def run_training():
with tf.Graph().as_default():
images, labels = inputs(train=True,
batch_size=FLAGS.batch_size,
num_epochs=FLAGS.num_epochs)
logits = mnist.inference(images, FLAGS.hidden1, FLAGS.hidden2)
loss = mnist.loss(logits, labels)
train_op = mnist.training(loss, FLAGS.learning_rate)
init_op = tf.group(tf.global_variables_initializer(),
tf.local_variables_initializer())
sess = tf.Session()
sess.run(init_op)
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess=sess, coord=coord)
try:
step = 0
while not coord.should_stop():
start_time = time.time()
_, loss_value = sess.run([train_op, loss])
duration = time.time() - start_time
if step % 100 == 0:
print('Step %d: loss = %.2f (%.3f sec)' %
(step, loss_value, duration))
step += 1
except tf.errors.OutOfRangeError:
print('Done training for %d epochs, %d steps.' %
(FLAGS.num_epochs, step))
finally:
coord.request_stop()
coord.join(threads)
sess.close()
def train():
filenames = tf.placeholder(tf.string, [None])
dataset = tf.data.TFRecordDataset(filenames)
dataset = dataset.map(mnist.parse_data)
dataset = dataset.shuffle(buffer_size=50000)
dataset = dataset.batch(FLAGS.batch_size)
dataset = dataset.repeat()
iterator = dataset.make_initializable_iterator()
global_step = tf.train.get_or_create_global_step()
images, labels = iterator.get_next()
logits, pred = mnist.inference(images, training=True)
loss = mnist.loss(logits, labels)
train_op = mnist.train(loss, global_step)
with tf.train.MonitoredTrainingSession(
checkpoint_dir=FLAGS.train_dir,
hooks=[tf.train.StopAtStepHook(last_step=FLAGS.max_step), tf.train.NanTensorHook(loss)],
save_checkpoint_steps=100
) as mon_sess:
mon_sess.run(iterator.initializer, feed_dict={filenames: ['train_img.tfrecords']})
while not mon_sess.should_stop():
_, train_loss, train_step, label = mon_sess.run([train_op, loss, global_step, labels])
if train_step % 100 == 0:
print('step: {}, loss: {}'.format(train_step, train_loss))
def main():
fraction = 0.4
min_after_dequeue = int(mnist.NUM_EXAMPLES_PER_EPOCH * fraction)
images, labels = mnist_inputs(min_after_dequeue)
validation_images, validation_labels = mnist_inputs(2000,
train=False,
num_epochs=None)
with tf.variable_scope("inference") as scope:
logits = mnist.inference(images)
scope.reuse_variables()
validation_logits = mnist.inference(validation_images)
loss = mnist.loss(logits, labels)
tf.scalar_summary("cross_entropy", loss)
accuracy = mnist.accuracy(validation_logits, validation_labels)
tf.scalar_summary("validation_accuracy", accuracy)
train_op = mnist.train(loss)
sess = tf.Session()
sess.run(tf.initialize_local_variables())
sess.run(tf.initialize_all_variables())
tf.train.start_queue_runners(sess=sess)
merge = tf.merge_all_summaries()
writer = tf.train.SummaryWriter(
"/home/windows98/PycharmProjects/mnist/Summary/")
for index in range(NUM_STEPS):
_, loss_value, summary = sess.run([train_op, loss, merge])
writer.add_summary(summary, index + 1)
# accuracy_score, summary = sess.run([accuracy, summary])
# writer.add_summary(summary, index+1)
print("step:" + str(index + 1) + " loss: " + str(loss_value))
def evaluate(dataset):
"""Evaluate model on Dataset for a number of steps."""
with tf.Graph().as_default():
# Graph creation
batch_size = dataset.num_examples
images_placeholder, labels_placeholder = mnist.placeholder_inputs(
batch_size)
logits = mnist.inference(images_placeholder, train=False)
validation_accuracy = tf.reduce_sum(
mnist.evaluation(logits,
labels_placeholder)) / tf.constant(batch_size)
validation_loss = mnist.loss(logits, labels_placeholder)
# Reference to sess and saver
sess = tf.Session()
saver = tf.train.Saver()
# Create summary writer
graph_def = tf.get_default_graph().as_graph_def()
summary_writer = tf.summary.FileWriter(FLAGS.eval_dir,
graph_def=graph_def)
step = -1
while True:
step = do_eval(saver,
summary_writer,
validation_accuracy,
validation_loss,
images_placeholder,
labels_placeholder,
dataset,
prev_global_step=step)
if FLAGS.run_once:
break
time.sleep(FLAGS.eval_interval_secs)
def train():
images, labels = mnist.inputs(['train_img.tfrecords'], mnist.TRAIN_EXAMPLES_NUM,
FLAGS.batch_size, shuffle=True)
global_step = tf.train.get_or_create_global_step()
logits, pred = mnist.inference(images, training=True)
loss = mnist.loss(logits, labels)
train_op = mnist.train(loss, global_step)
saver = tf.train.Saver()
with tf.Session() as sess:
init_op = tf.group(
tf.local_variables_initializer(),
tf.global_variables_initializer())
sess.run(init_op)
ckpt = os.path.join(FLAGS.train_dir, 'model.ckpt')
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess, coord=coord)
for i in range(1, FLAGS.max_step + 1):
_, train_loss, predict, label = sess.run([train_op, loss, pred, labels])
# print(predict, '\n', label)
if i % 100 == 0:
print('step: {}, loss: {}'.format(i, train_loss))
# print(predict, '\n', label)
saver.save(sess, ckpt, global_step=i)
coord.request_stop()
coord.join(threads)
def run_training():
data_sets = data_mnist.read_data_sets()
with tf.Graph().as_default():
images_placeholder, labels_placeholder = placeholder_inputs(FLAGS.batch_size)
logits = mnist.inference(images_placeholder, FLAGS.hidden1, FLAGS.hidden2)
loss = mnist.loss(logits, labels_placeholder)
train_op = mnist.training(loss, FLAGS.learning_rate)
eval_correct = mnist.evaluation(logits, labels_placeholder)
summary_op = tf.merge_all_summaries()
saver = tf.train.Saver()
sess = tf.Session()
sess.run(tf.initialize_all_variables())
summary_writer = tf.train.SummaryWriter(FLAGS.train_dir, sess.graph)
# Start the training loop.
for step in xrange(FLAGS.max_steps):
start_time = time.time()
feed_dict = fill_feed_dict(data_sets.train, images_placeholder, labels_placeholder)
_, loss_value = sess.run([train_op, loss], feed_dict=feed_dict)
duration = time.time() - start_time
if step % 100 == 0:
print('Step %d: loss = %.2f (%.3f sec)' % (step, loss_value, duration))
summary_str = sess.run(summary_op, feed_dict=feed_dict)
summary_writer.add_summary(summary_str, step)
summary_writer.flush()
if (step + 1) % 1000 == 0 or (step + 1) == FLAGS.max_steps:
checkpoint_file = os.path.join(FLAGS.train_dir, 'checkpoint')
saver.save(sess, checkpoint_file, global_step=step)
do_eval(sess,eval_correct, images_placeholder, labels_placeholder, data_sets.train)
do_eval(sess, eval_correct, images_placeholder, labels_placeholder, data_sets.validation)
do_eval(sess, eval_correct, images_placeholder, labels_placeholder, data_sets.test)
def main():
images, labels = inputs()
reshaped_images = tf.reshape(images, [
mnist.BATCH_SIZE, mnist.IMAGE_HEIGHT, mnist.IMAGE_WIDTH,
mnist.IMAGE_DEPTH
])
logits = mnist.inference(reshaped_images)
loss = mnist.loss(logits, labels)
accuracy = mnist.accuracy(logits, labels)
train_op = mnist.train(loss)
init = tf.initialize_all_variables()
with tf.Session() as sess:
sess.run(init)
for index in range(NUM_STEPS):
batch_x, batch_y = mnist_data.train.next_batch(mnist.BATCH_SIZE)
_, loss_value = sess.run([train_op, loss],
feed_dict={
images: batch_x,
labels: batch_y
})
print("step:" + str(index + 1) + " loss: " + str(loss_value))
if (index + 1) % 10 == 0:
validation_x, validation_y = mnist_data.validation.next_batch(
mnist.BATCH_SIZE)
accuracy_score = sess.run(accuracy,
feed_dict={
images: validation_x,
labels: validation_y
})
print("accuracy : " + str(accuracy_score))
def run_training():
data_sets = input_data.read_data_sets(FLAGS.input_data_dir,
FLAGS.fake_data)
with tf.Graph().as_default():
images_placeholder, labels_placeholder = placeholder_inputs(
FLAGS.batch_size)
logits = mnist.inference(images_placeholder, FLAGS.hidden1,
FLAGS.hidden2)
loss = mnist.loss(logits, labels_placeholder)
train_op = mnist.training(loss, FLAGS.learning_rate)
eval_correct = mnist.evaluation(logits, labels_placeholder)
# Build the summary tensor based on the TF collection of Summaries
summary = tf.summary.merge_all()
init = tf.global_variables_initializer()
# Create a saver for writing training checkpoints
saver = tf.train.Saver()
sess = tf.Session()
# Instantiate a SummaryWriter to output summries and the Graph
summary_writer = tf.summary.FileWriter(FLAGS.log_dir, sess.graph)
sess.run(init)
for step in xrange(FLAGS.max_steps):
start_time = time.time()
feed_dict = fill_feed_dict(data_sets.train, images_placeholder,
labels_placeholder)
_, loss_value = sess.run([train_op, loss], feed_dict=feed_dict)
duration = time.time() - start_time
if step % 100 == 0:
print('Step %d: loss = %.2f (%.3f sec)' %
(step, loss_value, duration))
# Update the events file
summary_str = sess.run(summary, feed_dict=feed_dict)
summary_writer.add_summary(summary_str, step)
summary_writer.flush()
# Save a checkpoint and evaluate the model periodically.
if (step + 1) % 1000 == 0 or (step + 1) == FLAGS.max_steps:
checkpoint_file = os.path.join(FLAGS.log_dir, 'model.ckpt')
saver.save(sess, checkpoint_file, global_step=step)
# Evaluate against the training set.
print('Training Data Eval:')
do_eval(sess, eval_correct, images_placeholder,
labels_placeholder, data_sets.train)
# Evaluate against the validation set.
print('Validation Data Eval:')
do_eval(sess, eval_correct, images_placeholder,
labels_placeholder, data_sets.validation)
# Evaluate aginst the test set.
print('Test Data Eval:')
do_eval(sess, eval_correct, images_placeholder,
labels_placeholder, data_sets.test)
def train():
"""Train CIFAR-10 for a number of steps."""
with tf.Graph().as_default():
global_step = tf.train.get_or_create_global_step()
# Get images and labels for CIFAR-10.
# Force input pipeline to CPU:0 to avoid operations sometimes ending up on
# GPU and resulting in a slow down.
with tf.device('/cpu:0'):
images, labels = mnist.distorted_inputs()
# Build a Graph that computes the logits predictions from the
# inference model.
logits = mnist.inference(images)
# Calculate loss.
loss = mnist.loss(logits, labels)
# Build a Graph that trains the model with one batch of examples and
# updates the model parameters.
train_op = mnist.train(loss, global_step)
class _LoggerHook(tf.train.SessionRunHook):
"""Logs loss and runtime."""
def begin(self):
self._step = -1
self._start_time = time.time()
def before_run(self, run_context):
self._step += 1
return tf.train.SessionRunArgs(loss) # Asks for loss value.
def after_run(self, run_context, run_values):
if self._step % FLAGS.log_frequency == 0:
current_time = time.time()
duration = current_time - self._start_time
self._start_time = current_time
loss_value = run_values.results
examples_per_sec = FLAGS.log_frequency * FLAGS.batch_size / duration
sec_per_batch = float(duration / FLAGS.log_frequency)
format_str = (
'%s: step %d, loss = %.2f (%.1f examples/sec; %.3f '
'sec/batch)')
print(format_str % (datetime.now(), self._step, loss_value,
examples_per_sec, sec_per_batch))
with tf.train.MonitoredTrainingSession(
checkpoint_dir=FLAGS.train_dir,
hooks=[
tf.train.StopAtStepHook(last_step=FLAGS.max_steps),
tf.train.NanTensorHook(loss),
_LoggerHook()
],
config=tf.ConfigProto(log_device_placement=FLAGS.
log_device_placement)) as mon_sess:
while not mon_sess.should_stop():
mon_sess.run(train_op)
def run_training():
data_sets = input_data.read_data_sets(fake_data)
with tf.Graph().as_default():
image_placeholder, label_placeholder = placeholder_inputs(batch_size)
logits = mnist.inference(image_placeholder, hidden1_unit, hidden2_unit)
#计算损失
loss = mnist.loss(logits, label_placeholder)
#训练
train_op = mnist.training(loss, 0.01)
#计算正确分类数
eval_correct = mnist.evaluation(logits, label_placeholder)
#合并默认图中的所有summaries操作
summary_op = tf.merge_all_summaries()
#用于保存网络中的变量
saver = tf.train.Saver()
sess = tf.Session()
#初始化所有变量
sess.run(tf.initialize_all_variables())
summary_writter = tf.train.SummaryWriter(train_dir, sess.graph)
for step in xrange(max_steps):
start_time = time.time()
#获取feed_dict字典
feed_dict = fill_placeholder(data_sets.train, image_placeholder,
label_placeholder)
_, loss_value = sess.run([train_op, loss], feed_dict=feed_dict)
duration = time.time() - start_time
if step % 100 == 0:
print(
'Step %d:loss=%.2f (%.3f sec) %(step,loss_value,duration)')
summary_str = sess.run(summary_op, feed_dict=feed_dict)
summary_writter.add_summary(summary_str, step)
summary_writter.flush()
if step % 1000 == 0:
saver.save(sess, train_dir, global_step=step)
print('Training Data Eval:')
do_eval(sess, eval_correct, image_placeholder,
label_placeholder, data_sets.train)
print("Validation Data Eval")
do_eval(sess, eval_correct, image_placeholder,
label_placeholder, data_sets.validation)
print("Test Data Eval:")
do_eval(sess, eval_correct, image_placeholder,
label_placeholder, data_sets.test)
def train():
with tf.Graph().as_default():
global_step = tf.train.get_or_create_global_step()
with tf.device('/cpu:0'):
images, labels = mnist.distorted_inputs()
print(global_step)
#with tf.device('/gpu:0'):
logits = mnist.inference(images)
#with tf.device('/gpu:0'):
loss = mnist.loss(logits, labels)
#with tf.device('/gpu:0'):
train_op = mnist.train(loss, global_step)
class _LoggerHook(tf.train.SessionRunHook):
def begin(self):
self._step = -1
self._start_time = time.time()
def before_run(self, run_context):
self._step += 1
return tf.train.SessionRunArgs(loss)
def after_run(self, run_context, run_values):
if self._step % FLAGS.log_frequency == 0:
current_time = time.time()
duration = current_time - self._start_time
self._start_time = current_time
loss_value = run_values.results
examples_per_sec = FLAGS.log_frequency * FLAGS.batch_size / duration
sec_per_batch = float(duration / FLAGS.log_frequency)
format_str = (
'%s: step %d, loss = %.2f (%.1f examples/sec; %.3f '
'sec/batch)')
print(format_str % (datetime.now(), self._step, loss_value,
examples_per_sec, sec_per_batch))
with tf.train.MonitoredTrainingSession(
checkpoint_dir=FLAGS.train_dir,
hooks=[
tf.train.StopAtStepHook(last_step=FLAGS.max_steps),
tf.train.NanTensorHook(loss),
_LoggerHook()
],
config=tf.ConfigProto(log_device_placement=FLAGS.
log_device_placement)) as mon_sess:
while not mon_sess.should_stop():
mon_sess.run(train_op)
def run_training():
data_sets = input_data.read_data_sets('MNIST_data', FLAGS.fake_data)
with tf.Graph().as_default():
image_placeholder, label_placeholder = placeholder_inputs(FLAGS.batch_size)
logits = mnist.inference(image_placeholder, FLAGS.hidden1_unit, FLAGS.hidden2_unit)
loss = mnist.loss(logits, label_placeholder)
train_op = mnist.training(loss, FLAGS.learning_rate)
eval_correct = mnist.evaluation(logits, label_placeholder)
summary_op = tf.summary.merge_all()
saver = tf.train.Saver()
sess = tf.Session()
sess.run(tf.global_variables_initializer())
summary_writer = tf.summary.FileWriter(FLAGS.train_dir, sess.graph)
for step in range(FLAGS.max_steps):
start_time = time.time()
feed_dict = fill_placeholder(data_sets.train, image_placeholder, label_placeholder)
_, loss_value = sess.run([train_op, loss], feed_dict=feed_dict)
duration = time.time() - start_time
if step % 100 == 0 :
print('Step %d: loss = %.2f (%.3f sec)' % (step, loss_value, duration))
summary_str = sess.run(summary_op, feed_dict = feed_dict)
summary_writer.add_summary(summary_str, step)
summary_writer.flush()
if step % 1000 == 0:
saver.save(sess, FLAGS.train_dir, global_step=step)
print('Training Data Eval: ')
do_eval(sess, eval_correct, image_placeholder, label_placeholder, data_sets.train)
print('Validation Data Eval: ')
do_eval(sess, eval_correct, image_placeholder, label_placeholder, data_sets.validation)
print('Test Data Eval:')
do_eval(sess, eval_correct, image_placeholder, label_placeholder, data_sets.test)
def run_training():
data_sets = input_data.read_data_sets(FLAGS.input_data_dir, FLAGS.fake_data)
with tf.Graph().as_default():
images_placeholder, labels_placeholder = placeholder_inputs(FLAGS.batch_size)
logits = mnist.inference(images_placeholder, FLAGS.hidden1, FLAGS.hidden2) # 模型的部分
loss = mnist.loss(logits, labels_placeholder)
train_op = mnist.training(loss, FLAGS.learning_rate)
eval_correct = mnist.evaluation(logits, labels_placeholder)
summary = tf.summary.merge_all()
init = tf.global_variables_initializer()
saver = tf.train.Saver()
sess = tf.Session()
summary_writer = tf.summary.FileWriter(FLAGS.log_dir, sess.graph)
sess.run(init)
for step in xrange(FLAGS.max_steps):
start_time = time.time()
feed_dict = fill_feed_dict(data_sets.train, images_placeholder, labels_placeholder)
_, loss_value = sess.run([train_op, loss], feed_dict=feed_dict)
duration = time.time() - start_time
if step % 100 == 0:
print 'Setp %d:loss =%.2f (%.3f sec)' % (step, loss_value, duration)
summary_str = sess.run(summary, feed_dict=feed_dict)
summary_writer.add_summary(summary_str, step)
summary_writer.flush()
if (step + 1) % 1000 == 0 or (step + 1) == FLAGS.max_steps:
checkpoint_file = os.path.join(FLAGS.log_dir, 'model.ckpt')
saver.save(sess, checkpoint_file, global_step=step)
print('Training Data Eval:')
do_eval(sess, eval_correct, images_placeholder, labels_placeholder, data_sets.train)
print "Validation Data Eval:"
do_eval(sess, eval_correct, images_placeholder, labels_placeholder, data_sets.validation)
print "Test Data Eval:"
do_eval(sess, eval_correct, images_placeholder, labels_placeholder, data_sets.test)
def run_training():
data_sets = input_data.read_data_sets(FLAGS.input_data_dir,
FLAGS.fake_data)
with tf.Graph().as_default():
# 初始化输入占位符
images_placeholder, labels_placeholder = placeholder_inputs(
FLAGS.batch_size)
# 构建神经网络
logits = mnist.inference(images_placeholder, FLAGS.hidden1,
FLAGS.hidden2)
# 添加损失函数
loss = mnist.loss(logits, labels_placeholder)
# 训练
train_op = mnist.training(loss, FLAGS.learning_rate)
# 正确率计算
eval_correct = mnist.evaluation(logits, labels_placeholder)
init = tf.global_variables_initializer()
sess = tf.Session()
sess.run(init)
start_time = time.time()
for step in range(FLAGS.max_steps):
feed_dict = fill_feed_dict(data_sets.train, images_placeholder,
labels_placeholder)
_, loss_value = sess.run([train_op, loss], feed_dict=feed_dict)
if step % 100 == 0:
duration = time.time() - start_time
print('Step %d: loss = %.2f (%.3f sec)' %
(step, loss_value, duration))
start_time = time.time()
if (step + 1) % 1000 == 0 or (step + 1) == FLAGS.max_steps:
print('Training Data Eval:')
do_eval(sess, eval_correct, images_placeholder,
labels_placeholder, data_sets.train)
print('Validation Data Eval:')
do_eval(sess, eval_correct, images_placeholder,
labels_placeholder, data_sets.validation)
print('Test Data Eval:')
do_eval(sess, eval_correct, images_placeholder,
labels_placeholder, data_sets.test)
print(FLAGS)
def train_and_validation():
training_dataset = tf.data.TFRecordDataset(['./train_img.tfrecords'])
validation_dataset = tf.data.TFRecordDataset(['./validation_img.tfrecords'])
test_dataset = tf.data.TFRecordDataset(['./test_img.tfrecords'])
training_dataset = training_dataset.map(mnist.parse_data)
training_dataset = training_dataset.shuffle(50000).batch(FLAGS.batch_size).repeat()
validation_dataset = validation_dataset.map(mnist.parse_data).batch(FLAGS.batch_size)
test_dataset = test_dataset.map(mnist.parse_data).batch(FLAGS.batch_size)
iterator = tf.data.Iterator.from_structure(output_types=training_dataset.output_types,
output_shapes=training_dataset.output_shapes)
training_init_op = iterator.make_initializer(training_dataset)
validation_init_op = iterator.make_initializer(validation_dataset)
test_init_op = iterator.make_initializer(test_dataset)
images, labels = iterator.get_next()
training = tf.placeholder(dtype=tf.bool)
logits, pred = mnist.inference(images, training=training)
loss = mnist.loss(logits, labels)
top_k_op = tf.nn.in_top_k(logits, labels, 1)
global_step = tf.train.get_or_create_global_step()
train_op = mnist.train(loss, global_step)
saver = tf.train.Saver()
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
sess.run(training_init_op)
print('begin to train!')
ckpt = os.path.join(FLAGS.train_dir, 'model.ckpt')
train_step = 0
while train_step < FLAGS.max_step:
_, train_loss, step, label = sess.run([train_op, loss, global_step, labels], feed_dict={training: True})
train_step += 1
if train_step % 100 == 0:
saver.save(sess, ckpt, train_step)
if train_step % 1000 == 0:
precision = evaluate(sess, top_k_op, training, mnist.TRAIN_EXAMPLES_NUM)
print('step: {}, loss: {}, training precision: {}'.format(train_step, train_loss, precision))
sess.run(validation_init_op)
precision = evaluate(sess, top_k_op, training, mnist.VALIDATION_EXAMPLES_NUM)
print('step: {}, loss: {}, validation precision: {}'.format(train_step, train_loss, precision))
sess.run(training_init_op)
sess.run(test_init_op)
precision = evaluate(sess, top_k_op, training, mnist.TEST_EXAMPLES_NUM)
print('finally test precision: {}'.format(precision))
def run_training():
with tf.Graph().as_default():
# 输入images和labels
images, labels = inputs(train=True,
batch_size=FLAGS.batch_size,
num_epochs=FLAGS.num_epochs)
# 构建一个从推理模型来预测数据的图
logits = mnist.inference(images, FLAGS.hidden1, FLAGS.hidden2)
loss = mnist.loss(logits, labels) # 定义损失函数
# Add to the Graph operations that train the model.
train_op = mnist.training(loss, FLAGS.learning_rate)
# 初始化参数，string_input_producer内部创建了一个epoch计数变量，
# 归入tf.GraphKeys.LOCAL_VARIABLES集合中，必须单独用initialize_local_variable()初始化
init_op = tf.group(tf.global_variables_initializer(),
tf.local_variables_initializer())
sess = tf.Session()
sess.run(init_op)
# Start input enqueue threads
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess=sess, coord=coord)
try:
step = 0
while not coord.should_stop(): # 进入死循环
start_time = time.time()
_, loss_value = sess.run([train_op, loss])
duration = time.time() - start_time
# 每100次训练输出一次结果
if step % 100 == 0:
print('Step %d: loss = %.2f (%.3f sec)' %
(step, loss_value, duration))
step += 1
except tf.errors.OutOfRangeError:
print('Done training for %d epochs, %d steps.' %
(FLAGS.num_epochs, step))
finally:
coord.request_stop() # 通知其他线程关闭
coord.join(threads)
sess.close()
def run_training():
data_sets = input_data.read_data_sets(data_dir)
with tf.Graph().as_default():
input_holder, label_holder = generate_placeholder(50, input_size)
logits = mnist.inference(input_holder, input_size, 128, 32,
label_classes)
loss = mnist.loss(logits, label_holder)
train_op = mnist.training(loss, 0.01)
eval_correct = mnist.evaluation(logits, label_holder)
summary = tf.summary.merge_all()
init = tf.global_variables_initializer()
saver = tf.train.Saver()
sess = tf.Session()
summary_writer = tf.summary.FileWriter(log_dir, sess.graph)
sess.run(init)
for step in range(2000):
start_time = time.time()
feed_dict = fill_feed_dict(data_sets.train, input_holder,
label_holder, 50)
_, loss_value = sess.run([train_op, loss], feed_dict=feed_dict)
duration = time.time() - start_time
if step % 100 == 0:
print('Step %d: loss = %.2f (%.3f sec)' %
(step, loss_value, duration))
summary_str = sess.run(summary, feed_dict=feed_dict)
summary_writer.add_summary(summary_str, step)
summary_writer.flush()
if (step + 1) % 1000 == 0 or (step + 1) == 2000:
checkpoint_file = os.path.join(log_dir, 'model.ckpt')
saver.save(sess, checkpoint_file, global_step=step)
print('Training Data Eval:')
do_eval(sess, eval_correct, input_holder, label_holder,
data_sets.train, 50)
print('Validation Data Eval:')
do_eval(sess, eval_correct, input_holder, label_holder,
data_sets.validation, 50)
print('Test Data Eval:')
do_eval(sess, eval_correct, input_holder, label_holder,
data_sets.test, 50)
def run_training():
data_sets = data_mnist.read_data_sets()
with tf.Graph().as_default():
images_placeholder, labels_placeholder = placeholder_inputs(
FLAGS.batch_size)
logits = mnist.inference(images_placeholder, FLAGS.hidden1,
FLAGS.hidden2)
loss = mnist.loss(logits, labels_placeholder)
train_op = mnist.training(loss, FLAGS.learning_rate)
eval_correct = mnist.evaluation(logits, labels_placeholder)
summary_op = tf.merge_all_summaries()
saver = tf.train.Saver()
sess = tf.Session()
sess.run(tf.initialize_all_variables())
summary_writer = tf.train.SummaryWriter(FLAGS.train_dir, sess.graph)
# Start the training loop.
for step in xrange(FLAGS.max_steps):
start_time = time.time()
feed_dict = fill_feed_dict(data_sets.train, images_placeholder,
labels_placeholder)
_, loss_value = sess.run([train_op, loss], feed_dict=feed_dict)
duration = time.time() - start_time
if step % 100 == 0:
print('Step %d: loss = %.2f (%.3f sec)' %
(step, loss_value, duration))
summary_str = sess.run(summary_op, feed_dict=feed_dict)
summary_writer.add_summary(summary_str, step)
summary_writer.flush()
if (step + 1) % 1000 == 0 or (step + 1) == FLAGS.max_steps:
checkpoint_file = os.path.join(FLAGS.train_dir, 'checkpoint')
saver.save(sess, checkpoint_file, global_step=step)
do_eval(sess, eval_correct, images_placeholder,
labels_placeholder, data_sets.train)
do_eval(sess, eval_correct, images_placeholder,
labels_placeholder, data_sets.validation)
do_eval(sess, eval_correct, images_placeholder,
labels_placeholder, data_sets.test)
def run_training():
data_sets = read_data_sets('/tmp/tensorflow/mnist/input_data', False)
with tf.Graph().as_default():
images_placeholder, labels_placeholder = placeholder_inputs(
100) # 每100张送进去计算一次所以这里的向量维度100
logits = mnist.inference(images_placeholder, 128, 32)
loss = mnist.loss(logits, labels_placeholder)
train_op = mnist.training(loss, 0.01)
summary = tf.summary.merge_all()
saver = tf.train.Saver()
init = tf.global_variables_initializer() #前面都是定义变量，这里对该图中的变量进行初始化
sess = tf.Session()
summary_writer = tf.summary.FileWriter("logs/", sess.graph)
sess.run(init)
for step in xrange(2000): # 每一步取100张，一共2000部，取20万张
feed_dict = fill_feed_dict(data_sets.train, images_placeholder,
labels_placeholder)
_, loss_value = sess.run([train_op, loss], feed_dict=feed_dict)
#sess.run（）函数对loss进行计算，得到lossValue用于查看拟合度是否上身，主要的计算在于执行train_op，返回值_表示匿名，既不关心（应为train_op只是一个计算过程，所以没有返回实体，所以打印并没有区别）
if step % 100 == 0:
print(loss_value)
summary_str = sess.run(summary, feed_dict=feed_dict)
summary_writer.add_summary(summary_str, step)
summary_writer.flush()
# print(train_op)
save_path = saver.save(sess, "myMnistNet/save_net.ckpt")
print(save_path)
def mnist_training():
x = tf.placeholder(tf.float32,
shape=[None, IMG_SIZE * IMG_SIZE],
name='x_ph')
y_ = tf.placeholder(tf.float32, shape=[None, NUM_CLASSES], name='y_ph')
logits = mnist.inference(x)
loss = mnist.loss(logits, y_)
train_op = mnist.training(loss, LR)
eval_correct = mnist.evaluation(logits, y_)
summary = tf.summary.merge_all()
init = tf.global_variables_initializer()
saver = tf.train.Saver(max_to_keep=cfg.MNIST.RUN.models_to_save)
sess = tf.Session()
# Instantiate a SummaryWriter to output summaries and the Graph.
#summary_writer = tf.summary.FileWriter(, sess.graph)
sess.run(init)
print "*****TRAINING STARTED*******"
for i in range(MAX_ITER):
if i % 100 == 0 and i > 0:
print('Step %d: loss = %.2f' % (i, loss_val))
#targets=sess.run(tf.cast(mnist_db.test.labels,tf.int32))
#prediction=sess.run(eval_correct,feed_dict={x:mnist_db.test.images,y_:targets})
#print('Step %d: loss = %.2f, accuracy %.4f' % (i, loss_val,prediction))
saver.save(sess,
os.path.join(cfg.MNIST.RUN.models_dir, 'model'),
global_step=i)
batch = mnist_db.train.next_batch(BATCH_SIZE)
_, loss_val = sess.run([train_op, loss],
feed_dict={
x: batch[0],
y_: batch[1]
})
saver.save(
sess,
os.path.join(cfg.MNIST.RUN.models_dir, cfg.MNIST.RUN.last_model_name))
def tower_loss(scope, images, labels):
# Build inference Graph.
logits = mnist.inference(images, None, None, None, train=True)
# Build the portion of the Graph calculating the losses. Note that we will
# assemble the total_loss using a custom function below.
_ = mnist.loss(logits, labels)
# Assemble all of the losses for the current tower only.
losses = tf.get_collection('losses', scope)
# Calculate the total loss for the current tower.
total_loss = tf.add_n(losses, name='total_loss')
# Attach a scalar summary to all individual losses and the total loss; do the
# same for the averaged version of the losses.
for l in losses + [total_loss]:
# Remove 'tower_[0-9]/' from the name in case this is a multi-GPU training
# session. This helps the clarity of presentation on tensorboard.
loss_name = re.sub('%s_[0-9]*/' % mnist.TOWER_NAME, '', l.op.name)
tf.summary.scalar(loss_name, l)
return total_loss
def run_training():
"""Train MNIST for a number of steps."""
# Get the sets of images and labels for training, validation, and
# test on MNIST.
data_sets = input_data.read_data_sets(FLAGS.train_dir, FLAGS.fake_data)
# Tell TensorFlow that the model will be built into the default Graph.
with tf.Graph().as_default():
# Generate placeholders for the images and labels.
images_placeholder, labels_placeholder = placeholder_inputs(
FLAGS.batch_size)
# Build a Graph that computes predictions from the inference model.
logits = mnist.inference(images_placeholder,
FLAGS.hidden1,
FLAGS.hidden2)
# Add to the Graph the Ops for loss calculation.
loss = mnist.loss(logits, labels_placeholder)
# Add to the Graph the Ops that calculate and apply gradients.
train_op = mnist.training(loss, FLAGS.learning_rate)
# Add the Op to compare the logits to the labels during evaluation.
eval_correct = mnist.evaluation(logits, labels_placeholder)
# Build the summary operation based on the TF collection of Summaries.
summary_op = tf.merge_all_summaries()
# Create a saver for writing training checkpoints.
saver = tf.train.Saver()
# Create a session for running Ops on the Graph.
sess = tf.Session()
# Run the Op to initialize the variables.
init = tf.initialize_all_variables()
sess.run(init)
# Instantiate a SummaryWriter to output summaries and the Graph.
summary_writer = tf.train.SummaryWriter(FLAGS.train_dir,
graph_def=sess.graph_def)
# And then after everything is built, start the training loop.
for step in xrange(FLAGS.max_steps):
start_time = time.time()
# Fill a feed dictionary with the actual set of images and labels
# for this particular training step.
feed_dict = fill_feed_dict(data_sets.train,
images_placeholder,
labels_placeholder)
# Run one step of the model. The return values are the activations
# from the `train_op` (which is discarded) and the `loss` Op. To
# inspect the values of your Ops or variables, you may include them
# in the list passed to sess.run() and the value tensors will be
# returned in the tuple from the call.
_, loss_value = sess.run([train_op, loss],
feed_dict=feed_dict)
duration = time.time() - start_time
# Write the summaries and print an overview fairly often.
if step % 100 == 0:
# Print status to stdout.
print('Step %d: loss = %.2f (%.3f sec)' % (step, loss_value, duration))
# Update the events file.
summary_str = sess.run(summary_op, feed_dict=feed_dict)
summary_writer.add_summary(summary_str, step)
# Save a checkpoint and evaluate the model periodically.
if (step + 1) % 1000 == 0 or (step + 1) == FLAGS.max_steps:
saver.save(sess, FLAGS.train_dir, global_step=step)
# Evaluate against the training set.
print('Training Data Eval:')
do_eval(sess,
eval_correct,
images_placeholder,
labels_placeholder,
data_sets.train)
# Evaluate against the validation set.
print('Validation Data Eval:')
do_eval(sess,
eval_correct,
images_placeholder,
labels_placeholder,
data_sets.validation)
# Evaluate against the test set.
print('Test Data Eval:')
do_eval(sess,
eval_correct,
images_placeholder,
labels_placeholder,
data_sets.test)
def run_training():
# 获取数据
data_sets = input_data.read_data_sets(FLAGS.input_data_dir, FLAGS.fake_data)
# 在默认Graph下运行.
with tf.Graph().as_default():
# 配置graph
images_placeholder, labels_placeholder = placeholder_inputs(FLAGS.batch_size)
# logits是shape为(batch_size，NUM_CLASSES)，表示每条数据预测后未归一化的概率分布
logits = mnist.inference(images_placeholder,
FLAGS.hidden1,
FLAGS.hidden2)
# 损失函数
loss = mnist.loss(logits, labels_placeholder)
train_op = mnist.training(loss, FLAGS.learning_rate)
eval_correct = mnist.evaluation(logits, labels_placeholder)
# 汇聚tensor
summary = tf.summary.merge_all()
# 建立初始化机制
init = tf.global_variables_initializer()
# 建立保存机制
saver = tf.train.Saver()
# 建立Session
sess = tf.Session()
# 建立一个SummaryWriter输出汇聚的tensor
summary_writer = tf.summary.FileWriter(FLAGS.log_dir, sess.graph)
# 开始执行
# 执行变量
sess.run(init)
# 开始训练，2000次循环
for step in xrange(FLAGS.max_steps):
start_time = time.time()
#获取当次循环的数据
feed_dict = fill_feed_dict(data_sets.train,
images_placeholder,
labels_placeholder)
# 丢弃了train数据, 记录loss数据, sess.run(train_op)是最简单的训练代码
# run(self, fetches, feed_dict=None, options=None, run_metadata=None)
# fetches格式很自由，详情见help(tf.Session.run)
_, loss_value = sess.run([train_op, loss], feed_dict=feed_dict)
duration = time.time() - start_time
# 每训练100次输出当前损失，并记录数据
if step % 100 == 0:
print('Step %d: loss = %.2f (%.3f sec)' % (step, loss_value, duration))
summary_str = sess.run(summary, feed_dict=feed_dict)
summary_writer.add_summary(summary_str, step)
summary_writer.flush()
# 每1000次测试模型
if (step + 1) % 1000 == 0 or (step + 1) == FLAGS.max_steps:
checkpoint_file = os.path.join(FLAGS.log_dir, 'model.ckpt')
saver.save(sess, checkpoint_file, global_step=step)
# Evaluate against the training set.
print('Training Data Eval:')
do_eval(sess,
eval_correct,
images_placeholder,
labels_placeholder,
data_sets.train)
# Evaluate against the validation set.
print('Validation Data Eval:')
do_eval(sess,
eval_correct,
images_placeholder,
labels_placeholder,
data_sets.validation)
# Evaluate against the test set.
print('Test Data Eval:')
do_eval(sess,
eval_correct,
images_placeholder,
labels_placeholder,
data_sets.test)
def run_training():
"""
Train MNIST for a number of steps
"""
data_sets = input_data.read_data_sets(FLAGS.train_dir, FLAGS.fake_data)
with tf.Graph().as_default():
images_placeholder, labels_placeholder = placeholder_inputs(FLAGS.batch_size)
logits = mnist.inference(
images_placeholder,
FLAGS.hidden1,
FLAGS.hidden2
)
loss = mnist.loss(logits, labels_placeholder)
train_op = mnist.training(loss, FLAGS.learning_rate)
eval_correct = mnist.evaluation(logits, labels_placeholder)
summary_op = tf.merge_all_summaries()
saver = tf.train.Saver()
sess = tf.Session()
init = tf.initialize_all_variables()
sess.run(init)
summary_writer = tf.training.summary_io.SummaryWriter(
FLAGS.train_dir,
graph_def = sess.graph_def
)
for step in xrange(FLAGS.max_steps):
start_time = time.time()
feed_dict = fill_feed_dict(
data_sets.train,
images_placeholder,
labels_placeholder
)
_, loss_value = sess.run([train_op, loss], feed_dict = feed_dict)
duration = time.time() - start_time
if step % 100 == 0:
print('Step %d: loss = %.2f (%.3f sec)' % (step, loss_value, duration))
summary_str = sess.run(summary_op, feed_dict=feed_dict)
summary_writer.add_summary(summary_str, step)
if (step + 1) % 1000 == 0 or (step + 1) == FLAGS.max_steps:
saver.save(sess, FLAGS.train_dir, global_step = step)
print('Evaluating on training data...')
do_eval(
sess,
eval_correct,
images_placeholder,
labels_placeholder,
data_sets.train
)
print('Evaluating on validation data...')
do_eval(
sess,
eval_correct,
images_placeholder,
labels_placeholder,
data_sets.validation
)
print('Evaluating on test data...')
do_eval(
sess,
eval_correct,
images_placeholder,
labels_placeholder,
data_sets.test
)
def tower_loss(scope):
images, labels = mnist_inputs(min_after_dequeue=MIN_AFTER_DEQUEUE)
logits = mnist.inference(images)
_ = mnist.loss(logits, labels)
loss = tf.get_collection("loss", scope=scope)[0]
return loss
def run_training():
"""
TRAIN MNIST for a number of steps
"""
# Get the sets of images and labels for training, validation, and
# test on MNIST
data_sets = input_data.read_data_sets(FLAGS.input_data_dir,
FLAGS.fake_data)
# Tell TensorFlow that the model will be built into the default Graph
with tf.Graph().as_default():
# Generate placeholders for the images and labels
images_placeholder, labels_placeholder = placeholder_inputs(
FLAGS.batch_size)
# Build a Graph that computes predictions from the inference model
logits = mnist.inference(images_placeholder, FLAGS.hidden1,
FLAGS.hidden2)
# Add to the Graph the Ops for loss calculation
loss = mnist.loss(logits, labels_placeholder)
# Add to the Graph the Ops that calculate and apply gradients
train_op = mnist.training(loss, FLAGS.learning_rate)
# Add the Op to compare the logits to the labels during evaluation
eval_correct = mnist.evaluation(logits, labels_placeholder)
# Build the summary Tensor based on the TF collection of Summaries
summary = tf.summary.merge_all()
# Add the variable initializer Op
init = tf.global_variables_initializer()
# Create a saver for writing training checkpoints
saver = tf.train.Saver()
# Create a session for running Ops on the Graph
sess = tf.Session()
# Instantiate a SummaryWriter to output summaries and the Graph
summary_writer = tf.summary.FileWriter(FLAGS.log_dir, sess.graph)
# And then after everything is built:
# Run the Op to initialize the variables
sess.run(init)
# Start the training loop
for step in xrange(FLAGS.max_steps):
start_time = time.time()
"""
TODO: Add progress bar
"""
# Fill a feed dictionary with the actual set of images and labels
# for t his particular training step
feed_dict = fill_feed_dict(data_sets.train, images_placeholder,
labels_placeholder)
# Run one step of the model. The return values are the activations
# from the 'train_op' (which is discarded) and the 'loss' Op.
# To inspect the values of your Ops or variables, you m ay include
# them in the list passed to sess.run() and the value tensors
# will be returned in the tuple from the calculate
_, loss_value = sess.run([train_op, loss], feed_dict=feed_dict)
duration = time.time() - start_time
# Write the summaries and print an overview fairly often
if step % 100 == 0:
# Print status to stdout
print('Step %d: loss = %.2f (%.3f)' %
(step, loss_value, duration))
#Update the events file
summary_str = sess.run(summary, feed_dict=feed_dict)
summary_writer.add_summary(summary_str, step)
summary_writer.flush()
# Save a checkpoint and evaluate the model periodically
if (step + 1) % 1000 == 0 or (step + 1) == FLAGS.max_steps:
checkpoint_file = os.path.join(FLAGS.log_dir, 'model.ckpt')
saver.save(sess, checkpoint_file, global_step=step)
# Evaluate against the training set
print('Training Data Eval:')
do_eval(sess, eval_correct, images_placeholder,
labels_placeholder, data_sets.train)
# Evaluate against the validationset
print('Validation Data Eval:')
do_eval(sess, eval_correct, images_placeholder,
labels_placeholder, data_sets.validation)
# Evaluate against the test set
print('Test Data Eval:')
do_eval(sess, eval_correct, images_placeholder,
labels_placeholder, data_sets.test)
def run_training():
"""Train MNIST for a number of steps."""
# Get the sets of images and labels for training, validation, and test on MNIST.
data_sets = aymericdamien.input_data.read_data_sets(FLAGS.train_dir, FLAGS.fake_data)
# Tell TensorFlow that the model will be built into the default Graph.
with tf.Graph().as_default():
# Generate placeholders for the images and labels
images_placeholder, labels_placeholder = placeholder_inputs(FLAGS.batch_size)
# Build a graph the Ops for loss calculation
logits = mnist.inference(
images_placeholder,
FLAGS.hidden1,
FLAGS.hidden2
)
# Add to the graph ops for loss calculation
loss = mnist.loss(logits, labels_placeholder)
# Add to the Graph the Ops that calculate and apply gradients.
train_op = mnist.training(loss, FLAGS.learning_rate)
# Add the Op to compare the logits to the labels during evaluation
eval_correct = mnist.evaluation(logits, labels_placeholder)
# Build the summary operation based on the TF collection of Summaries.
summary_op = tf.merge_all_summaries()
# Create a saver for writing training checkpoints
saver = tf.train.Saver()
# create a session for running ops on the graph
session = tf.Session()
# Run the Op to initialize the variables
init = tf.initialize_all_variables()
session.run(init)
# Instantiate a SummaryWriter to output summaries and the Graph
summary_writer = tf.train.SummaryWriter(
FLAGS.train_dir,
graph_def=session.graph_def
)
# After everything is built start the training loop
for step in xrange(FLAGS.max_steps):
start_time = time()
# Fill the feed dictionary with the actual set of images and labels
# for this training step
feed_dict = fill_feed_dict(
data_sets.train,
images_placeholder,
labels_placeholder
)
# run one step of the model, return values are the activations from
# the 'train_op' (which is discarded) and the 'loss' Op. To inspect
# the values of your Ops or Variables, you may include them in the list
# passed to the session.Run() and the value tensors will be returned in
# the tuple from the call
_, loss_value = session.run(
[train_op, loss],
feed_dict=feed_dict
)
duration = time() - start_time
if step % 100 == 0:
# print status update
print('Step %d: loss = %.2f (%.3f sec)' % (step, loss_value, duration))
# Update the events file.
summary_str = session.run(summary_op, feed_dict=feed_dict)
summary_writer.add_summary(summary_str, step)
# Save a checkpoint and evaluate the model periodically
if (step + 1) % 1000 == 0 or (step + 1) == FLAGS.max_steps:
saver.save(session, FLAGS.train_dir, global_step=step)
# Evaluate against the training set.
print('Training Data Eval:')
do_evaluation(
session,
eval_correct,
images_placeholder,
labels_placeholder,
data_sets.train)
# Evaluate against the validation set.
print('Validation Data Eval:')
do_evaluation(
session,
eval_correct,
images_placeholder,
labels_placeholder,
data_sets.validation)
# Evaluate against the test set.
print('Test Data Eval:')
do_evaluation(
session,
eval_correct,
images_placeholder,
labels_placeholder,
data_sets.test)
def run_training():
# Get the sets of images and labels for training, validation, and test on MNIST.
data_sets = input_data.read_data_sets(FLAGS.input_data_dir,
FLAGS.fake_data)
# Tell TensorFlow that the model will be built into the default Graph.
with tf.Graph().as_default():
image_placeholder, labels_placeholder = placeholder_inputs(
FLAGS.batch_size, mnist.IMAGE_PIEXLS)
logits = mnist.inference(image_placeholder, FLAGS.hidden_unt1,
FLAGS.hidden_unt2)
loss = mnist.loss(logits, labels_placeholder)
train_op = mnist.training(loss, learning_rate=FLAGS.lr)
eval_correct = mnist.evalution(logits, labels_placeholder)
summary = tf.summary.merge_all()
# Add the variable initializer Op.
init = tf.global_variables_initializer()
# Create a saver for writing training checkpoints.
saver = tf.train.Saver()
with tf.Session() as sess:
# Instantiate a SummaryWriter to output summaries and the Graph.
summary_writer = tf.summary.FileWriter(FLAGS.log_dir, sess.graph)
# Run the Op to initialize the variables.
sess.run(init)
feed_dict = fill_feed_dict(data_sets.train, image_placeholder,
labels_placeholder)
for step in xrange(FLAGS.max_steps):
start_time = time.time()
_, loss_value = sess.run([train_op, loss], feed_dict=feed_dict)
duration = time.time() - start_time
if step % 100 == 0:
# Print status to stdout.
print('Step %d: loss = %.2f (%.3f sec)' %
(step, loss_value, duration))
# Update the events file.
summary_str = sess.run(summary, feed_dict=feed_dict)
summary_writer.add_summary(summary_str, step)
summary_writer.flush()
# Save a checkpoint and evaluate the model periodically.
if (step + 1) % 1000 == 0 or (step + 1) == FLAGS.max_steps:
checkpoint_file = os.path.join(FLAGS.log_dir, 'model.ckpt')
saver.save(sess, checkpoint_file, global_step=step)
# Evaluate against the training set.
print('Training Data Eval:')
do_eval(sess, eval_correct, image_placeholder,
labels_placeholder, data_sets.train)
# Evaluate against the validation set.
print('Validation Data Eval:')
do_eval(sess, eval_correct, image_placeholder,
labels_placeholder, data_sets.validation)
# Evaluate against the test set.
print('Test Data Eval:')
do_eval(sess, eval_correct, image_placeholder,
labels_placeholder, data_sets.test)
def run_training(learning_rate=FLAGS.learning_rate,
momentum=FLAGS.momentum,
max_norm=FLAGS.max_norm,
weight_decay=FLAGS.weight_decay,
keep_prob=FLAGS.keep_prob,
keep_input=FLAGS.keep_input,
beta2=FLAGS.beta2,
num_layers=FLAGS.num_layers):
"""Train MNIST for a number of steps."""
# Get the sets of images and labels for training, validation, and
# test on MNIST.
data_sets = input_data.read_data_sets(FLAGS.train_dir, FLAGS.fake_data)
# Tell TensorFlow that the model will be built into the default Graph.
with tf.Graph().as_default():
# Generate placeholders for the images and labels.
images_placeholder = tf.placeholder(tf.float32, shape=(None,
mnist.IMAGE_PIXELS), name='images')
labels_placeholder = tf.placeholder(tf.int32, shape=[None], name='labels')
keep_prob_pl = tf.placeholder(tf.float32, name='keep_prob_pl')
keep_input_pl = tf.placeholder(tf.float32, name='keep_input_pl')
learning_rate_pl = tf.placeholder(tf.float32, name='learning_rate_pl')
def fill_feed_dict(data_set, batch_size=FLAGS.batch_size):
# Create the feed_dict for the placeholders filled with the next
# `batch size ` examples.
images_feed, labels_feed = data_set.next_batch(batch_size,
FLAGS.fake_data)
feed_dict = {
images_placeholder: images_feed,
labels_placeholder: labels_feed,
keep_prob_pl: keep_prob,
keep_input_pl: keep_input,
learning_rate_pl: learning_rate
}
return feed_dict
def fill_feed_dict_eval(data_set):
return {
images_placeholder: data_set._images,
labels_placeholder: data_set._labels,
keep_prob_pl: 1.0,
keep_input_pl: 1.0,
}
# Build a Graph that computes predictions from the inference model.
with tf.variable_scope('feed_forward_model') as scope:
logits, bn = mnist.inference(images_placeholder,
FLAGS.hidden1,
num_layers,
weight_decay,
keep_prob_pl,
keep_input_pl,
max_norm)
# Add to the Graph the Ops for loss calculation.
loss = mnist.loss(logits, labels_placeholder)
#loss_eval = mnist.loss( logits_eval, labels_placeholder)
# Add to the Graph the Ops that calculate and apply gradients.
train_op = mnist.training(loss, learning_rate_pl, momentum, beta2)
with tf.control_dependencies([train_op]):
train_op = tf.group(*[b.get_assigner() for b in bn])
# Add the Op to compare the logits to the labels during evaluation.
eval_correct = mnist.evaluation(logits, labels_placeholder)
results = tf.placeholder( tf.float32, [4])
summarize_evaluation = tf.scalar_summary(['correct_train', 'loss_train', 'correct_test', 'loss_test'], results)
# Build the summary operation based on the TF collection of Summaries.
summary_op = tf.merge_all_summaries()
# Create a saver for writing training checkpoints.
saver = tf.train.Saver(max_to_keep=2)
train_loss = test_loss = 0
train_cor = test_cor = 0.97
previous_test_loss = None
first_step = 0
# Create a session for running Ops on the Graph.
sess = tf.Session()
summary_writer = tf.train.SummaryWriter(FLAGS.train_dir, sess.graph_def)
restore_path = tf.train.latest_checkpoint("/Users/mikowals/projects/mnist")
if restore_path:
saver.restore(sess, restore_path)
first_step = int(restore_path.split('/')[-1].split('-')[-1])
print('retored variables from ', restore_path)
else:
# Run the Op to initialize the variables.
print('initializing variables')
init = tf.initialize_all_variables()
sess.run(init)
# And then after everything is built, start the training loop.
for step in range(first_step,FLAGS.max_steps):
start_time = time.time()
# Fill a feed dictionary with the actual set of images and labels
# for this particular training step.
feed_dict = fill_feed_dict(data_sets.train)
# Run one step of the model. The return values are the activations
# from the `train_op` (which is discarded) and the `loss` Op. To
# inspect the values of your Ops or variables, you may include them
# in the list passed to sess.run() and the value tensors will be
# returned in the tuple from the call.
_, loss_value = sess.run([train_op, loss],
feed_dict=feed_dict)
duration = time.time() - start_time
# Write the summaries and print an overview fairly often.
# Save a checkpoint and evaluate the model periodically.
if (step + 1) % 1000 == 0 or (step + 1) == FLAGS.max_steps:
saver.save(sess, FLAGS.train_dir, global_step=step)
# Evaluate against the training set.
# Evaluate against the validation set.
print('training Data Eval:')
feed_dict = fill_feed_dict_eval(data_sets.train)
train_cor, train_loss = sess.run([eval_correct, loss], feed_dict=feed_dict)
train_cor = train_cor / data_sets.train.num_examples
print(train_cor, train_loss)
print('Validation Data Eval:')
feed_dict = fill_feed_dict_eval(data_sets.validation)
test_cor, test_loss = sess.run([eval_correct, loss], feed_dict=feed_dict)
test_cor = test_cor / data_sets.validation.num_examples
print (test_cor, test_loss )
#if previous_test_loss and test_loss > previous_test_loss:
# learning_rate = learning_rate * 0.6
#if previous_test_loss and test_loss < previous_test_loss:
# learning_rate = learning_rate * 1.02
#previous_test_loss = test_loss
if step > 1000 and step % 100 == 0:
# Print status to stdout.
print('Step %d: loss = %.2f (%.3f sec)' % (step, loss_value, duration))
# Update the events file.
feed_dict[results] = [
train_cor,
train_loss,
test_cor,
test_loss]
summary_str = sess.run(summary_op, feed_dict=feed_dict)
summary_writer.add_summary(summary_str, step)
return -test_cor
def train_model():
"""
This method deals with training the model in mnist-model.py by running optimization over a number of steps.
"""
#read input da(ta first
#read_data_sets() function will ensure that the correct data has been downloaded to your local training folder
#and then unpack that data to return a dictionary of DataSet instances.
#FLAGS.fake_data can be ignored as it is used-for unit testing purposes
print("Training Started! ")
data = input_data.read_data_sets(FLAGS.input_data_dir, FLAGS.fake_data)
#We need to specify that our model will be used with the Default global Tensor Flow graph.
#A default global TF graph tf.Graph() is a collection of ops that may be executed as a group
with tf.Graph().as_default():
# Generate placeholders for the images and labels.
images_ph, labels_ph = input_placeholders(FLAGS.batch_size)
# Build a Graph that computes predictions from the inference model.
logits = mnist.inference(images_ph, FLAGS.num_hidden1_nodes,
FLAGS.num_hidden1_nodes)
# Add the loss calculation op to the default Graph
loss = mnist.loss(logits, labels_ph)
# Add the minimization op the Graph
train_op = mnist.training(loss, FLAGS.eta)
# Add the Evaluation to test predictions to the Graph
eval_correct = mnist.evaluation(logits, labels_ph)
# Build the summary Tensor based on the TF collection of Summaries.
summary = tf.summary.merge_all()
# Add the variable initializer Op. to the Graph
init = tf.global_variables_initializer()
#Create a Tensor Flow Saver for writing Training Checkpoints
saver = tf.train.Saver()
#Now once all the build preparation is completed and all the ops are added to the Graph,
#we need to create a Session in order to run the computaional Graph
sess = tf.Session()
#We also need to create a TF Summary Writer in order to record all the summaries and the Graph
summary_writer = tf.summary.FileWriter(FLAGS.log_dir, sess.graph)
#Now all the required Ops are attached to the Default Graph and all is built,
#Start the session by initialing all the TF variables.
sess.run(init)
#Start the Training
for step in xrange(FLAGS.max_steps):
start_time = time.time()
#Update the feed_dict with the next batch of samples to train with.
feed_dict = fill_feed_dictionary(data.train, images_ph, labels_ph)
#Run one step of the training by running Ops train_op and loss
#No need to store the activations returned by the train_op minimization step
_, loss_val = sess.run([train_op, loss], feed_dict=feed_dict)
duration = time.time() - start_time
#Record all the training summaries generates and print the training progress/statistics
#after every 100th iteration
if step % 100 == 0:
# Print status to stdout.
print('Step %d: loss = %.2f (%.3f sec)' %
(step, loss_val, duration))
#Now we need to update the events file with summaries
#Run the summary Op attached to the Graph
#Everytime the summary is evaluated, new summaries are written into the events files
summary_str = sess.run(summary, feed_dict=feed_dict)
summary_writer.add_summary(summary_str, step)
summary_writer.flush()
#Save the Model at every 1000th iteration and perform evaluation on complete data
if (step + 1) % 1000 == 0 or (step + 1) == FLAGS.max_steps:
checkpoint_file = os.path.join(FLAGS.log_dir, 'model.ckpt')
saver.save(sess, checkpoint_file, global_step=step)
#Evaluate against the training set
print('Training Data Evaluation:')
do_eval(sess, eval_correct, images_ph, labels_ph, data.train)
print('Validation Data Evaluation:')
do_eval(sess, eval_correct, images_ph, labels_ph, data.validation)
# Evaluate against the test set.
print('Test Data Evaluation:')
do_eval(sess, eval_correct, images_ph, labels_ph, data.test)
def run_training():
"""Train MNIST for a number of steps."""
# Get the sets of images and labels for training, validation, and
# test on MNIST.
data_sets = input_data.read_data_sets(FLAGS.train_dir, FLAGS.fake_data)
# labels and images are properties of the DataSet class, which is
# defined at tensorflow/contrib/learn/python/learn/datasets/mnist.py
# numpy.savetxt("/tmp/xx.csv", data_sets.train.labels, delimiter=",")
# numpy.savetxt("/tmp/yy.csv", data_sets.test.images, delimiter=",")
# sys.exit(1)
# Tell TensorFlow that the model will be built into the default Graph.
with tf.Graph().as_default():
# Generate placeholders for the images and labels.
images_placeholder, labels_placeholder = placeholder_inputs(FLAGS.batch_size)
# Build a Graph that computes predictions from the inference model.
logits = mnist.inference(images_placeholder, FLAGS.hidden1, FLAGS.hidden2)
# Add to the Graph the Ops for loss calculation.
loss = mnist.loss(logits, labels_placeholder)
# Add to the Graph the Ops that calculate and apply gradients.
train_op = mnist.training(loss, FLAGS.learning_rate)
# Add the Op to compare the logits to the labels during evaluation.
eval_correct = mnist.evaluation(logits, labels_placeholder)
# Build the summary operation based on the TF collection of Summaries.
summary_op = tf.merge_all_summaries()
# Add the variable initializer Op.
init = tf.initialize_all_variables()
# Create a saver for writing training checkpoints.
saver = tf.train.Saver()
# Create a session for running Ops on the Graph.
sess = tf.Session()
# Instantiate a SummaryWriter to output summaries and the Graph.
summary_writer = tf.train.SummaryWriter(FLAGS.train_dir, sess.graph)
# And then after everything is built:
# Run the Op to initialize the variables.
sess.run(init)
# Start the training loop.
for step in xrange(FLAGS.max_steps):
start_time = time.time()
# Fill a feed dictionary with the actual set of images and labels
# for this particular training step.
feed_dict = fill_feed_dict(data_sets.train, images_placeholder, labels_placeholder)
# Run one step of the model. The return values are the activations
# from the `train_op` (which is discarded) and the `loss` Op. To
# inspect the values of your Ops or variables, you may include them
# in the list passed to sess.run() and the value tensors will be
# returned in the tuple from the call.
_, loss_value = sess.run([train_op, loss], feed_dict=feed_dict)
duration = time.time() - start_time
# Write the summaries and print an overview fairly often.
if step % 100 == 0:
# Print status to stdout.
print("Step %d: loss = %.2f (%.3f sec)" % (step, loss_value, duration))
# Update the events file.
summary_str = sess.run(summary_op, feed_dict=feed_dict)
summary_writer.add_summary(summary_str, step)
summary_writer.flush()
# Save a checkpoint and evaluate the model periodically.
if (step + 1) % 1000 == 0 or (step + 1) == FLAGS.max_steps:
checkpoint_file = os.path.join(FLAGS.train_dir, "checkpoint")
saver.save(sess, checkpoint_file, global_step=step)
# Evaluate against the training set.
print("Training Data Eval:")
do_eval(sess, eval_correct, images_placeholder, labels_placeholder, data_sets.train)
# Evaluate against the validation set.
print("Validation Data Eval:")
do_eval(sess, eval_correct, images_placeholder, labels_placeholder, data_sets.validation)
# Evaluate against the test set.
print("Test Data Eval:")
do_eval(sess, eval_correct, images_placeholder, labels_placeholder, data_sets.test)
if (step + 1) == FLAGS.max_steps:
float_formatter = lambda x: "%.2f" % x
numpy.set_printoptions(formatter={"float_kind": float_formatter})
feed_dict = fill_feed_dict(data_sets.test, images_placeholder, labels_placeholder)
# output with softmax
# output = sess.run(tf.nn.softmax(logits), feed_dict=feed_dict)
# output without softmax
output = sess.run(tf.argmax(logits, dimension=1), feed_dict=feed_dict)
numpy.savetxt("/tmp/outputX.csv", output, delimiter=",")
def run_training():
"""Train MNIST for a number of steps."""
# Get the sets of images and labels for training, validation, and
# test on MNIST.
data_sets=reader(patchlength=0,\
maxlength=300,\
embedding_size=100,\
num_verbs=10,\
allinclude=False,\
shorten=False,\
shorten_front=False,\
testflag=False,\
passnum=0,\
dpflag=False)
# Tell TensorFlow that the model will be built into the default Graph.
with tf.Graph().as_default():
# Generate placeholders for the images and labels.
images_placeholder, labels_placeholder = placeholder_inputs(
FLAGS.batch_size)
# Build a Graph that computes predictions from the inference model.
logits, keep_prob = mnist.inference(images_placeholder, FLAGS.hidden1,
FLAGS.hidden2)
# Add to the Graph the Ops for loss calculation.
loss = mnist.loss(logits, labels_placeholder)
# Add to the Graph the Ops that calculate and apply gradients.
train_op = mnist.training(loss, FLAGS.learning_rate)
# Add the Op to compare the logits to the labels during evaluation.
eval_correct = mnist.evaluation(logits, labels_placeholder)
# Build the summary Tensor based on the TF collection of Summaries.
summary = tf.summary.merge_all()
# Add the variable initializer Op.
init = tf.global_variables_initializer()
# Create a saver for writing training checkpoints.
saver = tf.train.Saver()
# Create a session for running Ops on the Graph.
sess = tf.Session()
# Instantiate a SummaryWriter to output summaries and the Graph.
summary_writer = tf.summary.FileWriter(FLAGS.log_dir, sess.graph)
# And then after everything is built:
# Run the Op to initialize the variables.
with tf.Session() as session:
sess.run(init)
if True:
model_file = tf.train.latest_checkpoint(FLAGS.log_dir)
saver.restore(sess, model_file)
# Start the training loop.
start_time = time.time()
for step in xrange(FLAGS.max_steps):
# Fill a feed dictionary with the actual set of images and labels
# for this particular training step.
inputs, answers = data_sets.list_tags(FLAGS.batch_size,
test=False)
# print(len(inputs),len(inputs[0]),inputs[0])
# input()
inputs2 = []
for i in range(len(inputs)):
inputs2.append(inputs[i] / 255)
# print(len(inputs2),len(inputs2[0]),inputs2[0])
# input()
feed_dict = {
images_placeholder: inputs2,
labels_placeholder: answers,
keep_prob: 0.5
}
# Run one step of the model. The return values are the activations
# from the `train_op` (which is discarded) and the `loss` Op. To
# inspect the values of your Ops or variables, you may include them
# in the list passed to sess.run() and the value tensors will be
# returned in the tuple from the call.
_, loss_value, logi = sess.run([train_op, loss, logits],
feed_dict=feed_dict)
duration = time.time() - start_time
# Write the summaries and print an overview fairly often.
if step % 100 == 0:
# Print status to stdout.
print('Step %d: loss = %.2f (%.3f sec)' %
(step, loss_value, duration))
# print(logi)
# print(answers)
for i0 in range(FLAGS.batch_size):
lgans = np.argmax(logi[i0])
if (lgans != answers[i0] and False):
for tt in range(784):
if (tt % 28 == 0): print(' ')
if (inputs[i0][tt] != 0):
print('1', end=' ')
else:
print('0', end=' ')
# print('np',np.argmax(i),answers,answers[i0],'np')
print(lgans, answers[i0])
# Update the events file.
summary_str = sess.run(summary, feed_dict=feed_dict)
summary_writer.add_summary(summary_str, step)
summary_writer.flush()
if (step + 1) % 500 == 0 or (step + 1) == FLAGS.max_steps:
#print('Training Data Eval:')
do_eval(sess, eval_correct, data_sets, FLAGS.batch_size,
images_placeholder, labels_placeholder, keep_prob)
do_evalfake(sess, eval_correct, data_sets,
FLAGS.batch_size, images_placeholder,
labels_placeholder, logits, keep_prob)
# Save a checkpoint and evaluate the model periodically.
#if (step + 1) % 1000 == 0 or (step + 1) == FLAGS.max_steps:
checkpoint_file = os.path.join(FLAGS.log_dir, 'model.ckpt')
saver.save(sess, checkpoint_file, global_step=step)
print('saved to', checkpoint_file)
'''
def train(target, dataset, cluster_spec):
"""Train Inception on a dataset for a number of steps."""
# Number of workers and parameter servers are infered from the workers and ps
# hosts string.
num_workers = len(cluster_spec.as_dict()['worker'])
num_parameter_servers = len(cluster_spec.as_dict()['ps'])
# If no value is given, num_replicas_to_aggregate defaults to be the number of
# workers.
if FLAGS.num_replicas_to_aggregate == -1:
num_replicas_to_aggregate = num_workers
else:
num_replicas_to_aggregate = FLAGS.num_replicas_to_aggregate
# Both should be greater than 0 in a distributed training.
assert num_workers > 0 and num_parameter_servers > 0, (
' num_workers and '
'num_parameter_servers'
' must be > 0.')
# Choose worker 0 as the chief. Note that any worker could be the chief
# but there should be only one chief.
is_chief = (FLAGS.task_id == 0)
# Ops are assigned to worker by default.
with tf.device(
tf.train.replica_device_setter(
worker_device='/job:worker/task:%d' % FLAGS.task_id,
cluster=cluster_spec)):
# Create a variable to count the number of train() calls. This equals the
# number of updates applied to the variables. The PS holds the global step.
global_step = tf.Variable(0, name="global_step", trainable=False)
# Calculate the learning rate schedule.
num_batches_per_epoch = (dataset.num_examples / FLAGS.batch_size)
# Decay steps need to be divided by the number of replicas to aggregate.
# This was the old decay schedule. Don't want this since it decays too fast with a fixed learning rate.
decay_steps = int(num_batches_per_epoch * FLAGS.num_epochs_per_decay /
num_replicas_to_aggregate)
# New decay schedule. Decay every few steps.
#decay_steps = int(num_batches_per_epoch * FLAGS.num_epochs_per_decay / num_workers)
# Decay the learning rate exponentially based on the number of steps.
lr = tf.train.exponential_decay(FLAGS.initial_learning_rate,
global_step,
decay_steps,
FLAGS.learning_rate_decay_factor,
staircase=True)
images, labels = mnist.placeholder_inputs(FLAGS.batch_size)
# Number of classes in the Dataset label set plus 1.
# Label 0 is reserved for an (unused) background class.
logits = mnist.inference(images)
# Add classification loss.
total_loss = mnist.loss(logits, labels)
# Create an optimizer that performs gradient descent.
opt = tf.train.AdamOptimizer(lr)
# Use V2 optimizer
if FLAGS.interval_method or FLAGS.worker_times_cdf_method:
opt = TimeoutReplicasOptimizer(
opt,
global_step,
replicas_to_aggregate=num_replicas_to_aggregate,
total_num_replicas=num_workers)
else:
opt = WeightedGradsOptimizer(
# opt = tf.train.SyncReplicasOptimizer(
opt,
replicas_to_aggregate=num_replicas_to_aggregate,
total_num_replicas=num_workers)
# Compute gradients with respect to the loss.
grads = opt.compute_gradients(total_loss)
#===============================================================================================
# batch_idx_placeholder = tf.placeholder(dtype=tf.int32, shape=(int(num_workers),))
# worker_kill_placeholder = tf.placeholder(dtype=tf.int32, shape=(FLAGS.num_worker_kill,))
matrix_placeholder = tf.placeholder(dtype=tf.float32,
shape=((int(num_batches_per_epoch),
int(num_workers))))
'''
weight_vec_placeholder = tf.placeholder(dtype=tf.float32,
shape=(num_workers,))
grad_list = [x[0] for x in grads]
new_grad_list = []
for g_idx in range(len(grad_list)):
grad_on_worker = grad_list[g_idx]
weight = tf.slice(weight_vec_placeholder, [i], [1])
new_grad_list.append(tf.mul(grad_on_worker, weight))
for x_idx in range(len(grads)):
x = grads[x_idx]
x[0] = new_grad_list[x_idx]
'''
#===============================================================================================
if FLAGS.interval_method or FLAGS.worker_times_cdf_method:
# apply_gradients_op = opt.apply_gradients(grads, FLAGS.task_id, global_step=global_step, collect_cdfs=FLAGS.worker_times_cdf_method,
# batch_idx_list=batch_idx_placeholder, worker_kill_list=worker_kill_placeholder,
# num_workers=int(num_workers), num_batches_per_epoch=int(num_batches_per_epoch))
apply_gradients_op = opt.apply_gradients(
grads,
FLAGS.task_id,
global_step=global_step,
collect_cdfs=FLAGS.worker_times_cdf_method,
matrix_to_solve=matrix_placeholder,
num_batches_per_epoch=int(num_batches_per_epoch))
else:
apply_gradients_op = opt.apply_gradients(grads,
global_step=global_step)
with tf.control_dependencies([apply_gradients_op]):
train_op = tf.identity(total_loss, name='train_op')
# Get chief queue_runners, init_tokens and clean_up_op, which is used to
# synchronize replicas.
# More details can be found in sync_replicas_optimizer.
chief_queue_runners = [opt.get_chief_queue_runner()]
init_tokens_op = opt.get_init_tokens_op()
#clean_up_op = opt.get_clean_up_op()
# Create a saver.
saver = tf.train.Saver()
# Build the summary operation based on the TF collection of Summaries.
summary_op = tf.summary.merge_all()
# Build an initialization operation to run below.
init_op = tf.global_variables_initializer()
test_print_op = logging_ops.Print(0, [0], message="Test print success")
# We run the summaries in the same thread as the training operations by
# passing in None for summary_op to avoid a summary_thread being started.
# Running summaries and training operations in parallel could run out of
# GPU memory.
if is_chief:
local_init_op = opt.chief_init_op
else:
local_init_op = opt.local_step_init_op
local_init_opt = [local_init_op]
ready_for_local_init_op = opt.ready_for_local_init_op
sv = tf.train.Supervisor(
is_chief=is_chief,
local_init_op=local_init_op,
ready_for_local_init_op=ready_for_local_init_op,
logdir=FLAGS.train_dir,
init_op=init_op,
summary_op=None,
global_step=global_step,
saver=saver,
save_model_secs=FLAGS.save_interval_secs)
tf.logging.info('%s Supervisor' % datetime.now())
sess_config = tf.ConfigProto(
allow_soft_placement=True,
log_device_placement=FLAGS.log_device_placement)
# Get a session.
sess = sv.prepare_or_wait_for_session(target, config=sess_config)
# Start the queue runners.
queue_runners = tf.get_collection(tf.GraphKeys.QUEUE_RUNNERS)
sv.start_queue_runners(sess, queue_runners)
tf.logging.info('Started %d queues for processing input data.',
len(queue_runners))
if is_chief:
if not FLAGS.interval_method or FLAGS.worker_times_cdf_method:
sv.start_queue_runners(sess, chief_queue_runners)
sess.run(init_tokens_op)
# TIMEOUT client overseer.
# Even if not using timeout, we want to wait until all machines are ready.
timeout_client, timeout_server = launch_manager(sess, FLAGS)
# Train, checking for Nans. Concurrently run the summary operation at a
# specified interval. Note that the summary_op and train_op never run
# simultaneously in order to prevent running out of GPU memory.
next_summary_time = time.time() + FLAGS.save_summaries_secs
begin_time = time.time()
cur_iteration = -1
iterations_finished = set()
if FLAGS.task_id == 0 and FLAGS.interval_method:
opt.start_interval_updates(sess, timeout_client)
#the result of normal eqiation waited to be solved like min||Ax - b||^2
# b = np.ones(int(num_batches_per_epoch))
interval = np.arange(0, int(num_batches_per_epoch))
idx_list = np.random.choice(interval, int(num_workers), replace=False)
while not sv.should_stop():
sys.stdout.flush()
tf.logging.info("A new iteration...")
cur_iteration += 1
#sess.run([opt._wait_op], options=tf.RunOptions(timeout_in_ms=10000))
#sess.run([opt._wait_op])
#sess.run([test_print_op])
if FLAGS.worker_times_cdf_method:
sess.run([opt._wait_op])
timeout_client.broadcast_worker_dequeued_token(cur_iteration)
start_time = time.time()
feed_dict = mnist.fill_feed_dict(dataset, images, labels,
FLAGS.batch_size)
run_options = tf.RunOptions()
run_metadata = tf.RunMetadata()
#===============================================================================================
# interval_2 = np.arange(0, int(num_workers))
# workers_to_kill = np.random.choice(interval_2, FLAGS.num_worker_kill, replace=False)
LS_start_time = time.time()
interval_2 = np.arange(0, int(num_workers))
workers_to_kill = np.random.choice(interval_2,
FLAGS.num_worker_kill,
replace=False)
#interval_2 = np.arange(0, WORKER_NUM)
#workers_to_kill = np.random.choice(interval_2, NUM_WORKER_KILL, replace=False)
A = np.zeros((int(num_workers), int(num_batches_per_epoch)))
for i in range(A.shape[0]):
if i == A.shape[0] - 1:
A[i][idx_list[i]] = 1
A[i][idx_list[0]] = 1
else:
A[i][idx_list[i]] = 1
A[i][idx_list[i + 1]] = 1
for i in range(len(idx_list)):
element = idx_list[i]
if element == A.shape[1] - 1:
idx_list[i] = 0
else:
idx_list[i] += 1
for k in workers_to_kill:
A[k] = 0
A_for_calc = np.transpose(A)
# x = np.dot(np.linalg.pinv(A_for_calc), b)
# tf.logging.info("workers killed this iteration:")
# tf.logging.info(str(workers_to_kill))
# tf.logging.info("The matrix to solve:")
# for item in A_for_calc:
# tf.logging.info(str(item))
# tf.logging.info("Solution of LS:")
# tf.logging.info(str(x))
# LS_duration = time.time() - LS_start_time
# tf.logging.info("LS run time: %s" % str(LS_duration))
#===============================================================================================
if FLAGS.timeline_logging:
run_options.trace_level = tf.RunOptions.FULL_TRACE
run_options.output_partition_graphs = True
#timeout_ms = random.randint(300, 1200)
#tf.logging.info("SETTING TIMEOUT FOR %d ms" % timeout_ms)
#run_options.timeout_in_ms = 1000 * 60 * 1
# Increment current iteration
# Two more tiem in placeholder feed_dict
feed_dict[matrix_placeholder] = A_for_calc
tf.logging.info("RUNNING SESSION... %f" % time.time())
loss_value, step = sess.run([train_op, global_step],
feed_dict=feed_dict,
run_metadata=run_metadata,
options=run_options)
tf.logging.info("DONE RUNNING SESSION...")
if FLAGS.worker_times_cdf_method:
timeout_client.broadcast_worker_finished_computing_gradients(
cur_iteration)
assert not np.isnan(loss_value), 'Model diverged with loss = NaN'
# Log the elapsed time per iteration
finish_time = time.time()
# Create the Timeline object, and write it to a json
if FLAGS.timeline_logging:
tl = timeline.Timeline(run_metadata.step_stats)
ctf = tl.generate_chrome_trace_format()
with open(
'%s/worker=%d_timeline_iter=%d.json' %
(FLAGS.train_dir, FLAGS.task_id, step), 'w') as f:
f.write(ctf)
if step > FLAGS.max_steps:
break
duration = time.time() - start_time
examples_per_sec = FLAGS.batch_size / float(duration)
format_str = ('Worker %d: %s: step %d, loss = %f'
'(%.1f examples/sec; %.3f sec/batch)')
tf.logging.info(format_str %
(FLAGS.task_id, datetime.now(), step, loss_value,
examples_per_sec, duration))
# Determine if the summary_op should be run on the chief worker.
if is_chief and next_summary_time < time.time(
) and FLAGS.should_summarize:
tf.logging.info('Running Summary operation on the chief.')
summary_str = sess.run(summary_op)
sv.summary_computed(sess, summary_str)
tf.logging.info('Finished running Summary operation.')
# Determine the next time for running the summary.
next_summary_time += FLAGS.save_summaries_secs
if is_chief:
tf.logging.info('Elapsed Time: %f' % (time.time() - begin_time))
# Stop the supervisor. This also waits for service threads to finish.
sv.stop()
# Save after the training ends.
if is_chief:
saver.save(sess,
os.path.join(FLAGS.train_dir, 'model.ckpt'),
global_step=global_step)
def main(_):
# build a model
data = input_data.read_data_sets(FLAGS.data_dir, one_hot=False,
fake_data=False)
# make placeholders
images = tf.placeholder(tf.float32, [FLAGS.batch_size, mnist.IMAGE_PIXELS],
name='inputs')
labels = tf.placeholder(tf.int32, [FLAGS.batch_size], name='labels')
# build model up to inference
logits = mnist.inference(images, FLAGS.hidden_size, FLAGS.num_layers,
do_weightnorm=FLAGS.weightnorm,
do_batchnorm=FLAGS.batchnorm,
train=True)
if not FLAGS.batchnorm:
eval_logits = logits
else:
eval_logits = mnist.inference(images, FLAGS.hidden_size, FLAGS.num_layers,
do_weightnorm=FLAGS.weightnorm,
do_batchnorm=FLAGS.batchnorm,
train=False)
# get a loss function
loss = mnist.loss(logits, labels, 'train_xent')
eval_loss = mnist.loss(eval_logits, labels, 'eval_xent')
# add a sumary of this to track the training loss
# get training ops
train_op, gstep = mnist.training(loss, FLAGS.learning_rate, FLAGS.momentum)
# get an op to return precision on a batch
eval_op = mnist.evaluation(eval_logits, labels)
valid_var = tf.Variable(0, 'validation performance')
valid_summ = tf.scalar_summary('validation accuracy', valid_var)
# get summary op
summarise = tf.merge_all_summaries()
with tf.Session() as sess:
writer = tf.train.SummaryWriter(FLAGS.logdir, sess.graph_def)
tf.initialize_all_variables().run()
# do some training
print('nb: {} steps per epoch'.format(
data.train.num_examples // FLAGS.batch_size))
print('Step 0/{}.'.format(FLAGS.max_steps), end='')
for i in range(FLAGS.max_steps):
if (i+1) % 5 == 0:
# write summaries, check on validation set
if (i+1) % 100 == 0:
valid_perf = evaluate(sess,
data.validation,
logits,
[eval_op, eval_loss],
FLAGS.batch_size,
images,
labels,
gstep,
writer)
print()
summ_str, _, _ = sess.run([summarise, loss, train_op],
fill_feed(data.train, images, labels,
FLAGS.batch_size))
writer.add_summary(summ_str, gstep.eval(session=sess))
else:
# do a step of training
loss_val, _ = sess.run([loss, train_op],
fill_feed(data.train, images, labels,
FLAGS.batch_size))
print('\rStep {} (loss {})'.format(i+1, loss_val), end='', flush=True)
print('\n~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~')
print('Test evaluation:')
evaluate(sess, data.test, logits, [eval_op, eval_loss], FLAGS.batch_size, images, labels, gstep, writer)
print('\n~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~')
def train(dataset, testset):
"""Train on dataset for a number of steps."""
with tf.Graph().as_default(), tf.device('/gpu:0'):
# Create a variable to count the number of train() calls. This equals the
# number of batches processed * FLAGS.num_gpus.
global_step = tf.Variable(0, name="global_step", trainable=False)
# Calculate the learning rate schedule.
num_batches_per_epoch = (dataset.num_examples / FLAGS.batch_size)
decay_steps = int(num_batches_per_epoch * FLAGS.num_epochs_per_decay)
# Decay the learning rate exponentially based on the number of steps.
lr = tf.train.exponential_decay(FLAGS.initial_learning_rate,
global_step,
decay_steps,
FLAGS.learning_rate_decay_factor,
staircase=True)
# Create an optimizer that performs gradient descent.
opt = tf.train.AdamOptimizer(lr)
#fetch the data batch from training set
images, labels = mnist.placeholder_inputs(FLAGS.batch_size)
logits = mnist.inference(images)
#calc the loss and gradients
total_loss = mnist.loss(logits, labels)
grads = opt.compute_gradients(total_loss)
# Apply the gradients to adjust the shared variables.
apply_gradients_op = opt.apply_gradients(grads,
global_step=global_step)
with tf.control_dependencies([apply_gradients_op]):
train_op = tf.identity(total_loss, name='train_op')
# Create a saver.
saver = tf.train.Saver(tf.global_variables())
# Build the summary operation from the last tower summaries.
summary_op = tf.summary.merge_all()
# For testing trained model
test_size = testset.num_examples
test_images_placeholder, test_labels_placeholder = mnist.placeholder_inputs(
FLAGS.batch_size)
# logits_test = mnist.inference(test_images_placeholder, train=False)
#pred = mnist.predictions(logits_test)
validation_accuracy = tf.reduce_sum(mnist.evaluation(
logits, labels)) / tf.constant(FLAGS.batch_size)
# Build an initialization operation to run below.
init = tf.global_variables_initializer()
# Start running operations on the Graph. allow_soft_placement must be set to
# True to build towers on GPU, as some of the ops do not have GPU
# implementations.
sess = tf.Session(config=tf.ConfigProto(
allow_soft_placement=True,
log_device_placement=FLAGS.log_device_placement))
sess.run(init)
# Start the queue runners.
tf.train.start_queue_runners(sess=sess)
for step in range(FLAGS.max_steps):
feed_dict = mnist.fill_feed_dict(dataset, images, labels,
FLAGS.batch_size)
start_time = time.time()
_, loss_value, acc = sess.run(
[train_op, total_loss, validation_accuracy],
feed_dict=feed_dict)
# acc = sess.run(validation_accuracy, feed_dict=feed_dict_test)
duration = time.time() - start_time
assert not np.isnan(loss_value), 'Model diverged with loss = NaN'
examples_per_sec = FLAGS.batch_size / float(duration)
format_str = ('%s: step %d, loss = %.2f (%.1f examples/sec; %.3f '
'sec/batch); acc=%.4f')
tf.logging.info(format_str % (datetime.now(), step, loss_value,
examples_per_sec, duration, acc))
if step % 500 == 0 or (step + 1) == FLAGS.max_steps:
checkpoint_path = os.path.join(FLAGS.train_dir, 'model.ckpt')
saver.save(sess, checkpoint_path, global_step=step)
def run_training():
"""Train MNIST for a number of steps."""
# Tell TensorFlow that the model will be built into the default Graph.
with tf.Graph().as_default():
# Input images and labels.
images, labels = inputs(train=True,
batch_size=FLAGS.batch_size,
num_epochs=FLAGS.num_epochs)
# Build a Graph that computes predictions from the inference model.
logits = mnist.inference(images, FLAGS.hidden1, FLAGS.hidden2)
# Add to the Graph the loss calculation.
loss = mnist.loss(logits, labels)
print('---------------------------- ' + str(logits))
# Add to the Graph operations that train the model.
train_op = mnist.training(loss, FLAGS.learning_rate)
# The op for initializing the variables.
init_op = tf.group(tf.global_variables_initializer(),
tf.local_variables_initializer())
# Create a session for running operations in the Graph.
sess = tf.Session()
# Initialize the variables (the trained variables and the
# epoch counter).
sess.run(init_op)
# Start input enqueue threads.
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess=sess, coord=coord)
try:
step = 0
while step != 3:
start_time = time.time()
# Run one step of the model. The return values are
# the activations from the `train_op` (which is
# discarded) and the `loss` op. To inspect the values
# of your ops or variables, you may include them in
# the list passed to sess.run() and the value tensors
# will be returned in the tuple from the call.
_, loss_value = sess.run([train_op, loss])
#duration = time.time() - start_time
# Print an overview fairly often.
#if step % 100 == 0:
# print('Step %d: loss = %.2f (%.3f sec)' % (step, loss_value,
# duration))
step += 1
except tf.errors.OutOfRangeError:
print('Done training for %d epochs, %d steps.' %
(FLAGS.num_epochs, step))
finally:
# When done, ask the threads to stop.
coord.request_stop()
# Wait for threads to finish.
coord.join(threads)
sess.close()
def run_training():
"""Train MNIST for a number of steps."""
# Get the sets of images and labels for training, validation, and
# test on MNIST.
data_sets = input_data.read_data_sets(FLAGS.data_dir, FLAGS.fake_data)
# Tell TensorFlow that the model will be built into the default Graph.
with tf.Graph().as_default():
# Generate placeholders for the images and labels.
images_placeholder, labels_placeholder = placeholder_inputs(
FLAGS.batch_size)
# Build a Graph that computes predictions from the inference model.
logits = mnist.inference(images_placeholder)
# Add to the Graph the Ops for loss calculation.
loss = mnist.loss(logits, labels_placeholder)
# Add to the Graph the Ops that calculate and apply gradients.
train_op = mnist.training(loss, FLAGS.batch_size)
# Add the Op to compare the logits to the labels during evaluation.
eval_correct = mnist.evaluation(logits, labels_placeholder)
# Build the summary operation based on the TF collection of Summaries.
summary_op = tf.summary.merge_all()
# Create a saver for writing training checkpoints.
saver = tf.train.Saver()
# Create a session for running Ops on the Graph.
sess = tf.Session()
# Run the Op to initialize the variables.
init = tf.global_variables_initializer()
sess.run(init)
# Instantiate a SummaryWriter to output summaries and the Graph.
summary_writer = tf.summary.FileWriter(FLAGS.train_dir,
graph_def=sess.graph_def)
# And then after everything is built, start the training loop.
for step in range(FLAGS.max_steps):
start_time = time.time()
# Fill a feed dictionary with the actual set of images and labels
# for this particular training step.
feed_dict = fill_feed_dict(data_sets.train, images_placeholder,
labels_placeholder)
# Run one step of the model. The return values are the activations
# from the `train_op` (which is discarded) and the `loss` Op. To
# inspect the values of your Ops or variables, you may include them
# in the list passed to sess.run() and the value tensors will be
# returned in the tuple from the call.
_, loss_value = sess.run([train_op, loss], feed_dict=feed_dict)
duration = time.time() - start_time
# Write the summaries and print an overview fairly often.
if step % 100 == 0:
# Print status to stdout.
print('Step %d: loss = %.2f (%.3f sec)' %
(step, loss_value, duration))
# Update the events file.
summary_str = sess.run(summary_op, feed_dict=feed_dict)
summary_writer.add_summary(summary_str, step)
# Save a checkpoint and evaluate the model periodically.
if (step + 1) % 1000 == 0 or (step + 1) == FLAGS.max_steps:
saver.save(sess, FLAGS.train_dir, global_step=step)
# Evaluate against the training set.
print('Training Data Eval:')
do_eval(sess, eval_correct, images_placeholder,
labels_placeholder, data_sets.train)
# Evaluate against the validation set.
print('Validation Data Eval:')
do_eval(sess, eval_correct, images_placeholder,
labels_placeholder, data_sets.validation)
# Evaluate against the test set.
print('Test Data Eval:')
do_eval(sess, eval_correct, images_placeholder,
labels_placeholder, data_sets.test)
def run_training(settings: Settings) -> float:
tf.gfile.MakeDirs(settings.log_dir)
data_sets = copy.deepcopy(DATASETS)
with tf.Graph().as_default():
images_placeholder, labels_placeholder = placeholder_inputs(
settings.batch_size, )
logits = mnist.inference(
images_placeholder,
settings.hidden1,
settings.hidden2,
)
loss = mnist.loss(logits, labels_placeholder)
train_op = mnist.training(loss, settings.learning_rate)
eval_correct = mnist.evaluation(logits, labels_placeholder)
summary = tf.summary.merge_all()
init = tf.global_variables_initializer()
saver = tf.train.Saver()
sess = tf.Session()
summary_writer = tf.summary.FileWriter(settings.log_dir, sess.graph)
sess.run(init)
for step in range(settings.max_steps):
feed_dict = fill_feed_dict(
data_sets.train,
images_placeholder,
labels_placeholder,
settings,
)
_, loss_value = sess.run([train_op, loss], feed_dict=feed_dict)
# Write the summaries and print an overview fairly often.
if step % 100 == 0:
# Update the events file.
summary_str = sess.run(summary, feed_dict=feed_dict)
summary_writer.add_summary(summary_str, step)
summary_writer.flush()
# Save a checkpoint and evaluate the model periodically.
if (step + 1) % 1000 == 0 or (step + 1) == settings.max_steps:
checkpoint_file = os.path.join(settings.log_dir, 'model.ckpt')
saver.save(sess, checkpoint_file, global_step=step)
# Evaluate against the training set.
# print('Training Data Eval:')
do_eval(
sess,
eval_correct,
images_placeholder,
labels_placeholder,
data_sets.train,
settings,
)
# Evaluate against the validation set.
# print('Validation Data Eval:')
do_eval(
sess,
eval_correct,
images_placeholder,
labels_placeholder,
data_sets.validation,
settings,
)
# Evaluate against the test set.
# print('Test Data Eval:')
acc = do_eval(
sess,
eval_correct,
images_placeholder,
labels_placeholder,
data_sets.test,
settings,
)
return 1 - acc