def main(argv=None): # pylint: disable=unused-argument
cifar10.maybe_download_and_extract()
if tf.gfile.Exists(FLAGS.train_dir):
tf.gfile.DeleteRecursively(FLAGS.train_dir)
tf.gfile.MakeDirs(FLAGS.train_dir)
train()
cifar10_eval.evaluate()
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 = cifar10.distorted_inputs()
# Build a Graph that computes the logits predictions from the
# inference model.
logits = cifar10.inference(images)
# Calculate loss.
loss = cifar10.loss(logits, labels)
# Build a Graph that trains the model with one batch of examples and
# updates the model parameters.
train_op = cifar10.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()
self._start_time = current_time
for step in range(0, FLAGS.max_steps + 1, FLAGS.log_frequency):
print(str(step))
with tf.train.MonitoredTrainingSession(
checkpoint_dir=FLAGS.train_dir,
hooks=[
tf.train.StopAtStepHook(last_step=step),
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)
# evaluate test data
cifar10_eval.evaluate()
# evaluate train data
evaluate()
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))
if self._step % 1000 == 0:
cifar10_eval.evaluate()
def main(argv=None): # pylint: disable=unused-argument
total_time = time.time()
data_load_time = time.time()
cifar10.maybe_download_and_extract()
if tf.gfile.Exists(FLAGS.train_dir):
tf.gfile.DeleteRecursively(FLAGS.train_dir)
tf.gfile.MakeDirs(FLAGS.train_dir)
data_load_time = time.time() - data_load_time
train_time = time.time()
train()
train_time = time.time() - train_time
test_time = time.time()
cifar10_eval.evaluate()
test_time = time.time() - test_time
total_time = time.time() - total_time
print_time('Data load', data_load_time)
print_time('Train', train_time)
print_time('Test', test_time)
print_time('Total', total_time)
def main(argv=None): # pylint: disable=unused-argument
total_time = time.time()
data_load_time = time.time()
cifar10.maybe_download_and_extract()
if tf.gfile.Exists(FLAGS.train_dir):
tf.gfile.DeleteRecursively(FLAGS.train_dir)
tf.gfile.MakeDirs(FLAGS.train_dir)
data_load_time = time.time() - data_load_time
train_time = time.time()
train()
train_time = time.time() - train_time
test_time = time.time()
cifar10_eval.evaluate()
test_time = time.time() - test_time
total_time = time.time() - total_time
print_time('Data load', data_load_time)
print_time('Train', train_time)
print_time('Test', test_time)
print_time('Total', total_time)
def main():
# get all images in format of [r*1024,g*1024,b*1024]
# store them in dataset[0~9999]
save = open("check.csv", "w")
a = csv.writer(save)
f = open("test_batch.bin", "rb")
data = f.read()
for i in range(len(data) / 3073):
label.append(struct.unpack("B", data[i * 3073])[0])
datatmp = data[i * 3073 + 1:i * 3073 + 3073]
dataset.append(list(struct.unpack("B" * (len(datatmp)), datatmp[:])))
f.close()
# add error
for err_rate in range(1, 300, 10):
print(err_rate)
#for k in range(len(dataset)):
for k in range(100):
#show(dataset[k])
err.append(addError(dataset[k], err_rate, Gradient(dataset[k])))
#show(err[k])
# show image
# save back
out = open("/tmp/cifar10_data/cifar-10-batches-bin/random_batch_1.bin",
"wb")
#for j in range(len(dataset)):
for j in range(100):
back = [label[j]] + list(err[j])
back = np.array(back, np.uint8)
out.write(back)
out.close()
a.writerows([[str(err_rate), str(cifar10_eval.evaluate())]])
del err[:]
save.close()
def train():
"""Train CIFAR-10 for a number of steps."""
with tf.Graph().as_default():
global_step = tf.Variable(0, trainable=False)
# Get images and labels for CIFAR-10.
images, labels = cifar10.distorted_inputs()
# Build a Graph that computes the logits predictions from the
# inference model.
logits = cifar10.inference(images)
# Calculate loss.
loss = cifar10.loss(logits, labels)
# Build a Graph that trains the model with one batch of examples and
# updates the model parameters.
train_op = cifar10.train(loss, global_step)
# Create a saver.
saver = tf.train.Saver(tf.all_variables())
# Build the summary operation based on the TF collection of Summaries.
summary_op = tf.contrib.deprecated.merge_all_summaries()
# Build an initialization operation to run below.
init = tf.initialize_all_variables()
# Start running operations on the Graph.
sess = tf.Session(config=tf.ConfigProto(
log_device_placement=FLAGS.log_device_placement))
sess.run(init)
# Start the queue runners.
tf.train.start_queue_runners(sess=sess)
# summary_writer = tf.train.SummaryWriter(FLAGS.train_dir,
# graph_def=sess.graph_def)
for step in xrange(FLAGS.max_steps):
start_time = time.time()
_, loss_value = sess.run([train_op, loss])
duration = time.time() - start_time
assert not np.isnan(loss_value), 'Model diverged with loss = NaN'
if step % 10 == 0:
num_examples_per_step = FLAGS.batch_size
examples_per_sec = num_examples_per_step / duration
sec_per_batch = float(duration)
format_str = (
'%s: step %d, loss = %.2f (%.1f examples/sec; %.3f '
'sec/batch)')
print(format_str % (datetime.now(), step, loss_value,
examples_per_sec, sec_per_batch))
printfile(loss_value)
# Save the model checkpoint periodically.
if step % 1000 == 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)
if step % 1000 == 0:
cifar10_eval.evaluate()
def train():
"""Train CIFAR-10 for a number of steps."""
with tf.Graph().as_default(), tf.device('/cpu:0'):
# Create a variable to count the number of train() calls. This equals the
# number of batches processed * FLAGS.num_gpus.
global_step = tf.get_variable(
'global_step', [],
initializer=tf.constant_initializer(0), trainable=False)
# Calculate the learning rate schedule.
num_batches_per_epoch = (cifar10.NUM_EXAMPLES_PER_EPOCH_FOR_TRAIN /
(FLAGS.batch_size * FLAGS.num_gpus))
decay_steps = int(num_batches_per_epoch * FLAGS.lr_decay_epochs)
# Decay the learning rate exponentially based on the number of steps.
lr = tf.train.exponential_decay(FLAGS.learning_rate,
global_step,
decay_steps,
cifar10.LEARNING_RATE_DECAY_FACTOR,
staircase=True)
# Create an optimizer that performs gradient descent.
opt = tf.train.AdamOptimizer(lr)
# Get images and labels for CIFAR-10.
images, labels = cifar10.distorted_inputs()
batch_queue = tf.contrib.slim.prefetch_queue.prefetch_queue(
[images, labels], capacity=2 * FLAGS.num_gpus)
# Calculate the gradients for each model tower.
tower_grads = []
with tf.variable_scope(tf.get_variable_scope()):
for i in xrange(FLAGS.num_gpus):
with tf.device('/gpu:%d' % i):
with tf.name_scope('%s_%d' % (cifar10.TOWER_NAME, i)) as scope:
# Dequeues one batch for the GPU
image_batch, label_batch = batch_queue.dequeue()
# Calculate the loss for one tower of the CIFAR model. This function
# constructs the entire CIFAR model but shares the variables across
# all towers.
loss = tower_loss(scope, image_batch, label_batch)
# Reuse variables for the next tower.
tf.get_variable_scope().reuse_variables()
# Retain the summaries from the final tower.
summaries = tf.get_collection(tf.GraphKeys.SUMMARIES, scope)
#Added for BN - 25.7.17 Oran
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(update_ops):
grads = opt.compute_gradients(loss)
tower_grads.append(grads)
# We must calculate the mean of each gradient. Note that this is the
# synchronization point across all towers.
grads = average_gradients(tower_grads)
# Add a summary to track the learning rate.
summaries.append(tf.summary.scalar('learning_rate', lr))
# Add histograms for gradients.
for grad, var in grads:
if grad is not None:
summaries.append(tf.summary.histogram(var.op.name + '/gradients', grad))
# Apply the gradients to adjust the shared variables.
apply_gradient_op = opt.apply_gradients(grads, global_step=global_step)
# Add histograms for trainable variables.
for var in tf.trainable_variables():
summaries.append(tf.summary.histogram(var.op.name, var))
# Track the moving averages of all trainable variables.
variable_averages = tf.train.ExponentialMovingAverage(
cifar10.MOVING_AVERAGE_DECAY, global_step)
variables_averages_op = variable_averages.apply(tf.trainable_variables())
# Group all updates to into a single train op.
train_op = tf.group(apply_gradient_op, variables_averages_op)
# train_op = tf.group(variables_averages_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(summaries)
# 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)
summary_writer = tf.summary.FileWriter(FLAGS.train_dir, sess.graph)
step_reached = -1
max_steps = int(FLAGS.epochs * cifar10.NUM_EXAMPLES_PER_EPOCH_FOR_TRAIN / (FLAGS.batch_size * FLAGS.num_gpus))
best_precision = 0
f = open(FLAGS.train_dir + '/summary.txt', 'a')
# Load model if not a new run
if FLAGS.new_run==False:
ckpt = tf.train.get_checkpoint_state(FLAGS.train_dir)
if ckpt and ckpt.model_checkpoint_path:
saver.restore(sess, ckpt.model_checkpoint_path)
step_reached = ckpt.model_checkpoint_path.split('/')[-1].split('-')[-1]
for step in xrange(int(step_reached)+1,max_steps):
start_time = time.time()
_, loss_value = sess.run([train_op, loss])
duration = time.time() - start_time
assert not np.isnan(loss_value), 'Model diverged with loss = NaN'
if step % 1000 == 0:
num_examples_per_step = FLAGS.batch_size * FLAGS.num_gpus
examples_per_sec = num_examples_per_step / duration
sec_per_batch = duration / FLAGS.num_gpus
format_str = ('%s: step %d, loss = %.2f (%.1f examples/sec; %.3f '
'sec/batch)')
print (format_str % (datetime.now(), step, loss_value,
examples_per_sec, sec_per_batch))
if step % 1000 == 0:
summary_str = sess.run(summary_op)
summary_writer.add_summary(summary_str, step)
# Save the model checkpoint periodically.
if step % 1000 == 0 or (step + 1) == max_steps:
checkpoint_path = os.path.join(FLAGS.train_dir, 'model.ckpt')
saver.save(sess, checkpoint_path, global_step=step)
if (step % 4000 == 0) and (step > 28000):
precision = []
for i in xrange(6):
cifar10.draw_weights(sess)
precision += [cifar10_eval.evaluate()]
if precision[i]>best_precision:
best_precision = precision[i]
os.system('cp '+FLAGS.train_dir+'/weights/* '+FLAGS.train_dir+'/best_weights/')
print('Average precision: '+str(round(np.mean(precision),3)))
f.write('step: '+str(step)+', average precision = '+str(round(np.mean(precision),3))+'\n')
# if step % 1 == 0: # Check CLT assumption
# a1_, b1_, image_batch_, conv3_1_ = sess.run([a, b, image_batch, conv3_1])
# # LAYER 1
# x = conv3_1_[0,2:5,2:5,:]
# activations = []
# for i in range(700):
# W = cifar10.draw_ternary_weight(a1_[:,:,:,10:12], b1_[:,:,:,10:12])
# W = W[:,:,:,0]
# activations += [np.sum(x*W)]
# activations = np.squeeze(np.array(activations))
#
# mu_ = a1_[:,:,:,10] - b1_[:,:,:,10]
# mu_bar = np.sum(x*mu_)
# sigma = a1_[:,:,:,10] + b1_[:,:,:,10] - mu_*mu_ + 0.001
# sigma_bar = np.sum(x*x*sigma)
# samples = np.random.normal(mu_bar, np.sqrt(sigma_bar), 700)
#
# max_=max(max(activations),max(samples))
# min_=min(min(activations),min(samples))
# plt.figure(1)
# plt.subplot(211)
# p_activations,_,_ = plt.hist(activations, bins=70, range=[min_,max_])
# plt.title('Step ' +str(step)+';Layer 1 Activations: mu='+str(round(np.mean(activations),3))+'; Sigma='+str(round(np.var(activations),3)))
# plt.subplot(212)
# g_activations,_,_ = plt.hist(samples, bins=70, range=[min_,max_])
# m = 0.5*(p_activations + g_activations)
# js = 0.5*scipy.stats.entropy(p_activations, m) + 0.5*scipy.stats.entropy(g_activations, m)
## jsd1+=[js]
## steps1+=[step]
# plt.title('Gaussian; mu='+str(round(mu_bar,3))+'; Sigma='+str(round(sigma_bar,3))+'; JS Divergence='+str(round(js,3)))
# plt.show()
# plt.savefig(FLAGS.train_dir + '/activations/' + str(step)+'.png')
# plt.close(1)
f.write('best precision = '+str(round(best_precision, 3))+'\n')
f.close()
def train():
"""Train CIFAR-10 for a number of steps."""
with tf.Graph().as_default(), tf.device('/cpu:0'):
# Create a variable to count the number of train() calls. This equals the
# number of batches processed * FLAGS.num_gpus.
global_step = tf.get_variable('global_step', [],
initializer=tf.constant_initializer(0),
trainable=False)
# Calculate the learning rate schedule.
num_batches_per_epoch = (cifar10.NUM_EXAMPLES_PER_EPOCH_FOR_TRAIN /
FLAGS.batch_size)
decay_steps = int(num_batches_per_epoch * cifar10.NUM_EPOCHS_PER_DECAY)
# Decay the learning rate exponentially based on the number of steps.
lr = tf.train.exponential_decay(cifar10.INITIAL_LEARNING_RATE,
global_step,
decay_steps,
cifar10.LEARNING_RATE_DECAY_FACTOR,
staircase=True)
# Create an optimizer that performs gradient descent.
opt = tf.train.MomentumOptimizer(lr, momentum=0.9, use_nesterov=False)
# opt = tf.train.AdamOptimizer()
# Calculate the gradients for each model tower.
tower_grads = []
with tf.variable_scope(tf.get_variable_scope()):
for i in xrange(FLAGS.num_gpus):
with tf.device('/gpu:%d' % i):
with tf.name_scope('%s_%d' %
(cifar10.TOWER_NAME, i)) as scope:
# Calculate the loss for one tower of the CIFAR model. This function
# constructs the entire CIFAR model but shares the variables across
# all towers.
loss = tower_loss(scope)
# Reuse variables for the next tower.
tf.get_variable_scope().reuse_variables()
# Retain the summaries from the final tower.
summaries = tf.get_collection(tf.GraphKeys.SUMMARIES,
scope)
# Calculate the gradients for the batch of data on this CIFAR tower.
grads = opt.compute_gradients(loss)
# Keep track of the gradients across all towers.
tower_grads.append(grads)
# We must calculate the mean of each gradient. Note that this is the
# synchronization point across all towers.
grads = average_gradients(tower_grads)
# Add a summary to track the learning rate.
summaries.append(tf.summary.scalar('learning_rate', lr))
# Add histograms for gradients.
for grad, var in grads:
if grad is not None:
summaries.append(
tf.summary.histogram(var.op.name + '/gradients', grad))
# Apply the gradients to adjust the shared variables.
# pdb.set_trace()
apply_gradient_op = opt.apply_gradients(grads, global_step=global_step)
# Add histograms for trainable variables.
for var in tf.trainable_variables():
summaries.append(tf.summary.histogram(var.op.name, var))
# Track the moving averages of all trainable variables.
variable_averages = tf.train.ExponentialMovingAverage(
cifar10.MOVING_AVERAGE_DECAY, global_step)
variables_averages_op = variable_averages.apply(
tf.trainable_variables())
# Group all updates to into a single train op.
train_op = tf.group(apply_gradient_op, variables_averages_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(summaries)
# Prune the network iterately.s
#pruning_ops = pruning(k = 1.25, scale=0.02)
# 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)
if FLAGS.reload:
ckpt = tf.train.get_checkpoint_state(FLAGS.train_dir)
if ckpt and ckpt.model_checkpoint_path:
# Restores from checkpoint
saver.restore(sess, ckpt.model_checkpoint_path)
# Assuming model_checkpoint_path looks something like:
# /my-favorite-path/cifar10_train/model.ckpt-0,
# extract global_step from it.
step = int(
ckpt.model_checkpoint_path.split('/')[-1].split('-')[-1])
else:
step = 1
else:
tf.gfile.DeleteRecursively(FLAGS.train_dir)
tf.gfile.MakeDirs(FLAGS.train_dir)
step = 1
summary_writer = tf.summary.FileWriter(FLAGS.train_dir, sess.graph)
while step <= FLAGS.max_steps:
start_time = time.time()
_, loss_value = sess.run([train_op, loss])
duration = time.time() - start_time
assert not np.isnan(loss_value), 'Model diverged with loss = NaN'
if step % 10 == 0:
num_examples_per_step = FLAGS.batch_size * FLAGS.num_gpus
examples_per_sec = num_examples_per_step / duration
sec_per_batch = duration / FLAGS.num_gpus
format_str = (
'%s: step %d, loss = %.2f (%.1f examples/sec; %.3f '
'sec/batch)')
print(format_str % (datetime.now(), step, loss_value,
examples_per_sec, sec_per_batch))
if step % 100 == 0:
summary_str = sess.run(summary_op)
summary_writer.add_summary(summary_str, step)
#pruning
if step % 3000 == 0:
# Init precision and compress
checkpoint_path = os.path.join(FLAGS.train_dir, 'model.ckpt')
saver.save(sess, checkpoint_path, global_step=step)
# Evalute precision and compress rate
initPrecision, initCompress = cifar10_eval.evaluate()
print("initPrecision: %.4f , initCompress: %.4f" % \
(initPrecision,initCompress))
scaleFounder = float(0.01)
while True:
# Prune the network iterately.s
pruning_ops = pruning(k=1.5, scale=scaleFounder)
sess.run(pruning_ops)
# Save the model for eval
# tip: no need for packup because pruning scale is monotonic increasing
checkpoint_path = os.path.join(FLAGS.train_dir,
'model.ckpt')
saver.save(sess, checkpoint_path, global_step=step)
# Evalute precision and compress rate
precision, compress = cifar10_eval.evaluate()
print("precision: %.4f , compress: %.4f" %
(precision, compress))
scaleFounder = scaleFounder + 0.01
if(
(FLAGS.precision_fact * (initPrecision - precision) -
FLAGS.compress_fact * (initCompress - compress) > 0 )or \
scaleFounder == 1.0
):
break
# Save the model checkpoint periodically.
if step % 500 == 0 or step == FLAGS.max_steps:
checkpoint_path = os.path.join(FLAGS.train_dir, 'model.ckpt')
saver.save(sess, checkpoint_path, global_step=step)
#cifar10_eval.evaluate()
step += 1
