def build(self):
self.X3 = tf.reshape(self.X3, shape=[-1, 28, 28, 1])
conv1 = conv2d(self.X3, self.weights['wc1'], self.biases['bc1'])
conv2 = conv2d(conv1, self.weights['wc2'], self.biases['bc2'])
conv2 = max_pool_2x2(conv2)
conv2 = dropout_layer(conv2, self.keep_prob3)
conv2 = max_pool_2x2(conv2)
conv2 = dropout_layer(conv2, self.keep_prob3)
# Flatten
conv3 = tf.layers.flatten(conv2)
# Dense
fc1 = tf.add(tf.matmul(conv3, self.weights['wd1']), self.biases['bd1'])
fc1 = tf.nn.relu(fc1)
# Dropout
fc1 = dropout_layer(fc1, self.keep_prob3)
# Dense
out = tf.add(tf.matmul(fc1, self.weights['out']), self.biases['out'])
predictions = tf.nn.softmax(out)
return out, predictions
def model3(x, weights, biases, dropout):
# # Make the tensorflow graph here
with tf.name_scope('reshape'):
x = tf.reshape(x, shape=[-1, 28, 28, 1])
with tf.name_scope('conv1'):
conv1 = conv2d(x, weights['wc1'], biases['bc1'])
conv1 = max_pool_2x2(conv1)
with tf.name_scope('conv2'):
conv2 = conv2d(conv1, weights['wc2'], biases['bc2'])
conv2 = max_pool_2x2(conv2)
with tf.name_scope('conv3'):
conv3 = conv2d(conv2, weights['wc3'], biases['bc3'])
conv3 = max_pool_2x2(conv3)
conv3 = tf.layers.flatten(conv3)
with tf.name_scope('fc1'):
fc1 = tf.add(tf.matmul(conv3, weights['wd1']), biases['bd1'])
fc1 = tf.nn.relu(fc1)
out = tf.add(tf.matmul(fc1, weights['out']), biases['out'], name="out")
return out
def discriminator(gan, inp, num, keep_prob, reuse=False):
gpu_num = 0
hidden_units = gan.h_adv
print(hidden_units)
print(keep_prob)
with tf.device('/gpu:%d' % gpu_num):
with tf.variable_scope('discriminator_%d' % num):
if reuse:
tf.get_variable_scope().reuse_variables()
with tf.variable_scope('conv0'):
h0 = conv2d(inp, [3, 3, gan.num_channels, hidden_units / 4], [hidden_units / 4],
stride=2, name='h0')
h0 = leaky_relu(0.2, h0)
h0 = tf.nn.dropout(h0, keep_prob)
with tf.variable_scope('conv1'):
h1 = conv2d(h0, [3, 3, hidden_units / 4, hidden_units / 2], [hidden_units / 2],
stride=2, name='h0')
h1 = leaky_relu(0.2, h1)
h1 = tf.nn.dropout(h1, keep_prob)
with tf.variable_scope('conv2'):
h2 = conv2d(h1, [3, 3, hidden_units / 2, hidden_units], [hidden_units],
stride=1, name='h0')
h2 = leaky_relu(0.2, h2)
with tf.variable_scope('reshape'):
shape = h2.get_shape().as_list()
num_units = shape[1] * shape[2] * shape[3]
flattened = tf.reshape(h2, [gan.batch_size, num_units])
with tf.variable_scope('prediction'):
pred = dense(flattened, [num_units, 1], [1])
return pred
def build(self):
self.X = tf.reshape(self.X, shape=[-1, 28, 28, 1])
conv1 = conv2d(self.X, self.weights["wc1"], self.biases['bc1'])
conv1 = max_pool_2x2(conv1, ksize=2)
conv2 = conv2d(conv1, self.weights['wc2'], self.biases['bc2'])
conv2 = max_pool_2x2(conv2, ksize=2)
fc1 = tf.reshape(conv2, [-1, self.weights['wd1'].get_shape().as_list()[0]])
fc1 = tf.add(tf.matmul(fc1, self.weights['wd1']), self.biases['bd1'])
fc1 = tf.nn.relu(fc1)
fc1 = tf.nn.dropout(fc1, self.keep_prob)
out = tf.add(tf.matmul(fc1, self.weights['out']), self.biases['out'])
predictions = tf.nn.softmax(out)
return out, predictions
def discriminator(gan, image, reuse=False, name='Discriminator'):
"""
Args:
gan : instance of a generative adversarial network
reuse : Whether you want to reuse variables from previous
share_params : Whether weights are tied in initial layers
gan.batch_size: The size of batch. Should be specified before training. [64]
gan.output_size: The resolution in pixels of the images. [64]
gan.df_dim: Dimension of gen filters in first conv layer. [64]
gan.dfc_dim: Dimension of gen units for for fully connected layer. [1024]
gan.c_dim: Dimension of image color. For grayscale input, set to 1. [3]
"""
# layers that don't share variable
d_bn1 = batch_norm(name='d_bn1')
d_bn2 = batch_norm(name='d_bn2')
with tf.variable_scope(name):
if reuse:
tf.get_variable_scope().reuse_variables()
h0 = prelu(conv2d(image, gan.c_dim, name='d_h0_conv', reuse=False),
name='d_h0_prelu',
reuse=False)
h1 = prelu(d_bn1(conv2d(h0, gan.df_dim, name='d_h1_conv', reuse=False),
reuse=reuse),
name='d_h1_prelu',
reuse=False)
h1 = tf.reshape(h1, [self.batch_size, -1])
# layers that share variables
h2 = prelu(d_bn2(linear(h1, gan.dfc_dim, 'd_h2_lin', reuse=False),
reuse=False),
name='d_h2_prelu',
reuse=False)
h3 = linear(h2, 1, 'd_h3_lin', reuse=False)
return tf.nn.sigmoid(h3), h3
def build(self):
self.X2 = tf.reshape(self.X2, shape=[-1, 28, 28, 1])
conv1 = conv2d(self.X2, self.weights['wc1'], self.biases['bc1'])
conv1 = max_pool_2x2(conv1)
conv2 = conv2d(conv1, self.weights['wc2'], self.biases['bc2'])
conv2 = max_pool_2x2(conv2)
conv3 = conv2d(conv2, self.weights['wc3'], self.biases['bc3'])
conv3 = max_pool_2x2(conv3)
conv3 = tf.layers.flatten(conv3)
fc1 = tf.add(tf.matmul(conv3, self.weights['wd1']), self.biases['bd1'])
fc1 = tf.nn.relu(fc1)
out = tf.add(tf.matmul(fc1, self.weights['out']), self.biases['out'])
predictions = tf.nn.softmax(out)
return out, predictions
def model1(x, weights, biases, dropout):
with tf.name_scope('reshape'):
x = tf.reshape(x, shape=[-1, 28, 28, 1])
with tf.name_scope('conv1'):
conv1 = conv2d(x, weights["wc1"], biases['bc1'])
conv1 = max_pool_2x2(conv1, ksize=2)
with tf.name_scope('conv2'):
conv2 = conv2d(conv1, weights['wc2'], biases['bc2'])
conv2 = max_pool_2x2(conv2, ksize=2)
with tf.name_scope('fc1'):
fc1 = tf.reshape(conv2, [-1, weights['wd1'].get_shape().as_list()[0]])
fc1 = tf.add(tf.matmul(fc1, weights['wd1']), biases['bd1'])
fc1 = tf.nn.relu(fc1)
fc1 = tf.nn.dropout(fc1, DROPOUT)
out = tf.add(tf.matmul(fc1, weights['out']), biases['out'], name="out")
return out