def latent_classifier(self, z):
net = self.arch['classifier']['latent']
x = tf.transpose(z, [0, 2, 1]) # [N, z_dim, n_frames]
x = tf.expand_dims(x, -1) # [N, z_dim, n_frames, 1]
for i, (o, k, s) in enumerate(zip(net['output'], net['kernel'], net['stride'])):
with tf.variable_scope('Conv-{}'.format(i)):
x_tanh = tf.layers.conv2d_transpose(x, o, k, s,
padding='same',
data_format='channels_first',
)
x_tanh = Layernorm(x_tanh, [1, 2, 3], 'Layernorm-{}-tanh'.format(i))
x_sigmoid = tf.layers.conv2d_transpose(x, o, k, s,
padding='same',
data_format='channels_first',
)
x_sigmoid = Layernorm(x_sigmoid, [1, 2, 3], 'Layernorm-{}-sigmoid'.format(i))
# GLU
with tf.variable_scope('GLU'):
x = tf.sigmoid(x_sigmoid) * tf.tanh(x_tanh)
# carefully design the architecture so that now x has shape [N, C, n_frames, 1]
batch_size, c, n_frames, w = x.get_shape().as_list()
x = tf.transpose(x, perm=[0, 2, 1, 3]) # [N, n_frames, C, 1]
x = tf.squeeze(x, axis=[-1]) # [N, n_frames, C]
x = tf.layers.dense(x, self.arch['y_dim']) # [N, n_frames, 1]
return tf.reduce_mean(x, axis=1) #[N, 1]
def _generator(self, z, y, is_training=None):
net = self.arch['generator']
h, w, c = net['hwc']
if y is not None:
y = tf.nn.embedding_lookup(self.y_emb, y)
x = self._merge([z, y], h * w * c)
else:
x = z
print "generate x shape:", x.get_shape().as_list()
x = tf.reshape(x, [-1, c, h, w]) # channel first
print "generate x reshape:", x.get_shape().as_list()
for i, (o, k, s) in enumerate(
zip(net['output'], net['kernel'], net['stride'])):
x = tf.layers.conv2d_transpose(
x,
o,
k,
s,
padding='same',
data_format='channels_first',
)
if i < len(net['output']) - 1:
x = Layernorm(x, [1, 2, 3], 'ConvT-LN{}'.format(i))
x = lrelu(x)
print("ConvT-LN{}, shape:".format(i), x.get_shape().as_list())
print "generate x shape:", x.get_shape().as_list()
return x
def _generator(self, z, y, net):
with tf.device("/gpu:3"):
x = tf.expand_dims(z, 1) # [N, 1, n_frames, z_dim]
x = tf.transpose(x, perm=[0, 3, 2, 1]) # [N, z_dim, n_frames, 1]
y = tf.nn.embedding_lookup(self.y_emb,
y) # [N, n_frames, y_emb_dim]
y = tf.expand_dims(y, 1) # [N, 1, n_frames, y_emb_dim]
for i, (o, k, s) in enumerate(
zip(net['output'], net['kernel'], net['stride'])):
with tf.variable_scope('Conv-{}'.format(i)):
# concat y along channel axis
y_transpose = tf.transpose(
y, perm=[0, 3, 2, 1]) # [N, y_emb_dim, n_frames, 1]
x = tf.concat(
[x, y_transpose],
axis=1) # [N, channels + y_emb_dim, n_frames, 1]
if i < len(net['output']) - 1:
x_tanh = tf.layers.conv2d_transpose(
x,
o,
k,
s,
padding='same',
data_format='channels_first',
)
x_tanh = Layernorm(x_tanh, [1, 2, 3],
'Layernorm-{}-tanh'.format(i))
x_sigmoid = tf.layers.conv2d_transpose(
x,
o,
k,
s,
padding='same',
data_format='channels_first',
)
x_sigmoid = Layernorm(x_sigmoid, [1, 2, 3],
'Layernorm-{}-sigmoid'.format(i))
# GLU
with tf.variable_scope('GLU'):
x = tf.sigmoid(x_sigmoid) * tf.tanh(x_tanh)
else:
x = tf.layers.conv2d_transpose(
x,
o,
k,
s,
padding='same',
data_format='channels_first',
)
x = tf.squeeze(x, axis=[-1]) # [N, C, n_frames]
x = tf.transpose(x, perm=[0, 2, 1]) # [N, n_frames, C]
return x
