def discriminate(self, net, reuse=None, training=True, with_fc=True):
"""Builds a discriminator network on top of inputs.
Args:
net: Input to the discriminator
reuse: Whether to reuse already defined variables
training: Whether in train or eval mode.
with_fc: Whether to include fully connected layers (used during unsupervised training)
Returns:
Resulting logits
"""
with tf.variable_scope('discriminator', reuse=reuse):
with slim.arg_scope(alexnet_argscope(activation=self.fc_activation, padding='SAME', training=training,
fix_bn=self.fix_bn)):
net = slim.conv2d(net, 96, kernel_size=[11, 11], stride=4, scope='conv_1', padding=self.pad,
normalizer_fn=None)
net = slim.max_pool2d(net, kernel_size=[3, 3], stride=2, scope='pool_1', padding=self.pad)
# net = tf.nn.lrn(net, depth_radius=2, alpha=0.00002, beta=0.75)
net = conv_group(net, 256, kernel_size=[5, 5], scope='conv_2')
net = slim.max_pool2d(net, kernel_size=[3, 3], stride=2, scope='pool_2', padding=self.pad)
# net = tf.nn.lrn(net, depth_radius=2, alpha=0.00002, beta=0.75)
net = slim.conv2d(net, 384, kernel_size=[3, 3], scope='conv_3')
net = conv_group(net, 384, kernel_size=[3, 3], scope='conv_4')
net = conv_group(net, 256, kernel_size=[3, 3], scope='conv_5')
if self.use_pool5:
net = slim.max_pool2d(net, kernel_size=[3, 3], stride=2, scope='pool_5', padding=self.pad)
encoded = net
drop_pred = None
if with_fc:
drop_pred = slim.conv2d(net, 1, kernel_size=[1, 1], activation_fn=None, normalizer_fn=None)
drop_pred = slim.flatten(drop_pred)
net = slim.flatten(net)
net = slim.fully_connected(net, 4096, scope='fc1')
net = slim.dropout(net, 0.5, is_training=training)
net = slim.fully_connected(net, 4096, scope='fc2')
net = slim.dropout(net, 0.5, is_training=training)
net = slim.fully_connected(net, 2, scope='fc3',
activation_fn=None,
normalizer_fn=None,
biases_initializer=tf.zeros_initializer())
return net, drop_pred, encoded
def discriminate(self, net, reuse=None, training=True):
with tf.variable_scope('discriminator', reuse=reuse):
with slim.arg_scope(
alexnet_argscope(activation=self.fc_activation,
padding='SAME',
training=training,
fix_bn=self.fix_bn)):
self.layers['input'] = net
net = slim.conv2d(net,
96,
kernel_size=[11, 11],
stride=4,
scope='conv_1',
padding=self.pad,
normalizer_fn=None)
self.layers['conv_1'] = net
net = slim.max_pool2d(net,
kernel_size=[3, 3],
stride=2,
scope='pool_1',
padding=self.pad)
net = tf.nn.lrn(net, depth_radius=2, alpha=0.00002, beta=0.75)
net = conv_group_nobn(net,
256,
kernel_size=[5, 5],
scope='conv_2')
self.layers['conv_2'] = net
net = slim.max_pool2d(net,
kernel_size=[3, 3],
stride=2,
scope='pool_2',
padding=self.pad)
net = tf.nn.lrn(net, depth_radius=2, alpha=0.00002, beta=0.75)
net = slim.conv2d(net,
384,
kernel_size=[3, 3],
scope='conv_3',
normalizer_fn=None)
self.layers['conv_3'] = net
net = conv_group_nobn(net,
384,
kernel_size=[3, 3],
scope='conv_4')
self.layers['conv_4'] = net
net = conv_group(net, 256, kernel_size=[3, 3], scope='conv_5')
self.layers['conv_5'] = net
if self.use_pool5:
net = slim.max_pool2d(net,
kernel_size=[3, 3],
stride=2,
scope='pool_5',
padding=self.pad)
encoded = net
drop_pred = slim.conv2d(net,
1,
kernel_size=[3, 3],
padding='SAME',
activation_fn=None,
normalizer_fn=None,
scope='drop_pred')
self.layers['drop_pred'] = net
drop_pred = slim.flatten(drop_pred)
enc_shape = net.get_shape().as_list()
net = slim.conv2d(net,
4096,
kernel_size=enc_shape[1:3],
padding='VALID',
scope='fc_1')
self.layers['fc_1'] = net
net = slim.dropout(net, 0.5, is_training=training)
net = slim.conv2d(net,
2,
kernel_size=[1, 1],
padding='VALID',
activation_fn=None,
normalizer_fn=None,
scope='fc_2')
self.layers['fc_2'] = net
net = slim.flatten(net)
return net, drop_pred, encoded