# -*- coding: utf-8 -*-
"""
@author: Qian Yue
"""
import tensorflow as tf
import numpy as np
import logging
import os
import shutil
from core.util import crop_to_shape
logging.basicConfig(level=logging.INFO, format='%(asctime)s %(message)s')
def create_ae_encoder(x, training, summaries=True, batch_size=32):
"""
Introduce the encoder part of the auto-encoder for anatomical constraint.
:param x: Input label tensor, expected dim [batch_size, height, width, n_class],
here roi = [240, 240]
:param training: Whether to return the output in training mode or in inference mode.
:param summaries: Flag if summaries should be created.
:returns: codes (compact representation of labels); the tensor shape of the last encoder convolution
"""
# dw_layers
with tf.variable_scope('encoder'):
in_node = x
# [batch_size, 240, 240, n_class]
with tf.variable_scope('dw_conv_layer1'):
conv1 = tf.layers.conv2d(in_node, filters=16, kernel_size=3, strides=2, padding='same',
use_bias=False, kernel_initializer=tf.contrib.layers.xavier_initializer(),
name='conv1')
bn1 = tf.layers.batch_normalization(conv1, training=training, name='bn1')
relu1 = tf.nn.relu(bn1, name='relu1')
conv2 = tf.layers.conv2d(relu1, filters=16, kernel_size=3, use_bias=False, padding='same',
kernel_initializer=tf.contrib.layers.xavier_initializer(),
name='conv2')
bn2 = tf.layers.batch_normalization(conv2, training=training, name='bn2')
relu2 = tf.nn.relu(bn2, name='relu2')
in_node = relu2
# [batch_size, 120, 120, 16]
with tf.variable_scope('dw_conv_layer2'):
conv1 = tf.layers.conv2d(in_node, filters=32, kernel_size=3, strides=2, padding='same',
use_bias=False, kernel_initializer=tf.contrib.layers.xavier_initializer(),
name='conv1')
bn1 = tf.layers.batch_normalization(conv1, training=training, name='bn1')
relu1 = tf.nn.relu(bn1, name='relu1')
conv2 = tf.layers.conv2d(relu1, filters=32, kernel_size=3, use_bias=False, padding='same',
kernel_initializer=tf.contrib.layers.xavier_initializer(),
name='conv2')
bn2 = tf.layers.batch_normalization(conv2, training=training, name='bn2')
relu2 = tf.nn.relu(bn2, name='relu2')
in_node = relu2
# [batch_size, 60, 60, 32]
with tf.variable_scope('dw_conv_layer3'):
conv1 = tf.layers.conv2d(in_node, filters=64, kernel_size=3, strides=2, padding='same',
use_bias=False, kernel_initializer=tf.contrib.layers.xavier_initializer(),
name='conv1')
bn1 = tf.layers.batch_normalization(conv1, training=training, name='bn1')
relu1 = tf.nn.relu(bn1, name='relu1')
conv2 = tf.layers.conv2d(relu1, filters=64, kernel_size=3, use_bias=False, padding='same',
kernel_initializer=tf.contrib.layers.xavier_initializer(),
name='conv2')
bn2 = tf.layers.batch_normalization(conv2, training=training, name='bn2')
relu2 = tf.nn.relu(bn2, name='relu2')
in_node = relu2
# [batch_size, 30, 30, 64]
with tf.variable_scope('dw_conv_layer4'):
conv = tf.layers.conv2d(in_node, filters=1, kernel_size=3, strides=(3, 3), padding='same',
use_bias=False, kernel_initializer=tf.contrib.layers.xavier_initializer(),
name='conv')
bn = tf.layers.batch_normalization(conv, training=training, name='bn')
relu = tf.nn.relu(bn, name='relu')
in_node = relu
# [batch_size, 10, 10, 1]
with tf.variable_scope('fc_layer'):
in_node = tf.reshape(in_node, [batch_size, 100])
codes = tf.layers.dense(in_node, units=64,
kernel_initializer=tf.contrib.layers.xavier_initializer(),
name='codes')
# [batch_size, 64]
if summaries:
with tf.name_scope('summaries'):
tf.summary.histogram('codes', codes)
return codes
def create_ae_decoder(x, training, n_class):
"""
Construct the decoder part of the auto-encoder for anatomical constraint.
:param x: Input codes tensor, expected a vector of length of codes. [batch_size, 64]
:param training: Whether to return the output in training mode or in inference mode.
:param n_class: The number of output classes.
:return: Decodes representing the reconstructed label maps.
"""
with tf.variable_scope('decoder'):
in_node = x
with tf.variable_scope('fc_layer'):
fc = tf.layers.dense(in_node, units=100, activation=tf.nn.relu,
kernel_initializer=tf.contrib.layers.xavier_initializer(),
name='fc')
in_node = tf.reshape(fc, [-1, 10, 10, 1])
# [batch_size, 10, 10, 1]
with tf.variable_scope('up_conv_layer4'):
deconv = tf.layers.conv2d_transpose(in_node, filters=64, kernel_size=7,
strides=(3, 3), padding='same', use_bias=False,
kernel_initializer=tf.contrib.layers.xavier_initializer(),
name='deconv')
bn1 = tf.layers.batch_normalization(deconv, training=training, name='bn1')
relu1 = tf.nn.relu(bn1, name='relu1')
conv = tf.layers.conv2d(relu1, filters=64, kernel_size=3, padding='same',
use_bias=False, kernel_initializer=tf.contrib.layers.xavier_initializer(),
name='conv')
bn2 = tf.layers.batch_normalization(conv, training=training, name='bn2')
relu2 = tf.nn.relu(bn2, name='relu2')
in_node = relu2
# [batch_size, 30, 30, 64]
with tf.variable_scope('up_conv_layer3'):
deconv = tf.layers.conv2d_transpose(in_node, filters=32, kernel_size=4,
strides=2, padding='same', use_bias=False,
kernel_initializer=tf.contrib.layers.xavier_initializer(),
name='deconv')
bn1 = tf.layers.batch_normalization(deconv, training=training, name='bn1')
relu1 = tf.nn.relu(bn1, name='relu1')
conv = tf.layers.conv2d(relu1, filters=32, kernel_size=3, padding='same',
use_bias=False, kernel_initializer=tf.contrib.layers.xavier_initializer(),
name='conv')
bn2 = tf.layers.batch_normalization(conv, training=training, name='bn2')
relu2 = tf.nn.relu(bn2, name='relu2')
in_node = relu2
# [batch_size, 60, 60, 32]
with tf.variable_scope('up_conv_layer2'):
deconv = tf.layers.conv2d_transpose(in_node, filters=16, kernel_size=4,
strides=2, padding='same', use_bias=False,
kernel_initializer=tf.contrib.layers.xavier_initializer(),
name='deconv')
bn1 = tf.layers.batch_normalization(deconv, training=training, name='bn1')
relu1 = tf.nn.relu(bn1, name='relu1')
conv = tf.layers.conv2d(relu1, filters=16, kernel_size=3, padding='same',
use_bias=False, kernel_initializer=tf.contrib.layers.xavier_initializer(),
name='conv')
bn2 = tf.layers.batch_normalization(conv, training=training, name='bn2')
relu2 = tf.nn.relu(bn2, name='relu2')
in_node = relu2
# [batch_size, 120, 120, 16]
with tf.variable_scope('up_conv_layer1'):
deconv = tf.layers.conv2d_transpose(in_node, filters=16, kernel_size=4,
strides=2, padding='same', use_bias=False,
kernel_initializer=tf.contrib.layers.xavier_initializer(),
name='deconv')
bn1 = tf.layers.batch_normalization(deconv, training=training, name='bn1')
relu1 = tf.nn.relu(bn1, name='relu1')
conv = tf.layers.conv2d(relu1, filters=n_class, kernel_size=3, padding='same',
kernel_initializer=tf.contrib.layers.xavier_initializer(),
name='conv')
# [batch_size, 240, 240, 4]
return conv
class AutoEncoder(object):
"""
An anatomical constraint auto-encoder implementation.
:param channels: (optional) number of channels in the input image
:param n_class: (optional) number of output labels
:param batch_size: size of training batch
"""
def __init__(self, channels=1, n_class=4, batch_size=1, optimizer="adam", cost_kwargs=None, opt_kwargs=None):
tf.reset_default_graph()
if cost_kwargs is None:
cost_kwargs = {}
if opt_kwargs is None:
opt_kwargs = {}
self.channels = channels
self.n_class = n_class
self.batch_size = batch_size
self.cost_kwargs = cost_kwargs
self.optimizer = optimizer
self.opt_kwargs = opt_kwargs
self.__labels = tf.placeholder("float", shape=[self.batch_size, 240, 240, n_class], name='labels')
self.__train_phase = tf.placeholder(tf.bool, name='train_phase')
with tf.variable_scope('autoencoder'):
self.__codes = create_ae_encoder(self.__labels, self.__train_phase, batch_size=self.batch_size)
# with shape [batch_size, 64]
self.__decodes = create_ae_decoder(self.__codes, self.__train_phase, self.n_class)
self.predictor = self._get_predictor(self.__decodes)
with tf.name_scope('cost_function'):
self.__cost = self._get_cost(self.__decodes, self.__labels)
def _get_predictor(self, logits):
"""
produce the probability maps from the final feature maps of the network
"""
return tf.nn.softmax(logits, axis=-1, name='probability_map')
def _get_cost(self, logits, labels):
"""
Construct the loss function of the auto-encoder.
"""
loss_map = tf.nn.softmax_cross_entropy_with_logits_v2(logits=logits, labels=labels,
name='cross_entropy_map')
cross_entropy = tf.reduce_mean(loss_map, name='cross_entropy')
return cross_entropy
def save(self, sess, model_path, latest_filename, **kwargs):
"""
Saves the current session to a checkpoint
:param sess: current session
:param model_path: path to file system location
:param latest_filename: Optional name for the protocol buffer file that will contains the list of most recent
checkpoints.
"""
saver = tf.train.Saver(**kwargs)
save_path = saver.save(sess, model_path, latest_filename=latest_filename)
return save_path
def restore(self, sess, model_path, **kwargs):
"""
Restores a session from a checkpoint
:param sess: current session instance
:param model_path: path to file system checkpoint location
"""
saver = tf.train.Saver(**kwargs)
saver.restore(sess, model_path)
logging.info("Model restored from file: %s" % model_path)
def _get_optimizer(self, training_iters, global_step):
"""
Construct optimizer based on the type of optimization
:param training_iters: decay step used for learning rate in momentum optimization
:param global_step: total frequency of optimization operation
:return train_op: optimization operation
"""
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
var_list = tf.trainable_variables(scope='autoencoder')
if self.optimizer == "momentum":
learning_rate = self.opt_kwargs.pop("learning_rate", 0.2)
decay_rate = self.opt_kwargs.pop("decay_rate", 0.95)
momentum = self.opt_kwargs.pop("momentum", 0.2)
self.__learning_rate_node = tf.train.exponential_decay(learning_rate=learning_rate,
global_step=global_step,
decay_steps=training_iters,
decay_rate=decay_rate,
staircase=True, name='learning_rate')
with tf.control_dependencies(update_ops):
train_op = tf.train.MomentumOptimizer(learning_rate=self.__learning_rate_node, momentum=momentum,
**self.opt_kwargs).minimize(self.__cost,
global_step=global_step,
var_list=var_list)
elif self.optimizer == "adam":
learning_rate = self.opt_kwargs.pop("learning_rate", 0.001)
self.__learning_rate_node = tf.Variable(learning_rate, name="learning_rate")
with tf.control_dependencies(update_ops):
train_op = tf.train.AdamOptimizer(learning_rate=self.__learning_rate_node,
**self.opt_kwargs).minimize(self.__cost,
global_step=global_step,
var_list=var_list)
else:
raise ValueError("Unknown optimizer type: " % self.optimizer)
return train_op
def _initialize(self, training_iters, model_path, restore):
"""
initialize optimization operation and model variables;
create model saving direction and summary operation
:param training_iters: decay step used for learning rate in momentum optimization
:param model_path: path to file system location
:param restore: whether to restore a previously trained model
:return init: global variables initializer
"""
global_step = tf.Variable(0, dtype=tf.int32, trainable=False, name='global_step')
tf.summary.scalar('loss', self.__cost)
self.summary_op = tf.summary.merge_all()
self.__optimizer = self._get_optimizer(training_iters, global_step)
init = tf.global_variables_initializer()
abs_model_path = os.path.abspath(model_path)
if not restore:
print("Removing '{:}'".format(abs_model_path))
shutil.rmtree(abs_model_path, ignore_errors=True)
if not os.path.exists(model_path):
print("Allocating '{:}'".format(abs_model_path))
os.makedirs(abs_model_path)
return init
def train(self, data_provider, model_path, training_iters=10, epochs=10, display_step=1, restore=False):
"""
Lauches the training process
:param data_provider: callable returning training data
:param model_path: path where to store checkpoints
:param training_iters: number of training mini batch iteration
:param epochs: number of epochs
:param display_step: number of steps till outputting stats
:param restore: Flag if previous model should be restored
"""
save_path = os.path.join(model_path, 'model.ckpt')
if epochs == 0:
return save_path
init = self._initialize(training_iters, model_path, restore)
with tf.Session() as sess:
sess.run(init)
if restore:
ckpt = tf.train.get_checkpoint_state(model_path)
if ckpt and ckpt.model_checkpoint_path:
var_list = tf.global_variables(scope='autoencoder') + self.__optimizer.variables() + [
tf.train.get_global_step()]
self.restore(sess, ckpt.model_checkpoint_path, var_list=var_list)
summary_writer = tf.summary.FileWriter(model_path, graph=sess.graph)
logging.info("Start Optimization!")
for epoch in range(epochs):
total_loss = 0.
for step in range((epoch * training_iters), ((epoch + 1) * training_iters)):
_, batch_y, _, _ = data_provider(self.batch_size)
decodes = sess.run(self.__decodes, feed_dict={self.__labels: batch_y,
self.__train_phase: False})
_, batch_loss = sess.run((self.__optimizer, self.__cost),
feed_dict={self.__labels: crop_to_shape(batch_y, decodes.shape),
self.__train_phase: True})
if step % display_step == 0:
logging.info("Iteration {:}, Mini-batch loss= {:.4f}".format(step, batch_loss))
summary_str = sess.run(self.summary_op, feed_dict={self.__labels: batch_y,
self.__train_phase: True})
summary_writer.add_summary(summary_str, step)
summary_writer.flush()
total_loss += batch_loss
logging.info("Epoch {:}, Average mini-batch loss= {:.4f}".format(epoch, total_loss / training_iters))
save_path = self.save(sess, save_path, "checkpoint")
logging.info("Optimization Finished!")
return save_path
Datasets
Models
