True)
out = tf.nn.sigmoid(out_logit)
return out, out_logit, h
def apply(self, z, y, is_training):
"""Build the generator network for the given inputs.
Args:
z: `Tensor` of shape [batch_size, z_dim] with latent code.
y: `Tensor` of shape [batch_size, num_classes] with one hot encoded
labels.
is_training: boolean, are we in train or eval model.
Returns:
A tensor of size [batch_size] + self._image_shape with values in [0, 1].
"""
shape_or_none = lambda t: None if t is None else t.shape
logging.info("[Generator] inputs are z=%s, y=%s", z.shape,
shape_or_none(y))
seed_size = 4
if self._embed_y:
y = ops.linear(y,
self._embed_y_dim,
scope="embed_y",
use_sn=False,
use_bias=False)
if y is not None:
y = tf.concat([z, y], axis=1)
z = y
in_channels, out_channels = self._get_in_out_channels()
num_blocks = len(in_channels)
# Map noise to the actual seed.
net = ops.linear(z,
in_channels[0] * seed_size * seed_size,
scope="fc_noise",
use_sn=self._spectral_norm)
# Reshape the seed to be a rank-4 Tensor.
net = tf.reshape(net, [-1, seed_size, seed_size, in_channels[0]],
name="fc_reshaped")
for block_idx in range(num_blocks):
scale = "none" if block_idx % 2 == 0 else "up"
block = self._resnet_block(name="B{}".format(block_idx + 1),
in_channels=in_channels[block_idx],
out_channels=out_channels[block_idx],
scale=scale)
net = block(net, z=z, y=y, is_training=is_training)
# At resolution 64x64 there is a self-attention block.
if scale == "up" and net.shape[1].value == 64:
logging.info("[Generator] Applying non-local block to %s",
net.shape)
net = ops.non_local_block(net,
"non_local_block",
use_sn=self._spectral_norm)
# Final processing of the net.
# Use unconditional batch norm.
logging.info("[Generator] before final processing: %s", net.shape)
net = ops.batch_norm(net, is_training=is_training, name="final_norm")
net = tf.nn.relu(net)
colors = self._image_shape[2]
if self._experimental_fast_conv_to_rgb:
net = ops.conv2d(net,
output_dim=128,
k_h=3,
k_w=3,
d_h=1,
d_w=1,
name="final_conv",
use_sn=self._spectral_norm)
net = net[:, :, :, :colors]
else:
net = ops.conv2d(net,
output_dim=colors,
k_h=3,
k_w=3,
d_h=1,
d_w=1,
name="final_conv",
use_sn=self._spectral_norm)
logging.info("[Generator] after final processing: %s", net.shape)
net = (tf.nn.tanh(net) + 1.0) / 2.0
return net