def build_graph(input_pl,
is_training,
keep_prob,
weight_decay=0.0,
bn_decay=None,
reuse_layers=True):
print("\nNetowrk Input: ", input_pl)
net = tf.image.resize_images(input_pl, [cfg.im_dim, cfg.im_dim])
net = tf.layers.conv2d(
inputs=net,
filters=32,
kernel_size=[5, 5],
# strides=(1, 1),
# padding='valid',
# data_format='channels_last',
activation=tf.nn.relu,
use_bias=True,
kernel_initializer=tf.contrib.layers.xavier_initializer(),
bias_initializer=tf.zeros_initializer(),
# kernel_regularizer=l2_reg,
name='conv2d_layer_1',
reuse=reuse_layers)
# implement batch normalization
if bn_decay:
net = do_bn(net, is_training)
net = tf.layers.conv2d(
inputs=net,
filters=48,
kernel_size=[5, 5],
activation=tf.nn.relu,
use_bias=True,
kernel_initializer=tf.contrib.layers.xavier_initializer(),
bias_initializer=tf.zeros_initializer(),
name='conv2d_layer_2',
reuse=reuse_layers)
net = tf_ops.max_pool2d(net, [3, 3], padding='VALID', scope='maxpool')
if bn_decay:
net = do_bn(net, is_training)
net = tf.layers.conv2d(
inputs=net,
filters=64,
kernel_size=[3, 3],
activation=tf.nn.relu,
use_bias=True,
kernel_initializer=tf.contrib.layers.xavier_initializer(),
bias_initializer=tf.zeros_initializer(),
name='conv2d_layer_3',
reuse=reuse_layers)
if bn_decay:
net = do_bn(net, is_training)
net = tf.layers.conv2d(
inputs=net,
filters=64,
kernel_size=[3, 3],
activation=tf.nn.relu,
use_bias=True,
kernel_initializer=tf.contrib.layers.xavier_initializer(),
bias_initializer=tf.zeros_initializer(),
name='conv2d_layer_4',
reuse=reuse_layers)
net = tf_ops.max_pool2d(net, [3, 3], padding='VALID', scope='maxpool')
if bn_decay:
net = do_bn(net, is_training)
net = tf.layers.conv2d(
inputs=net,
filters=128,
kernel_size=[3, 3],
activation=tf.nn.relu,
use_bias=True,
kernel_initializer=tf.contrib.layers.xavier_initializer(),
bias_initializer=tf.zeros_initializer(),
name='conv2d_layer_5',
reuse=reuse_layers)
print("\nFinal Conv Shape: ", net.shape)
if bn_decay:
net = do_bn(net, is_training)
net = tf.contrib.layers.flatten(net)
print("\nshape after flatenning: ", net.shape)
net = tf.layers.dense(
inputs=net,
units=512,
activation=tf.nn.relu,
use_bias=True,
kernel_initializer=tf.truncated_normal_initializer(stddev=0.1),
bias_initializer=tf.zeros_initializer(),
kernel_regularizer=None,
bias_regularizer=None,
activity_regularizer=None,
kernel_constraint=None,
bias_constraint=None,
trainable=True,
name='fc_1',
reuse=reuse_layers)
if bn_decay:
net = do_bn(net, is_training)
net = tf_ops.dropout(net,
keep_prob=keep_prob,
is_training=is_training,
scope='dp1')
net = tf.layers.dense(
inputs=net,
units=128,
# units=256,
activation=tf.nn.relu,
use_bias=True,
kernel_initializer=tf.truncated_normal_initializer(stddev=0.1),
bias_initializer=tf.zeros_initializer(),
name='fc_2',
reuse=reuse_layers)
embed_logits = tf.nn.l2_normalize(net, axis=-1)
class_logits = tf.layers.dense(
inputs=net,
units=cfg.num_classes,
activation=tf.nn.softmax,
# activation=tf.nn.sigmoid,
use_bias=True,
kernel_initializer=tf.truncated_normal_initializer(stddev=0.1),
bias_initializer=tf.zeros_initializer(),
name='classification_output',
reuse=reuse_layers)
tf.summary.histogram('embed_outputs', embed_logits)
print("\nShape of logits: ", embed_logits.shape)
return embed_logits, class_logits
def build_graph_multi(input_pl,
is_training,
weight_decay=0.0,
keep_prob=1.0,
bn_decay=None):
print("\nNetowrk Input: ", input_pl)
with tf.device('/device:GPU:1'):
net = tf.image.resize_images(input_pl, [cfg.im_dim, cfg.im_dim])
net = tf_ops.conv2d(net,
32, [5, 5],
padding='VALID',
stride=[1, 1],
bn=True,
is_training=is_training,
scope='conv1',
bn_decay=bn_decay)
net = tf.nn.relu(net)
net = do_bn(net, is_training)
net = tf_ops.conv2d(net,
32, [5, 5],
padding='VALID',
stride=[1, 1],
bn=True,
is_training=is_training,
scope='conv1_1',
bn_decay=bn_decay)
net = tf.nn.relu(net)
net = do_bn(net, is_training)
net = tf_ops.conv2d(net,
48, [5, 5],
padding='VALID',
stride=[1, 1],
bn=True,
is_training=is_training,
scope='conv2',
bn_decay=bn_decay)
net = tf.nn.relu(net)
net = do_bn(net, is_training)
net = tf_ops.conv2d(net,
48, [3, 3],
padding='VALID',
stride=[1, 1],
bn=True,
is_training=is_training,
scope='conv2_2',
bn_decay=bn_decay)
net = do_bn(net, is_training)
net = tf_ops.max_pool2d(net, [3, 3], padding='VALID', scope='maxpool')
net = tf_ops.conv2d(net,
64, [3, 3],
padding='VALID',
stride=[1, 1],
bn=True,
is_training=is_training,
scope='conv3',
bn_decay=bn_decay)
net = tf.nn.relu(net)
net = do_bn(net, is_training)
net = tf_ops.conv2d(net,
64, [3, 3],
padding='VALID',
stride=[1, 1],
bn=True,
is_training=is_training,
scope='conv4',
bn_decay=bn_decay)
net = tf.nn.relu(net)
net = do_bn(net, is_training)
with tf.device('/device:GPU:2'):
net = tf_ops.max_pool2d(net, [3, 3], padding='VALID', scope='maxpool')
net = tf_ops.conv2d(net,
128, [3, 3],
padding='VALID',
stride=[1, 1],
bn=True,
is_training=is_training,
scope='conv5',
bn_decay=bn_decay)
net = tf.nn.relu(net)
net = do_bn(net, is_training)
net = tf_ops.conv2d(net,
128, [3, 3],
padding='VALID',
stride=[1, 1],
bn=True,
is_training=is_training,
scope='conv6',
bn_decay=bn_decay)
print("\nfinal conv shape: ", net.shape)
net = tf.nn.relu(net)
net = do_bn(net, is_training)
# Symmetric function: max pooling
# net = tf_ops.max_pool2d(net, [3,3], padding='VALID', scope='maxpool')
# net = tf.reshape(net, [cfg.batch_size, -1])
net = tf.contrib.layers.flatten(net)
print("\nshape after flatenning: ", net.shape)
net = tf_ops.fully_connected(net,
1024,
bn=True,
is_training=is_training,
scope='fc1',
bn_decay=bn_decay)
net = tf.nn.relu(net)
net = do_bn(net, is_training)
net = tf_ops.dropout(net,
keep_prob=keep_prob,
is_training=is_training,
scope='dp1')
net = tf_ops.fully_connected(net,
512,
bn=True,
is_training=is_training,
scope='fc2',
bn_decay=bn_decay)
net = tf.nn.relu(net)
net = do_bn(net, is_training)
net = tf_ops.dropout(net,
keep_prob=keep_prob,
is_training=is_training,
scope='dp2')
net = tf_ops.fully_connected(net, cfg.num_classes, scope='fc3')
# net = tf.nn.sigmoid(net, name="output_node")
net = tf.nn.softmax(net, name="output_node")
print("\nShape of logits: ", net.shape)
return net
def build_graph_old(input_pl,
is_training,
weight_decay=0.0,
keep_prob=1.0,
bn_decay=None):
print("\nNetowrk Input: ", input_pl)
pool_num = int(cfg.num_points / 3)
net = tf_ops.conv2d(input_pl,
64, [1, 3],
padding='VALID',
stride=[1, 1],
bn=True,
is_training=is_training,
scope='conv1',
bn_decay=bn_decay)
print("\nshape after input: ", net.shape)
net = tf.nn.relu(net)
# skip_pool = tf_ops.max_pool2d(net, [pool_num, 1],
# padding='VALID', scope='maxpool')
net = tf_ops.conv2d(net,
64, [1, 1],
padding='VALID',
stride=[1, 1],
bn=True,
is_training=is_training,
scope='conv2',
bn_decay=bn_decay)
net = tf.nn.relu(net)
net = tf_ops.conv2d(net,
64, [1, 1],
padding='VALID',
stride=[1, 1],
bn=True,
is_training=is_training,
scope='conv3',
bn_decay=bn_decay)
net = tf.nn.relu(net)
net = tf_ops.conv2d(net,
128, [1, 1],
padding='VALID',
stride=[1, 1],
bn=True,
is_training=is_training,
scope='conv4',
bn_decay=bn_decay)
net = tf.nn.relu(net)
net = tf_ops.conv2d(net,
1024, [1, 1],
padding='VALID',
stride=[1, 1],
bn=True,
is_training=is_training,
scope='conv5',
bn_decay=bn_decay)
net = tf.nn.relu(net)
print("\nshape before max pool: ", net.shape)
# Symmetric function: max pooling
# net = tf_ops.max_pool2d(net, [pool_num, 1],
net = tf_ops.max_pool2d(net, [cfg.num_points, 1],
padding='VALID',
scope='maxpool')
# skip_multiply = tf.multiply(net, skip_pool)
# net = tf.add(net, skip_multiply)
print("\nshape after skip pool: ", net.shape)
net = tf.contrib.layers.flatten(net)
print("\nshape after flatenning: ", net.shape)
net = tf_ops.fully_connected(net,
512,
bn=True,
is_training=is_training,
scope='fc1',
bn_decay=bn_decay)
net = tf.nn.relu(net)
net = tf_ops.dropout(net,
keep_prob=keep_prob,
is_training=is_training,
scope='dp1')
net = tf_ops.fully_connected(net,
128,
bn=True,
is_training=is_training,
scope='fc2',
bn_decay=bn_decay)
net = tf.nn.relu(net)
net = tf_ops.dropout(net,
keep_prob=keep_prob,
is_training=is_training,
scope='dp2')
net = tf_ops.fully_connected(net, cfg.num_classes, scope='fc3')
net = tf.nn.relu(net, name="output_node")
print("\nShape of logits: ", net.shape)
return net
def build_graph(input_pl,
is_training,
weight_decay=0.0,
keep_prob=1.0,
bn_decay=None):
print("\nNetowrk Input: ", input_pl)
net = tf.image.resize_images(input_pl, [cfg.im_dim, cfg.im_dim])
net = tf_ops.conv2d(net,
32, [5, 5],
padding='VALID',
stride=[1, 1],
bn=True,
is_training=is_training,
scope='conv1',
bn_decay=bn_decay)
net = tf.nn.relu(net)
net = do_bn(net, is_training)
net = tf_ops.conv2d(net,
48, [5, 5],
padding='VALID',
stride=[1, 1],
bn=True,
is_training=is_training,
scope='conv1_1',
bn_decay=bn_decay)
net = tf.nn.relu(net)
net = tf_ops.max_pool2d(net, [3, 3], padding='VALID', scope='maxpool_1')
net = inception2d(net, 64, is_training, bn_decay, name='incp_1')
net = do_bn(net, is_training)
net = inception2d(net, 64, is_training, bn_decay, name='incp_2')
net = do_bn(net, is_training)
net = inception2d(net, 64, is_training, bn_decay, name='incp_3')
net = do_bn(net, is_training)
net = tf_ops.max_pool2d(net, [3, 3], padding='VALID', scope='maxpool_2')
net = inception2d(net, 96, is_training, bn_decay, name='incp_4')
net = do_bn(net, is_training)
net = inception2d(net, 96, is_training, bn_decay, name='incp_5')
net = do_bn(net, is_training)
net = inception2d(net, 128, is_training, bn_decay, name='incp_6')
net = do_bn(net, is_training)
net = tf_ops.max_pool2d(net, [3, 3], padding='VALID', scope='maxpool_3')
print("\nfinal conv shape: ", net.shape)
net = tf.contrib.layers.flatten(net)
print("\nshape after flatenning: ", net.shape)
net = tf_ops.fully_connected(net,
1024,
bn=True,
is_training=is_training,
scope='fc1',
bn_decay=bn_decay)
net = tf.nn.relu(net)
net = do_bn(net, is_training)
net = tf_ops.dropout(net,
keep_prob=keep_prob,
is_training=is_training,
scope='dp1')
net = tf_ops.fully_connected(net,
512,
bn=True,
is_training=is_training,
scope='fc2',
bn_decay=bn_decay)
net = tf.nn.relu(net)
net = do_bn(net, is_training)
net = tf_ops.dropout(net,
keep_prob=keep_prob,
is_training=is_training,
scope='dp2')
net = tf_ops.fully_connected(net, cfg.num_classes, scope='fc3')
# net = tf.nn.sigmoid(net, name="output_node")
net = tf.nn.softmax(net, name="output_node")
print("\nShape of logits: ", net.shape)
return net