def __calc_conv3(self):
self.conv3_1 = tf.nn.relu(conv(self.pool2, 3, 3, 256, 1, 1, name='conv3_1', data=self.weights_data,
retrain_layers=self.retrain_layers))
self.conv3_2 = tf.nn.relu(conv(self.conv3_1, 3, 3, 256, 1, 1, name='conv3_2', data=self.weights_data,
retrain_layers=self.retrain_layers))
self.conv3_3 = tf.nn.relu(conv(self.conv3_2, 3, 3, 256, 1, 1, name='conv3_3', data=self.weights_data,
retrain_layers=self.retrain_layers))
self.pool3 = max_pool(self.conv3_3, 2, 2, 2, 2, name='pool3')
def __calc_conv5(self):
self.conv5_1 = tf.nn.relu(conv(self.pool4, 3, 3, 512, 1, 1, name='conv5_1', data=self.weights_data,
retrain_layers=self.retrain_layers))
self.conv5_2 = tf.nn.relu(conv(self.conv5_1, 3, 3, 512, 1, 1, name='conv5_2', data=self.weights_data,
retrain_layers=self.retrain_layers))
self.conv5_3 = tf.nn.relu(conv(self.conv5_2, 3, 3, 512, 1, 1, name='conv5_3', data=self.weights_data,
retrain_layers=self.retrain_layers))
self.pool5 = max_pool(self.conv5_3, 2, 2, 2, 2, name='pool5')
def highway(x, size=None, activation=None,
num_layers=2, scope="highway", dropout=0.0, reuse=None):
with tf.variable_scope(scope, reuse):
if size is None:
size = x.shape.as_list()[-1]
else:
x = conv(x, size, name="input_projection", reuse=reuse)
for i in range(num_layers):
T = conv(x, size, bias=True, activation=tf.sigmoid,
name="gate_%d" % i, reuse=reuse)
H = conv(x, size, bias=True, activation=activation,
name="activation_%d" % i, reuse=reuse)
H = tf.nn.dropout(H, 1.0 - dropout)
x = H * T + x * (1.0 - T)
return x
def __identity_block(self, input_layer, name, data, num_filters):
with tf.name_scope('identity_block'):
x = conv(input_layer,
1,
1,
num_filters,
1,
1,
padding='SAME',
name='res' + name + '_branch2a',
data=self.weights_data,
retrain_layers=self.retrain_layers)
x = batch_norm_layer(x, data=data, name='bn' + name + '_branch2a')
x = tf.nn.relu(x)
x = conv(x,
3,
3,
num_filters,
1,
1,
padding='SAME',
name='res' + name + '_branch2b',
data=self.weights_data,
retrain_layers=self.retrain_layers)
x = batch_norm_layer(x, data=data, name='bn' + name + '_branch2b')
x = tf.nn.relu(x)
x = conv(x,
1,
1,
num_filters * 4,
1,
1,
padding='SAME',
name='res' + name + '_branch2c',
data=self.weights_data,
retrain_layers=self.retrain_layers)
x = batch_norm_layer(x, data=data, name='bn' + name + '_branch2c')
x = tf.add(x, input_layer)
return x
def __calc_conv3(self):
# 3rd Layer: Conv (w ReLu)
self.conv3 = tf.nn.relu(
conv(self.norm2,
3,
3,
384,
1,
1,
name='conv3',
data=self.weights_data,
retrain_layers=self.retrain_layers))
def __calc_conv4(self):
# 4th Layer: Conv (w ReLu) splitted into two groups
self.conv4 = tf.nn.relu(
conv(self.conv3,
3,
3,
384,
1,
1,
name='conv4',
groups=2,
data=self.weights_data,
retrain_layers=self.retrain_layers))
def __calc_conv1(self):
temp = conv(self.preprocessed_image,
7,
7,
64,
2,
2,
padding='VALID',
name='conv1',
data=self.weights_data,
retrain_layers=self.retrain_layers)
temp = batch_norm_layer(temp, 'bn_conv1', self.weights_data)
self.conv1 = tf.nn.relu(temp)
self.pool1 = max_pool(self.conv1,
3,
3,
2,
2,
padding='SAME',
name='pool1')
def __calc_conv5(self):
# 5th Layer: Conv (w ReLu) -> Pool splitted into two groups
self.conv5 = tf.nn.relu(
conv(self.conv4,
3,
3,
256,
1,
1,
name='conv5',
groups=2,
data=self.weights_data,
retrain_layers=self.retrain_layers))
self.pool5 = max_pool(self.conv5,
3,
3,
2,
2,
padding='VALID',
name='pool5')
def __calc_conv2(self):
# 2nd Layer: Conv (w ReLu) -> Pool -> Lrn with 2 groups
self.conv2 = tf.nn.relu(
conv(self.norm1,
5,
5,
256,
1,
1,
name='conv2',
groups=2,
data=self.weights_data,
retrain_layers=self.retrain_layers))
self.pool2 = max_pool(self.conv2,
3,
3,
2,
2,
padding='VALID',
name='pool2')
self.norm2 = lrn(self.pool2, 2, 2e-05, 0.75, name='norm2')
def __calc_conv1(self):
# 1st layer: conv (w relu) -> pool -> lrn
self.conv1 = tf.nn.relu(
conv(self.preprocessed_image,
11,
11,
96,
4,
4,
padding='VALID',
name='conv1',
data=self.weights_data,
retrain_layers=self.retrain_layers))
self.pool1 = max_pool(self.conv1,
3,
3,
2,
2,
padding='VALID',
name='pool1')
self.norm1 = lrn(self.pool1, 2, 2e-05, 0.75, name='norm1')
def __calc_conv2(self):
self.conv2_1 = tf.nn.relu(conv(self.pool1, 3, 3, 128, 1, 1, name='conv2_1', data=self.weights_data,
retrain_layers=self.retrain_layers))
self.conv2_2 = tf.nn.relu(conv(self.conv2_1, 3, 3, 128, 1, 1, name='conv2_2', data=self.weights_data,
retrain_layers=self.retrain_layers))
self.pool2 = max_pool(self.conv2_2, 2, 2, 2, 2, name='pool2')
def __calc_conv1(self):
self.conv1_1 = tf.nn.relu(conv(self.preprocessed_image, 3, 3, 64, 1, 1, name='conv1_1', data=self.weights_data,
retrain_layers=self.retrain_layers, padding='VALID'))
self.conv1_2 = tf.nn.relu(conv(self.conv1_1, 3, 3, 64, 1, 1, name='conv1_2', data=self.weights_data,
retrain_layers=self.retrain_layers, padding='VALID'))
self.pool1 = max_pool(self.conv1_2, 2, 2, 2, 2, name='pool1', padding='SAME')
def __conv_block(self,
input_layer,
name,
data,
num_filters,
stride_x=2,
stride_y=2):
with tf.name_scope('conv_block'):
x = conv(input_layer,
1,
1,
num_filters,
1,
1,
padding='SAME',
name='res' + name + '_branch2a',
data=self.weights_data,
retrain_layers=self.retrain_layers)
x = batch_norm_layer(x, data=data, name='bn' + name + '_branch2a')
x = tf.nn.relu(x)
x = conv(x,
3,
3,
num_filters,
stride_x,
stride_y,
padding='SAME',
name='res' + name + '_branch2b',
data=self.weights_data,
retrain_layers=self.retrain_layers)
x = batch_norm_layer(x, data=data, name='bn' + name + '_branch2b')
x = tf.nn.relu(x)
x = conv(x,
1,
1,
num_filters * 4,
1,
1,
padding='SAME',
name='res' + name + '_branch2c',
data=self.weights_data,
retrain_layers=self.retrain_layers)
x = batch_norm_layer(x, data=data, name='bn' + name + '_branch2c')
shortcut = conv(input_layer,
1,
1,
num_filters * 4,
stride_x,
stride_y,
padding='SAME',
name='res' + name + '_branch1',
data=self.weights_data,
retrain_layers=self.retrain_layers)
shortcut = batch_norm_layer(shortcut,
data=data,
name='bn' + name + '_branch1')
x = tf.add(x, shortcut)
return x
