def Discriminator(input, reuse=False):
with tf.variable_scope("Discriminator") as scope:
if reuse:
scope.reuse_variables()
image = tf.reshape(input, [-1, 28, 28, 1])
conv1 = tlib.Con2D(image, FLAGS.DIM, 5, 2, scope="conv1")
relu1 = tlib.leaky_relu(conv1)
conv2 = tlib.Con2D(relu1, 2 * FLAGS.DIM, 5, 2, scope="conv2")
relu2 = tlib.leaky_relu(conv2)
conv3 = tlib.Con2D(relu2, 4 * FLAGS.DIM, 5, 2, scope="conv3")
relu3 = tlib.leaky_relu(conv3)
out_put = tf.reshape(relu3, [-1, 4 * 4 * 4 * FLAGS.DIM])
fc1_source = tlib.fc(out_put, 1, scope="fc1")
fc2_class = tlib.fc(out_put, FLAGS.n_class, scope="fc2")
fc3_con = tlib.fc(out_put, 2, scope="fc3")
n_class_ = tf.nn.softmax(fc2_class, name="class")
return fc1_source, fc2_class, fc3_con, n_class_
def Generator_k(z, labels=None, reuse=False, nums=50):
with tf.variable_scope("Generator") as scope:
if reuse:
scope.reuse_variables()
if z is None:
z = tf.random_normal([nums, FLAGS.z_dim])
if labels is not None:
z = tf.concat([z, labels], 1)
z_labels = tlib.fc(z, 4 * 4 * 4 * FLAGS.DIM, scope="project")
bn1 = tlib.bn(z_labels, scope="bn1")
out_put = tf.nn.relu(bn1)
out_put = tf.reshape(out_put, [-1, 4, 4, 4 * FLAGS.DIM])
dconv1 = tlib.Con2D_transpose(out_put, [nums, 8, 8, 2 * FLAGS.DIM],
5,
2,
scope="conv2D_transpose1")
bnconv1 = tlib.bn(dconv1, scope="bn2")
h1 = tf.nn.relu(bnconv1)
dconv2 = tlib.Con2D_transpose(h1, [nums, 16, 16, FLAGS.DIM],
5,
2,
scope="conv2D_transpose2")
bnconv2 = tlib.bn(dconv2, scope="bn3")
h2 = tf.nn.relu(bnconv2)
dconv3 = tlib.Con2D_transpose(h2, [nums, 32, 32, FLAGS.input_channel],
5,
2,
scope="conv2D_transpose3")
h3 = tf.tanh(dconv3)
return tf.reshape(h3, [-1, FLAGS.Out_DIm])
def Generator(z,labels=None,reuse=False,nums=50):
with tf.variable_scope("Generator") as scope:
#z_labels = tf.concat([z,labels],1)
if reuse:
scope.reuse_variables()
if z is None:
z = tf.random_normal([nums,FLAGS.z_dim])
if labels is not None:
z = tf.concat([z,labels],1)
oh,ow =flags.FLAGS.out_height,flags.FLAGS.out_width
z_labels = tlib.fc(z,4*4*4*FLAGS.DIM,scope="project")
out_put = tf.nn.relu(z_labels)
out_put = tf.reshape(out_put,[-1,4,4,4*FLAGS.DIM])
dconv1 = tlib.Con2D_transpose(out_put,[nums,8,8,2*FLAGS.DIM],5,2,scope="conv2D_transpose1")
h1 = tf.nn.relu(dconv1)
h1 = h1[:,:7,:7,:]
dconv2=tlib.Con2D_transpose(h1,[nums,14,14,FLAGS.DIM],5,2,scope="conv2D_transpose2")
h2= tf.nn.relu(dconv2)
dconv3 = tlib.Con2D_transpose(h2,[nums,28,28,FLAGS.input_channel],5,2,scope="conv2D_transpose3")
h3 = tf.nn.sigmoid(dconv3)
return tf.reshape(h3,[-1,FLAGS.Out_DIm])
def Discriminator_k(input, reuse=False):
with tf.variable_scope("Discriminator") as scope:
if reuse:
scope.reuse_variables()
#image = tf.transpose(tf.reshape(input,[-1,3,32,32]),perm=[0,2,3,1])
conv1 = tlib.Con2D(input, FLAGS.DIM, 5, 2, scope="conv1")
relu1 = tlib.leaky_relu(conv1)
conv2 = tlib.Con2D(relu1, 2 * FLAGS.DIM, 5, 2, scope="conv2")
relu2 = tlib.leaky_relu(conv2)
conv3 = tlib.Con2D(relu2, 4 * FLAGS.DIM, 5, 2, scope="conv3")
relu3 = tlib.leaky_relu(conv3)
out_put = tf.reshape(relu3, [-1, 4 * 4 * 4 * FLAGS.DIM])
fc1 = tlib.fc(out_put, 1, scope="fc1")
return tf.reshape(fc1, [-1])
def Discriminator(input, reuse=False):
with tf.variable_scope("Discriminator") as scope:
if reuse:
scope.reuse_variables()
image = tf.reshape(input, [-1, 28, 28, 1])
conv1 = tlib.Con2D(image, FLAGS.DIM, 5, 2, scope="conv1")
relu1 = tlib.leaky_relu(conv1)
conv2 = tlib.Con2D(relu1, 2 * FLAGS.DIM, 5, 2, scope="conv2")
relu2 = tlib.leaky_relu(conv2)
conv3 = tlib.Con2D(relu2, 4 * FLAGS.DIM, 5, 2, scope="conv3")
relu3 = tlib.leaky_relu(conv3)
out_put = tf.reshape(relu3, [-1, 4 * 4 * 4 * FLAGS.DIM])
fc1 = tlib.fc(out_put, 1, scope="fc1")
return fc1