def dcgan_encoder(name,
inputs,
n_channels,
latent_dim,
is_training,
mode=None,
nonlinearity=LeakyReLU):
conv2d.set_weights_stdev(0.02)
deconv2d.set_weights_stdev(0.02)
linear.set_weights_stdev(0.02)
output = tf.reshape(inputs, [-1, n_channels, DIM, DIM])
output = Conv2D(name + '.1', 3, DIM, 5, output, stride=2)
output = nonlinearity(output)
output = Conv2D(name + '.2', DIM, 2 * DIM, 5, output, stride=2)
output = Normalize(name + '.BN2', [0, 2, 3], output, is_training, mode)
output = nonlinearity(output)
output = Conv2D(name + '.3', 2 * DIM, 4 * DIM, 5, output, stride=2)
output = Normalize(name + '.BN3', [0, 2, 3], output, is_training, mode)
output = nonlinearity(output)
output = Conv2D(name + '.4', 4 * DIM, 8 * DIM, 5, output, stride=2)
output = Normalize(name + '.BN4', [0, 2, 3], output, is_training, mode)
output = nonlinearity(output)
output = tf.reshape(output, [-1, 4 * 4 * 8 * DIM])
output = Linear(name + '.Output', 4 * 4 * 8 * DIM, latent_dim, output)
conv2d.unset_weights_stdev()
deconv2d.unset_weights_stdev()
linear.unset_weights_stdev()
return output
def high_capacity_encoder(name,
inputs,
n_channels,
latent_dim,
is_training,
mode=None,
nonlinearity=tf.nn.relu):
output = tf.reshape(inputs, [-1, n_channels, DIM, DIM])
output = Conv2D(name + '.0', n_channels, DIM, 5, output, stride=2)
output = Normalize(name + '.BN0', [0, 2, 3], output, is_training, mode)
output = nonlinearity(output)
output = Conv2D(name + '.1', DIM, DIM * 2, 5, output, stride=2)
output = Normalize(name + '.BN1', [0, 2, 3], output, is_training, mode)
output = nonlinearity(output)
output = Conv2D(name + '.2', DIM * 2, DIM * 4, 5, output, stride=2)
output = Normalize(name + '.BN2', [0, 2, 3], output, is_training, mode)
output = nonlinearity(output)
output = tf.reshape(output, [-1, DIM * 4 * 8 * 8])
output = Linear(name + '.FC', DIM * 4 * 8 * 8, DIM * 4 * 8, output)
output = Normalize(name + '.BNFC', [0], output, is_training, mode)
output = nonlinearity(output)
output = Linear(name + '.Output', DIM * 4 * 8, latent_dim, output)
return output
def high_capacity_decoder(name,
z,
n_channels,
is_training,
mode=None,
nonlinearity=tf.nn.relu):
output = Linear(name + '.Input',
z.get_shape().as_list()[1], DIM * 4 * 8 * 8, z)
output = Normalize(name + '.BN0', [0], output, is_training, mode)
output = nonlinearity(output)
output = tf.reshape(output, [-1, DIM * 4, 8, 8])
output = Deconv2D(name + '.1', DIM * 4, DIM * 4, 5, output)
output = Normalize(name + '.BN1', [0, 2, 3], output, is_training, mode)
output = nonlinearity(output)
output = Deconv2D(name + '.2', DIM * 4, DIM * 2, 5, output)
output = Normalize(name + '.BN2', [0, 2, 3], output, is_training, mode)
output = nonlinearity(output)
output = Deconv2D(name + '.3', DIM * 2, DIM // 2, 5, output)
output = Normalize(name + '.BN3', [0, 2, 3], output, is_training, mode)
output = nonlinearity(output)
output = Conv2D(name + '.4', DIM // 2, n_channels, 5, output)
output = tf.reshape(output, [-1, n_channels * DIM * DIM])
return output
def resnet_encoder_new(name,
inputs,
n_channels,
latent_dim,
is_training,
mode=None,
nonlinearity=tf.nn.relu):
output = tf.reshape(inputs, [-1, n_channels, DIM, DIM])
output = Conv2D(name + '.Input',
n_channels,
DIM,
3,
output,
he_init=False,
cpu=CPU)
output = ResidualBlock(name + '.Res1',
DIM,
2 * DIM,
3,
output,
is_training,
mode,
resample='down')
output = ResidualBlock(name + '.Res2',
2 * DIM,
4 * DIM,
3,
output,
is_training,
mode,
resample='down')
output = ResidualBlock(name + '.Res3',
4 * DIM,
8 * DIM,
3,
output,
is_training,
mode,
resample='down')
output = ResidualBlock(name + '.Res4',
8 * DIM,
8 * DIM,
3,
output,
is_training,
mode,
resample='down')
output = Normalize(name + '.BN.5', [0, 2, 3], output, is_training, mode)
output = nonlinearity(output)
output = tf.reshape(output, [-1, 4 * 4 * 8 * DIM])
output = Linear(name + '.Output', 4 * 4 * 8 * DIM, latent_dim, output)
return output
def resnet_decoder(name,
z,
n_channels,
is_training,
mode=None,
nonlinearity=tf.nn.relu):
output = Linear(name + '.Input',
z.get_shape().as_list()[1], 4 * 4 * 8 * DIM, z)
output = Normalize(name + '.BN0', [0], output, is_training, mode)
output = nonlinearity(output)
output = tf.reshape(output, [-1, 8 * DIM, 4, 4])
output = ResidualBlock(name + '.Res1',
8 * DIM,
8 * DIM,
3,
output,
is_training,
mode,
resample='up',
norm_inputs=True)
output = ResidualBlock(name + '.Res2',
8 * DIM,
4 * DIM,
3,
output,
is_training,
mode,
resample='up')
output = ResidualBlock(name + '.Res3',
4 * DIM,
2 * DIM,
3,
output,
is_training,
mode,
resample='up')
output = ResidualBlock(name + '.Res4',
2 * DIM,
1 * DIM,
3,
output,
is_training,
mode,
resample='up')
output = Normalize(name + '.BN5', [0, 2, 3], output, is_training, mode)
output = nonlinearity(output)
output = Conv2D(name + '.Output', DIM, n_channels, 3, output)
output = tf.reshape(output, [-1, n_channels * DIM * DIM])
return output
def ConvMeanPool(name,
input_dim,
output_dim,
filter_size,
inputs,
he_init=True,
biases=True):
output = Conv2D(name,
input_dim,
output_dim,
filter_size,
inputs,
he_init=he_init,
biases=biases,
cpu=CPU)
output = tf.add_n([
output[:, :, ::2, ::2], output[:, :, 1::2, ::2],
output[:, :, ::2, 1::2], output[:, :, 1::2, 1::2]
]) / 4.
return output
def UpsampleConv(name,
input_dim,
output_dim,
filter_size,
inputs,
he_init=True,
biases=True):
output = inputs
output = tf.concat([output, output, output, output], axis=1)
output = tf.transpose(output, [0, 2, 3, 1])
output = tf.depth_to_space(output, 2)
output = tf.transpose(output, [0, 3, 1, 2])
output = Conv2D(name,
input_dim,
output_dim,
filter_size,
output,
he_init=he_init,
biases=biases,
cpu=CPU)
return output