def create_acvideo_discriminator(clips,
actions,
ndf=64,
norm_layer='instance',
use_noise=False,
noise_sigma=None):
norm_layer = ops.get_norm_layer(norm_layer)
layers = []
paddings = [[0, 0], [0, 0], [1, 1], [1, 1], [0, 0]]
clips = clips * 2 - 1
clip_pairs = tf.concat([clips[:-1], clips[1:]], axis=-1)
clip_pairs = tile_concat([clip_pairs, actions[..., None, None, :]],
axis=-1)
clip_pairs = tf_utils.transpose_batch_time(clip_pairs)
with tf.variable_scope("acvideo_layer_1"):
h1 = noise(clip_pairs, use_noise, noise_sigma)
h1 = conv3d(tf.pad(h1, paddings),
ndf,
kernel_size=(3, 4, 4),
strides=(1, 2, 2),
padding='VALID',
use_bias=False)
h1 = lrelu(h1, 0.2)
layers.append(h1)
with tf.variable_scope("acvideo_layer_2"):
h2 = noise(h1, use_noise, noise_sigma)
h2 = conv3d(tf.pad(h2, paddings),
ndf * 2,
kernel_size=(3, 4, 4),
strides=(1, 2, 2),
padding='VALID',
use_bias=False)
h2 = norm_layer(h2)
h2 = lrelu(h2, 0.2)
layers.append(h2)
with tf.variable_scope("acvideo_layer_3"):
h3 = noise(h2, use_noise, noise_sigma)
h3 = conv3d(tf.pad(h3, paddings),
ndf * 4,
kernel_size=(3, 4, 4),
strides=(1, 2, 2),
padding='VALID',
use_bias=False)
h3 = norm_layer(h3)
h3 = lrelu(h3, 0.2)
layers.append(h3)
with tf.variable_scope("acvideo_layer_4"):
logits = conv3d(tf.pad(h3, paddings),
1,
kernel_size=(3, 4, 4),
strides=(1, 2, 2),
padding='VALID',
use_bias=False)
layers.append(logits)
return nest.map_structure(tf_utils.transpose_batch_time, layers)
def create_video_discriminator(clips, ndf=64, norm_layer='instance'):
norm_layer = ops.get_norm_layer(norm_layer)
layers = []
paddings = [[0, 0], [0, 0], [1, 1], [1, 1], [0, 0]]
clips = tf_utils.transpose_batch_time(clips)
with tf.variable_scope("video_layer_1"):
h1 = conv3d(tf.pad(clips, paddings),
ndf,
kernel_size=4,
strides=(1, 2, 2),
padding='VALID')
h1 = lrelu(h1, 0.2)
layers.append(h1)
with tf.variable_scope("video_layer_2"):
h2 = conv3d(tf.pad(h1, paddings),
ndf * 2,
kernel_size=4,
strides=(1, 2, 2),
padding='VALID')
h2 = norm_layer(h2)
h2 = lrelu(h2, 0.2)
layers.append(h2)
with tf.variable_scope("video_layer_3"):
h3 = conv3d(tf.pad(h2, paddings),
ndf * 4,
kernel_size=4,
strides=(1, 2, 2),
padding='VALID')
h3 = norm_layer(h3)
h3 = lrelu(h3, 0.2)
layers.append(h3)
with tf.variable_scope("video_layer_4"):
if h3.shape[1].value < 4:
kernel_size = (h3.shape[1].value, 4, 4)
else:
kernel_size = 4
logits = conv3d(h3,
1,
kernel_size=kernel_size,
strides=1,
padding='VALID')
layers.append(logits)
return nest.map_structure(tf_utils.transpose_batch_time, layers)
def conv3d(inputs, *args, **kwargs):
kwargs.setdefault('padding', 'VALID')
kwargs.setdefault('use_spectral_norm', True)
return ops.conv3d(tf.pad(inputs, paddings), *args, **kwargs)