def generator_m2(z, heads=10, training=True, reuse=True, name="generator"):
# Network Architecture is exactly same as in infoGAN (https://arxiv.org/abs/1606.03657)
# Architecture : FC1024_BR-FC7x7x128_BR-(64)4dc2s_BR-(1)4dc2s_S
with tf.variable_scope(name, reuse=reuse):
net = tf.nn.relu(
bn(linear(z, 1024, scope='g_fc1'),
is_training=training,
scope='g_bn1'))
net = tf.nn.relu(
bn(linear(net, 128 * 7 * 7, scope='g_fc2'),
is_training=training,
scope='g_bn2'))
net = tf.reshape(net, [64, 7, 7, 128])
net = tf.nn.relu(
bn(deconv2d(net, [64, 14, 14, 64], 4, 4, 2, 2, name='g_dc3'),
is_training=training,
scope='g_bn3'))
img_sets = []
for i in range(heads):
out = tf.nn.sigmoid(
deconv2d(net, [64, 28, 28, 1],
4,
4,
2,
2,
name='g_dc4_h' + str(i)))
img_sets.append(out)
return img_sets
def discriminator2_32X32_dc(x,
training=True,
reuse=True,
name="discriminator"):
# Network Architecture is exactly same as in infoGAN (https://arxiv.org/abs/1606.03657)
# Architecture : (64)4c2s-(128)4c2s_BL-FC1024_BL-FC1_S
with tf.variable_scope(name, reuse=reuse):
net = lrelu(conv2d(x, 64, 4, 4, 2, 2, name='d_conv1'))
net = lrelu(
bn(conv2d(net, 64 * 2, 4, 4, 2, 2, name='d_conv2'),
is_training=training,
scope='d_bn2'))
net = lrelu(
bn(conv2d(net, 64 * 4, 4, 4, 2, 2, name='d_conv3'),
is_training=training,
scope='d_bn3'))
net = lrelu(
bn(conv2d(net, 64 * 8, 4, 4, 2, 2, name='d_conv4'),
is_training=training,
scope='d_bn4'))
# net = tf.reshape(net, [batch_size, -1])
# net = lrelu(bn(linear(net, 1024, scope='d_fc3'), is_training=training, scope='d_bn3'))
# net = fc(net, 1024)
# net = lrelu(bn(net, is_training=training, scope='d_bn5'))
# net = fc(net, 1024)
# net = lrelu(bn(net, is_training=training, scope='d_bn3'))
# out_logit = linear(net, 1, scope='d_fc4')
# out_logit = linear(tf.reshape(net, [64, -1]), 1, 'd_h4_lin')
# out_logit = linear(net, 1, scope='d_fc4')
# out = tf.nn.sigmoid(out_logit)
out_logit = fc(net, 1)
return out_logit
def mgan_dis(x, training=True, reuse=True, name = "discriminator"):
with tf.variable_scope(name, reuse=reuse):
net = lrelu(conv2d(x, 128, 5, 5, 2, 2, name='d_conv1'))
net = lrelu(bn(conv2d(net, 256, 5, 5, 2, 2, name='d_conv2'), is_training=training, scope='d_bn2'))
net = lrelu(bn(conv2d(net, 512, 5, 5, 2, 2, name='d_conv3'), is_training=training, scope='d_bn3'))
# net = tf.reshape(net, [batch_size, -1])
# net = lrelu(bn(linear(net, 1024, scope='d_fc3'), is_training=training, scope='d_bn3'))
out_logit = fc(net,1)
# net = lrelu(bn(net, is_training=training, scope='d_bn3'))
# out_logit = linear(net, 1, scope='d_fc4')
return out_logit
def allconvnet_mnist(x,
out_dim=10,
name="classifier",
training=True,
reuse=True):
# w_init = tf.variance_scaling_initializer(scale=2., mode='fan_in')
bn = partial(batch_norm, is_training=training)
# conv_relu = partial(conv, activation_fn=relu)#, weights_initializer=w_init)
conv_bn_relu = partial(
conv, activation_fn=relu, normalizer_fn=bn,
biases_initializer=None) #, weights_initializer=w_init)
with tf.variable_scope(name, reuse=reuse):
# Convolutional Layer #1
y = conv_bn_relu(x, 64, 3, 1)
y = conv_bn_relu(y, 64, 3, 1)
y = conv_bn_relu(y, 64, 3, 2)
y = conv_bn_relu(y, 128, 3, 1)
y = conv_bn_relu(y, 128, 3, 1)
y = conv_bn_relu(y, 128, 3, 2)
y = conv_bn_relu(y, out_dim, 1, 1)
# print(y.shape)
y = tf.reduce_mean(y, [1, 2])
y = bn(y)
#restraint layer
y_max = tf.reduce_max(y, axis=1, keep_dims=True)
y = tf.exp(y - y_max)
y = y / tf.norm(y, ord=2, axis=1, keep_dims=True)
return y
def mgan_gen(z, heads=10,training=True, reuse=True, name = "generator"):
with tf.variable_scope(name, reuse=reuse):
net = tf.nn.relu(bn(linear(z, 4*4*512, scope='g_fc1'), is_training=training, scope='g_bn1'))
# net = tf.nn.relu(bn(linear(net, 128 * 7 * 7, scope='g_fc2'), is_training=training, scope='g_bn2'))
net = tf.reshape(net, [64, 4, 4, 512])
net = tf.nn.relu(
bn(deconv2d(net, [64, 8, 8, 256], 5, 5, 2, 2, name='g_dc2'), is_training=training,
scope='g_bn2'))
net = tf.nn.relu(
bn(deconv2d(net, [64, 16, 16, 256], 5, 5, 2, 2, name='g_dc3'), is_training=training,
scope='g_bn3'))
img_sets = []
for i in range(heads):
out = tf.nn.sigmoid(deconv2d(net, [64, 32, 32, 3], 5, 5, 2, 2, name='g_dc4_h' + str(i)))
img_sets.append(out)
return img_sets
def generator_m2_32X32_dc(z, heads=10, training=True, reuse=True, name="generator"):
# Network Architecture is exactly same as in infoGAN (https://arxiv.org/abs/1606.03657)
# Architecture : FC1024_BR-FC7x7x128_BR-(64)4dc2s_BR-(1)4dc2s_S
with tf.variable_scope(name, reuse=reuse):
net = tf.nn.relu(bn(linear(z, 2048, scope='g_fc1'), is_training=training, scope='g_bn1'))
net = tf.reshape(net, [64, 2, 2, 512])
net = tf.nn.relu(
bn(deconv2d(net, [64, 4, 4, 4*64], 4, 4, 2, 2, name='g_dc2'), is_training=training,
scope='g_bn2'))
net = tf.nn.relu(
bn(deconv2d(net, [64, 8, 8, 2 * 64], 4, 4, 2, 2, name='g_dc3'), is_training=training,
scope='g_bn3'))
net = tf.nn.relu(
bn(deconv2d(net, [64, 16, 16, 2 * 64], 4, 4, 2, 2, name='g_dc4'), is_training=training,
scope='g_bn4'))
img_sets = []
for i in range(heads):
out = tf.nn.tanh(deconv2d(net, [64, 32, 32, 3], 4, 4, 2, 2, name='g_dc5_h' + str(i)))
img_sets.append(out)
return img_sets
