def discriminator_base(inputs):
with tf.name_scope('discriminator_base'):
#net = layers.batch_norm(inputs, training, name='bn1')
net = layers.conv2d_layer(1,
inputs, [5, 5, 16],
lambda x: layers.lrelu(x, 0.2),
stride=2)
#net = layers.batch_norm(net, training, name='bn2')
net = layers.conv2d_layer(2,
net, [5, 5, 32],
lambda x: layers.lrelu(x, 0.2),
stride=2)
#net = layers.batch_norm(net, training, name='bn3')
net = layers.conv2d_layer(3,
net, [5, 5, 64],
lambda x: layers.lrelu(x, 0.2),
stride=2)
#net = layers.batch_norm(net, training, name='bn4')
net = layers.conv2d_layer(4,
net, [5, 5, 128],
lambda x: layers.lrelu(x, 0.2),
stride=2)
net = layers.max_pool2d(net, [2, 2])
#net = layers.batch_norm(net, training, name='bn5')
return net
def __init__(self):
conv_stride = [4, 4]
self.conv1 = layers.conv2d(filters=96, kernel=[11, 11], padding='SAME', name='conv1', activation='relu', normalization='local_response_normalization', stride=conv_stride)
self.conv2 = layers.conv2d(filters=256, kernel=[5, 5], padding='SAME', name='conv2', activation='relu', normalization="local_response_normalization", stride=[1, 1])
self.conv1 = layers.conv2d(filters=96, kernel=[11, 11], padding='SAME', name='conv1', activation='relu', normalization='local_response_normalization', stride=conv_stride)
self.conv2 = layers.conv2d(filters=256, kernel=[5, 5], padding='SAME', name='conv2', activation='relu', normalization="local_response_normalization", stride=[1, 1])
self.conv3 = layers.conv2d(filters=384, kernel=[3, 3], padding='SAME', name='conv3', activation='relu', stride=[1, 1])
self.conv4 = layers.conv2d(filters=384, kernel=[3, 3], padding='SAME', name='conv4', activation='relu', stride=[1, 1])
self.conv5 = layers.conv2d(filters=256, kernel=[3, 3], padding='SAME', name='conv5', activation='relu', stride=[1, 1])
self.fc6 = layers.dense(4096, activation='relu', dropout=True, name='fc6')
self.fc7 = layers.dense(4096, activation='relu', dropout=True, name='fc7')
self.fc8 = layers.dense(1000, activation='relu', name='fc8')
self.max_pool1 = layers.max_pool2d(ksize=[3, 3], stride=[2, 2])
self.max_pool2 = layers.max_pool2d(ksize=[3, 3], stride=[2, 2])
self.max_pool5 = layers.max_pool2d(ksize=[3, 3], stride=[2, 2])
def __call__(self, x,
is_training=True,
reuse=False,
*args,
**kwargs):
with tf.variable_scope(self.__class__.__name__) as vs:
if reuse:
vs.reuse_variables()
conv_params = {'is_training': is_training,
'activation_': 'relu',
'normalization': 'batch'}
x = conv_block(x, 64, **conv_params, dropout_rate=0.3)
x = conv_block(x, 64, **conv_params, dropout_rate=0.3)
x = conv_block(x, 128, **conv_params, dropout_rate=0.4)
x = conv_block(x, 128, **conv_params, dropout_rate=0.4)
x = conv_block(x, 256, **conv_params, dropout_rate=0.4)
x = conv_block(x, 256, **conv_params, dropout_rate=0.4)
x = conv_block(x, 256, **conv_params)
l1 = x
x = max_pool2d(x)
x = conv_block(x, 512, **conv_params, dropout_rate=0.4)
x = conv_block(x, 512, **conv_params, dropout_rate=0.4)
x = conv_block(x, 512, **conv_params)
l2 = x
x = max_pool2d(x)
x = conv_block(x, 512, **conv_params, dropout_rate=0.4)
x = conv_block(x, 512, **conv_params, dropout_rate=0.4)
x = conv_block(x, 512, **conv_params)
l3 = x
x = max_pool2d(x)
x = conv_block(x, 512, **conv_params, sampling='pool')
x = conv_block(x, 512, **conv_params, sampling='pool')
x = flatten(x)
g = dense(x, 512, activation_='relu')
x, attentions = attention_module([l1, l2, l3], g)
x = dense(x, self.nb_classes)
return x, attentions
def _siamese(self, img, scope, reuse=False):
with tf.variable_scope(scope, reuse=reuse):
x = img
h1 = conv2d(x,
64,
kernel_size=10,
strides=1,
name='conv1',
l2reg=1e-4,
training=self.is_training,
use_bn=True)
h1 = max_pool2d(h1, kernel_size=2, strides=2, name='pool1')
h1 = tf.layers.dropout(h1, rate=.4, training=self.is_training)
h2 = conv2d(h1,
128,
kernel_size=7,
strides=2,
name='conv2',
l2reg=1e-5,
training=self.is_training,
use_bn=True)
#h2 = max_pool2d(h2, kernel_size=2, strides=2, name='pool2')
h2 = tf.layers.dropout(h2, rate=.4, training=self.is_training)
h3 = conv2d(h2,
256,
kernel_size=4,
strides=2,
name='conv3',
l2reg=1e-5,
training=self.is_training,
use_bn=True)
#h3 = max_pool2d(h3, kernel_size=2, strides=2, name='pool3')
h3 = tf.layers.dropout(h3, rate=.4, training=self.is_training)
h4 = conv2d(h3,
512,
kernel_size=4,
strides=2,
name='conv4',
l2reg=1e-5,
training=self.is_training,
use_bn=True)
# h4 = max_pool2d(h4, kernel_size=2, strides=2, name='pool4')
h4 = tf.layers.dropout(h4, rate=.4, training=self.is_training)
h5 = conv2d(h4,
1024,
kernel_size=3,
strides=2,
name='conv5',
l2reg=1e-5,
training=self.is_training,
use_bn=True)
h5_flat = tf.layers.flatten(h5, name='flatten1')
out = fc(h5_flat,
2056,
name='out',
activation_fn=tf.nn.sigmoid,
l2reg=1e-6)
return out
padding='SAME',
name='conv4',
activation='relu',
stride=[1, 1])
conv5 = layers.conv2d(filters=256,
kernel=[3, 3],
padding='SAME',
name='conv5',
activation='relu',
stride=[1, 1])
fc6 = layers.dense(4096, activation='relu', dropout=True, name='fc6')
fc7 = layers.dense(4096, activation='relu', dropout=True, name='fc7')
fc8 = layers.dense(1000, activation='relu', name='fc8')
max_pool1 = layers.max_pool2d(ksize=[3, 3], stride=[2, 2])
max_pool2 = layers.max_pool2d(ksize=[3, 3], stride=[2, 2])
max_pool5 = layers.max_pool2d(ksize=[3, 3], stride=[2, 2])
def Forward(x):
out = conv1.forward(x)
# print('layer 1:', out.shape)
out = max_pool1.forward(out)
# print('layer max 1:', out.shape)
out = conv2.forward(out)
# print('layer conv2:', out.shape)
out = max_pool2.forward(out)