def lenet5(images, nclass):
# conv 1
net = layers.conv_2d_layer(images, [5, 5, images.get_shape()[3], 20], nonlinearity=None,
padding='SAME', name='conv_1', with_biases=False)
biases = layers.variable_on_cpu('biases_1', net.get_shape()[3], tf.constant_initializer(0.0), dtype=tf.float32)
net = tf.nn.bias_add(net, biases)
net = tf.nn.relu(net)
# max pool
net = tf.nn.max_pool(net, ksize=[1,2,2,1], strides=[1,2,2,1], padding='VALID')
# conv 2
net = layers.conv_2d_layer(net, [5, 5, net.get_shape()[3], 50], nonlinearity=None,
padding='SAME', name='conv_2', with_biases=False)
biases = layers.variable_on_cpu('biases_2', net.get_shape()[3], tf.constant_initializer(0.0), dtype=tf.float32)
net = tf.nn.bias_add(net, biases)
net = tf.nn.relu(net)
# max pool
net = tf.nn.max_pool(net, ksize=[1,2,2,1], strides=[1,2,2,1], padding='VALID')
# reshape
net = tf.reshape(net, [-1, (net.get_shape()[1]*net.get_shape()[2]*net.get_shape()[3]).value])
# dense 1
net = layers.dense_layer(net, net.get_shape()[1], 500,
nonlinearity=None, name='dense_1')
net = tf.nn.relu(net)
# dense 2
net = layers.dense_layer(net, net.get_shape()[1], nclass,
nonlinearity=None, name='dense_2')
return net
def conv_bn_rectify(net, num_filters, name, is_training, reuse, kl_weight=1.0):
with tf.variable_scope(name):
net = layers.conv_2d_layer(net, [3,3, net.get_shape()[3], num_filters], nonlinearity=None,
padding='SAME', name='conv', with_biases=False, reuse=reuse)
net = layers.sbp_dropout(net, 3, is_training, 'sbp', reuse, kl_weight=kl_weight)
biases = layers.variable_on_cpu('biases', net.get_shape()[3], tf.constant_initializer(0.0), dtype=tf.float32)
net = tf.nn.bias_add(net, biases)
net = tf.contrib.layers.batch_norm(net, scope=tf.get_variable_scope(), reuse=reuse, is_training=False,
center=True, scale=True)
net = tf.nn.relu(net)
return net
def net_vgglike(images, nclass, scale, is_training, reuse):
net = conv_bn_rectify(images, int(64*scale), 'conv_1', is_training, reuse, 10.0)
net = conv_bn_rectify(net, int(64*scale), 'conv_2', is_training, reuse, 10.0)
net = tf.nn.max_pool(net, ksize=[1,2,2,1], strides=[1,2,2,1], padding='VALID')
net = conv_bn_rectify(net, int(128*scale), 'conv_3', is_training, reuse, 5.0)
net = conv_bn_rectify(net, int(128*scale), 'conv_4', is_training, reuse, 5.0)
net = tf.nn.max_pool(net, ksize=[1,2,2,1], strides=[1,2,2,1], padding='VALID')
net = conv_bn_rectify(net, int(256*scale), 'conv_5', is_training, reuse)
net = conv_bn_rectify(net, int(256*scale), 'conv_6', is_training, reuse)
net = conv_bn_rectify(net, int(256*scale), 'conv_7', is_training, reuse)
net = tf.nn.max_pool(net, ksize=[1,2,2,1], strides=[1,2,2,1], padding='VALID')
net = conv_bn_rectify(net, int(512*scale), 'conv_8', is_training, reuse)
net = conv_bn_rectify(net, int(512*scale), 'conv_9', is_training, reuse)
net = conv_bn_rectify(net, int(512*scale), 'conv_10', is_training, reuse)
net = tf.nn.max_pool(net, ksize=[1,2,2,1], strides=[1,2,2,1], padding='VALID')
net = conv_bn_rectify(net, int(512*scale), 'conv_11', is_training, reuse)
net = conv_bn_rectify(net, int(512*scale), 'conv_12', is_training, reuse)
net = conv_bn_rectify(net, int(512*scale), 'conv_13', is_training, reuse)
net = tf.nn.max_pool(net, ksize=[1,2,2,1], strides=[1,2,2,1], padding='VALID')
net = tf.reshape(net, [-1, (net.get_shape()[1]*net.get_shape()[2]*net.get_shape()[3]).value])
net = layers.sbp_dropout(net, 1, is_training, 'sbp_dense_1', reuse)
net = layers.dense_layer(net, net.get_shape()[1], int(512*scale),
nonlinearity=None, name='dense_1', with_biases=False)
biases = layers.variable_on_cpu('biases_dense_1', net.get_shape()[1], tf.constant_initializer(0.0),
dtype=tf.float32)
net = tf.nn.bias_add(net, biases)
net = tf.contrib.layers.batch_norm(net, scope=tf.get_variable_scope(), reuse=reuse, is_training=False,
center=True, scale=True)
net = tf.nn.relu(net)
net = layers.sbp_dropout(net, 1, is_training, 'sbp_dense_2', reuse)
net = layers.dense_layer(net, net.get_shape()[1], nclass,
nonlinearity=None, name='dense_2', with_biases=False)
biases = layers.variable_on_cpu('biases_dense_2', net.get_shape()[1], tf.constant_initializer(0.0),
dtype=tf.float32)
net = tf.nn.bias_add(net, biases)
return net
def net_vgglike(images, nclass, num_filters, is_training, reuse):
net = conv_bn_rectify(images, num_filters[0], 'conv_1', is_training, reuse)
net = tf.contrib.layers.dropout(net,
keep_prob=0.7,
is_training=is_training)
net = conv_bn_rectify(net, num_filters[1], 'conv_2', is_training, reuse)
net = tf.nn.max_pool(net,
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='VALID')
net = conv_bn_rectify(net, num_filters[2], 'conv_3', is_training, reuse)
net = tf.contrib.layers.dropout(net,
keep_prob=0.6,
is_training=is_training)
net = conv_bn_rectify(net, num_filters[3], 'conv_4', is_training, reuse)
net = tf.nn.max_pool(net,
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='VALID')
net = conv_bn_rectify(net, num_filters[4], 'conv_5', is_training, reuse)
net = tf.contrib.layers.dropout(net,
keep_prob=0.6,
is_training=is_training)
net = conv_bn_rectify(net, num_filters[5], 'conv_6', is_training, reuse)
net = tf.contrib.layers.dropout(net,
keep_prob=0.6,
is_training=is_training)
net = conv_bn_rectify(net, num_filters[6], 'conv_7', is_training, reuse)
net = tf.nn.max_pool(net,
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='VALID')
net = conv_bn_rectify(net, num_filters[7], 'conv_8', is_training, reuse)
net = tf.contrib.layers.dropout(net,
keep_prob=0.6,
is_training=is_training)
net = conv_bn_rectify(net, num_filters[8], 'conv_9', is_training, reuse)
net = tf.contrib.layers.dropout(net,
keep_prob=0.6,
is_training=is_training)
net = conv_bn_rectify(net, num_filters[9], 'conv_10', is_training, reuse)
net = tf.nn.max_pool(net,
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='VALID')
net = conv_bn_rectify(net, num_filters[10], 'conv_11', is_training, reuse)
net = conv_bn_rectify(net, num_filters[11], 'conv_12', is_training, reuse)
net = conv_bn_rectify(net, num_filters[12], 'conv_13', is_training, reuse)
net = tf.nn.max_pool(net,
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='VALID')
net = tf.reshape(net, [
-1,
(net.get_shape()[1] * net.get_shape()[2] * net.get_shape()[3]).value
])
net = tf.contrib.layers.dropout(net,
keep_prob=0.5,
is_training=is_training)
net = layers.dense_layer(net,
net.get_shape()[1],
512,
nonlinearity=None,
name='dense_1',
with_biases=False)
biases = layers.variable_on_cpu('biases_dense_1',
net.get_shape()[1],
tf.constant_initializer(0.0),
dtype=tf.float32)
net = tf.nn.bias_add(net, biases)
net = tf.contrib.layers.batch_norm(net,
scope=tf.get_variable_scope(),
reuse=reuse,
is_training=is_training,
center=True,
scale=True)
net = tf.nn.relu(net)
net = tf.contrib.layers.dropout(net,
keep_prob=0.5,
is_training=is_training)
net = layers.dense_layer(net,
net.get_shape()[1],
nclass,
nonlinearity=None,
name='dense_2',
with_biases=False)
biases = layers.variable_on_cpu('biases_dense_2',
net.get_shape()[1],
tf.constant_initializer(0.0),
dtype=tf.float32)
net = tf.nn.bias_add(net, biases)
tf.add_to_collection('logits', net)
return net