def run_training():
# construct the graph
with tf.Graph().as_default():
# specify the training data file location
#trainfile = os.path.join(FLAGS.data_dir, TRAIN_FILE)
# read the images and labels
train_dataset, train_labels, img_size, img_size_labels = nn.inputs_ms(batch_size=FLAGS.batch_size, filename=TRAIN_FILE)
#images, labels = nn.inputs(batch_size=FLAGS.batch_size,
# num_epochs=FLAGS.num_epochs,
# filename=trainfile)
# run inference on the images
results = nn.inference_ms(train_dataset, img_size)
# calculate the loss from the results of inference and the labels
loss = nn.loss_ms(results, train_labels)
# setup the training operations
train_op = nn.training(loss, FLAGS.learning_rate, FLAGS.decay_steps,
FLAGS.decay_rate)
# setup the summary ops to use TensorBoard
summary_op = tf.merge_all_summaries()
# init to setup the initial values of the weights
init_op = tf.group(tf.initialize_all_variables(),
tf.initialize_local_variables())
# setup a saver for saving checkpoints
saver = tf.train.Saver()
# create the session
sess = tf.Session()
# specify where to write the log files for import to TensorBoard
summary_writer = tf.train.SummaryWriter(FLAGS.checkpoint_dir,
sess.graph)
# initialize the graph
sess.run(init_op)
# setup the coordinato and threadsr. Used for multiple threads to read data.
# Not strictly required since we don't have a lot of data but typically
# using multiple threads to read data improves performance
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess=sess, coord=coord)
# loop will continue until we run out of input training cases
try:
step = 0
while not coord.should_stop():
# start time and run one training iteration
start_time = time.time()
_, loss_value = sess.run([train_op, loss])
duration = time.time() - start_time
# print some output periodically
if step % 100 == 0:
print('OUTPUT: Step %d: loss = %.3f (%.3f sec)' % (step,
loss_value,
duration))
# output some data to the log files for tensorboard
summary_str = sess.run(summary_op)
summary_writer.add_summary(summary_str, step)
summary_writer.flush()
# less frequently output checkpoint files. Used for evaluating the model
if step % 1000 == 0:
checkpoint_path = os.path.join(FLAGS.checkpoint_dir,
MODEL_FILE)
saver.save(sess, checkpoint_path, global_step=step)
step += 1
# quit after we run out of input files to read
except tf.errors.OutOfRangeError:
print('OUTPUT: Done training for %d epochs, %d steps.' % (FLAGS.num_epochs,
step))
checkpoint_path = os.path.join(FLAGS.checkpoint_dir,
'model.ckpt')
saver.save(sess, checkpoint_path, global_step=step)
finally:
coord.request_stop()
# shut down the threads gracefully
coord.join(threads)
sess.close()
def run_training():
# construct the graph
with tf.Graph().as_default():
# specify the training data file location
trainfiles = []
for fi in TRAIN_FILES:
trainfiles.append(os.path.join(FLAGS.data_dir, fi))
# trainfile = os.path.join(FLAGS.data_dir, TRAIN_FILE)
# read the images and labels
x, y_ = nn.inputs(batch_size=FLAGS.batch_size,
num_epochs=FLAGS.num_epochs,
filenames=trainfiles,
ifeval=False)
keep_prob = tf.placeholder(tf.float32)
z_placeholder = tf.placeholder(tf.float32,
[FLAGS.batch_size, z_dimensions])
# run inference on the images
y_conv = nn.inference(x, np.array([65, 65, 65]), keep_prob,
FLAGS.batch_size)
# calculate the loss from the results of inference and the labels
loss = nn.loss(y_conv, y_)
# caculate the accuracy
accuracy = nn.evaluation(y_conv, y_)
# setup the training operations
train_op = nn.training(loss, FLAGS.learning_rate, FLAGS.decay_steps,
FLAGS.decay_rate)
# setup the summary ops to use TensorBoard
summary_op = tf.summary.merge_all()
# init to setup the initial values of the weights
#init_op = tf.group(tf.initialize_all_variables(), tf.initialize_local_variables())
init_op = tf.group(tf.global_variables_initializer(),
tf.local_variables_initializer())
# create the session
with tf.Session() as sess:
sess.run(init_op)
# setup a saver for saving checkpoints
saver = tf.train.Saver()
summary_writer = tf.summary.FileWriter(FLAGS.checkpoint_dir,
sess.graph)
# setup the coordinato and threadsr. Used for multiple threads to read data.
# Not strictly required since we don't have a lot of data but typically
# using multiple threads to read data improves performance
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess=sess, coord=coord)
start_training_time = time.time()
# loop will continue until we run out of input training cases
try:
step = 0
while not coord.should_stop():
# start time and run one training iteration
start_time = time.time()
_, l, acc = sess.run(
[train_op, loss, accuracy],
feed_dict={keep_prob: 0.5}) # Update the discriminator
duration = time.time() - start_time
# print some output periodically
if step % 20 == 0:
print(
'OUTPUT: Step %d: loss = %.3f (%.3f sec), accuracy = %.3f'
% (step, l, duration, acc))
# output some data to the log files for tensorboard
summary_str = sess.run(summary_op)
summary_writer.add_summary(summary_str, step)
summary_writer.flush()
# less frequently output checkpoint files. Used for evaluating the model
if step % 500 == 0:
checkpoint_path = os.path.join(check_save,
'model.ckpt')
saver.save(sess,
save_path=checkpoint_path,
global_step=step)
step += 1
# quit after we run out of input files to read
except tf.errors.OutOfRangeError:
print('OUTPUT: Done training for %d epochs, %d steps.' %
(FLAGS.num_epochs, step))
checkpoint_path = os.path.join(check_save, 'model.ckpt')
saver.save(sess, checkpoint_path, global_step=step)
finally:
coord.request_stop()
# shut down the threads gracefully
coord.join(threads)
sess.close()
end_training_time = time.time()
def run_training():
# construct the graph
with tf.Graph().as_default():
# specify the training data file location
trainfile = os.path.join(FLAGS.data_dir, TRAIN_FILE)
# read the images and labels
images, labels = nn.inputs(batch_size=FLAGS.batch_size,
num_epochs=FLAGS.num_epochs,
filename=trainfile)
# run inference on the images
results = nn.inference(images)
# calculate the loss from the results of inference and the labels
loss = nn.loss(results, labels)
# setup the training operations
train_op = nn.training(loss, FLAGS.learning_rate, FLAGS.decay_steps,
FLAGS.decay_rate)
# setup the summary ops to use TensorBoard
summary_op = tf.summary.merge_all()
# init to setup the initial values of the weights
init_op = tf.group(tf.global_variables_initializer(),
tf.local_variables_initializer())
# setup a saver for saving checkpoints
saver = tf.train.Saver()
# create the session
sess = tf.Session()
# specify where to write the log files for import to TensorBoard
summary_writer = tf.summary.FileWriter(FLAGS.checkpoint_dir,
sess.graph)
# initialize the graph
sess.run(init_op)
# setup the coordinato and threadsr. Used for multiple threads to read data.
# Not strictly required since we don't have a lot of data but typically
# using multiple threads to read data improves performance
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess=sess, coord=coord)
# loop will continue until we run out of input training cases
try:
step = 0
while not coord.should_stop():
# start time and run one training iteration
start_time = time.time()
_, loss_value = sess.run([train_op, loss])
duration = time.time() - start_time
# print some output periodically
if step % 100 == 0:
print('OUTPUT: Step %d: loss = %.3f (%.3f sec)' % (step,
loss_value,
duration))
# output some data to the log files for tensorboard
summary_str = sess.run(summary_op)
summary_writer.add_summary(summary_str, step)
summary_writer.flush()
# less frequently output checkpoint files. Used for evaluating the model
if step % 1000 == 0:
checkpoint_path = os.path.join(FLAGS.checkpoint_dir,
'model.ckpt')
saver.save(sess, checkpoint_path, global_step=step)
step += 1
# quit after we run out of input files to read
except tf.errors.OutOfRangeError:
print('OUTPUT: Done training for %d epochs, %d steps.' % (FLAGS.num_epochs,
step))
checkpoint_path = os.path.join(FLAGS.checkpoint_dir,
'model.ckpt')
saver.save(sess, checkpoint_path, global_step=step)
finally:
coord.request_stop()
# shut down the threads gracefully
coord.join(threads)
sess.close()
