def eval_one_epoch(args, sess, dataset, image_paths_placeholder,
labels_placeholder, is_training_placeholder, enqueue_op,
clones):
batch_size = args.batch_size * args.num_gpus
image_paths, num_per_class = all_val_entities(args, dataset)
print('eval image paths', len(image_paths))
nrof_origin_samples = len(image_paths)
assert (sum(num_per_class) == nrof_origin_samples)
#print(num_per_class)
#print(image_paths[0:10])
assert (args.batch_size % 3 == 0)
triplet_size = args.batch_size // 3
_a = int(math.ceil(len(image_paths) / batch_size))
nrof_samples = _a * batch_size
while nrof_samples > len(image_paths):
image_paths.append(image_paths[0])
#for p in image_paths:
# print(p)
#print('Running forward pass on sampled images: ', end='')
start_time = time.time()
nrof_examples = len(image_paths)
assert (nrof_examples % batch_size == 0)
labels_array = np.reshape(np.arange(nrof_examples), (-1, 3))
image_paths_array = np.reshape(np.expand_dims(np.array(image_paths), 1),
(-1, 3))
print(image_paths_array.shape)
print(labels_array.shape)
sess.run(
enqueue_op, {
image_paths_placeholder: image_paths_array,
labels_placeholder: labels_array
})
emb_array = np.zeros((nrof_examples, args.embedding_size),
dtype=np.float32)
nrof_batches = int(np.ceil(nrof_examples / batch_size))
print('eval batches', nrof_batches)
for i in xrange(nrof_batches):
if i % 10 == 0:
prt('running eval batch %d' % i)
ops = []
for clone in clones:
with tf.device(clone.device):
embeddings, labels, _ = clone.outputs
ops += [embeddings, labels]
ops_value = sess.run(ops, feed_dict={is_training_placeholder: False})
for k in xrange(args.num_gpus):
emb = ops_value[k * 2]
#prt(emb.shape)
lab = ops_value[k * 2 + 1]
#prt(lab.shape)
emb_array[lab, :] = emb
sys.stdout.flush()
print('%.3f' % (time.time() - start_time))
emb_array = emb_array[0:nrof_origin_samples, :]
score = top1_recall(emb_array, num_per_class)
print('top1 recall: %f' % score)
def do_training(train_op, init_fn=None, summary_op=None, lr=None):
global savers
graph = ops.get_default_graph()
with graph.as_default():
global_step = variables.get_or_create_global_step()
saver = tf_saver.Saver(max_to_keep=0)
with ops.name_scope('init_ops'):
init_op = tf_variables.global_variables_initializer()
ready_op = tf_variables.report_uninitialized_variables()
local_init_op = control_flow_ops.group(
tf_variables.local_variables_initializer(),
data_flow_ops.tables_initializer())
summary_writer = supervisor.Supervisor.USE_DEFAULT
with ops.name_scope('train_step'):
train_step_kwargs = {}
if not FLAGS.max_number_of_steps is None:
should_stop_op = math_ops.greater_equal(
global_step, FLAGS.max_number_of_steps)
else:
should_stop_op = constant_op.constant(False)
train_step_kwargs['should_stop'] = should_stop_op
if FLAGS.log_every_n_steps > 0:
train_step_kwargs['should_log'] = math_ops.equal(
math_ops.mod(global_step, FLAGS.log_every_n_steps), 0)
prefix = "loc/net"
lp = len(prefix)
vdic = {
"InceptionV2" + v.op.name[lp:]: v
for v in tf.trainable_variables()
if v.name.startswith(prefix) and v.name.find("Logits/") < 0
}
_saver = tf_saver.Saver(vdic)
savers.append(_saver)
for i in xrange(NUM_STN):
prefix = "stn%d/net" % i
lp = len(prefix)
vdic = {
"InceptionV2" + v.op.name[lp:]: v
for v in tf.trainable_variables()
if v.name.startswith(prefix) and v.name.find("Logits/") < 0
}
# saver = tf.train.Saver(vdic)
_saver = tf_saver.Saver(vdic)
savers.append(_saver)
prt("savers %d" % len(savers))
is_chief = True
logdir = FLAGS.train_dir
sv = supervisor.Supervisor(graph=graph,
is_chief=is_chief,
logdir=logdir,
init_op=init_op,
init_feed_dict=None,
local_init_op=local_init_op,
ready_for_local_init_op=None,
ready_op=ready_op,
summary_op=summary_op,
summary_writer=summary_writer,
global_step=global_step,
saver=saver,
save_summaries_secs=FLAGS.save_summaries_secs,
save_model_secs=FLAGS.save_interval_secs,
init_fn=init_fn)
if summary_writer is not None:
train_step_kwargs['summary_writer'] = sv.summary_writer
with sv.managed_session('', start_standard_services=False,
config=None) as sess:
logging.info('Starting Session.')
if is_chief:
if logdir:
sv.start_standard_services(sess)
elif startup_delay_steps > 0:
_wait_for_step(
sess, global_step,
min(startup_delay_steps, number_of_steps or sys.maxint))
sv.start_queue_runners(sess)
logging.info('Starting Queues.')
try:
while not sv.should_stop():
total_loss, global_step_value, should_stop = train_step(
sess, train_op, global_step, lr, train_step_kwargs)
current_epoch = int(
math.ceil(float(global_step_value) / FLAGS.steps_in_epoch))
if global_step_value > 0 and global_step_value % FLAGS.save_every_n_steps == 0:
sv.saver.save(sess,
sv.save_path,
global_step=sv.global_step)
if should_stop:
logging.info('Stopping Training.')
break
except errors.OutOfRangeError:
# OutOfRangeError is thrown when epoch limit per
# tf.train.limit_epochs is reached.
logging.info('Caught OutOfRangeError. Stopping Training.')
if logdir and sv.is_chief:
logging.info('Finished training! Saving model to disk.')
sv.saver.save(sess, sv.save_path, global_step=sv.global_step)
def train_step(sess, train_op, global_step, lr, train_step_kwargs):
"""Function that takes a gradient step and specifies whether to stop.
Args:
sess: The current session.
train_op: An `Operation` that evaluates the gradients and returns the
total loss.
global_step: A `Tensor` representing the global training step.
train_step_kwargs: A dictionary of keyword arguments.
Returns:
The total loss and a boolean indicating whether or not to stop training.
Raises:
ValueError: if 'should_trace' is in `train_step_kwargs` but `logdir` is not.
"""
start_time = time.time()
trace_run_options = None
run_metadata = None
if 'should_trace' in train_step_kwargs:
if 'logdir' not in train_step_kwargs:
raise ValueError(
'logdir must be present in train_step_kwargs when '
'should_trace is present')
if sess.run(train_step_kwargs['should_trace']):
trace_run_options = config_pb2.RunOptions(
trace_level=config_pb2.RunOptions.FULL_TRACE)
run_metadata = config_pb2.RunMetadata()
total_loss, lr_value, np_global_step = sess.run(
[train_op, lr, global_step],
options=trace_run_options,
run_metadata=run_metadata)
time_elapsed = time.time() - start_time
if run_metadata is not None:
tl = timeline.Timeline(run_metadata.step_stats)
trace = tl.generate_chrome_trace_format()
trace_filename = os.path.join(train_step_kwargs['logdir'],
'tf_trace-%d.json' % np_global_step)
logging.info('Writing trace to %s', trace_filename)
file_io.write_string_to_file(trace_filename, trace)
if 'summary_writer' in train_step_kwargs:
train_step_kwargs['summary_writer'].add_run_metadata(
run_metadata, 'run_metadata-%d' % np_global_step)
if 'should_log' in train_step_kwargs:
if sess.run(train_step_kwargs['should_log']):
logging.info('global step %d: loss = %.4f (%.3f sec/step)',
np_global_step, total_loss, time_elapsed)
prt('global step %d with lr %.4f: loss = %.4f (%.3f sec/step)' %
(np_global_step, lr_value, total_loss, time_elapsed))
# TODO(nsilberman): figure out why we can't put this into sess.run. The
# issue right now is that the stop check depends on the global step. The
# increment of global step often happens via the train op, which used
# created using optimizer.apply_gradients.
#
# Since running `train_op` causes the global step to be incremented, one
# would expected that using a control dependency would allow the
# should_stop check to be run in the same session.run call:
#
# with ops.control_dependencies([train_op]):
# should_stop_op = ...
#
# However, this actually seems not to work on certain platforms.
if 'should_stop' in train_step_kwargs:
should_stop = sess.run(train_step_kwargs['should_stop'])
else:
should_stop = False
return total_loss, np_global_step, should_stop
from tensorflow.python.ops import control_flow_ops
from tensorflow.python.ops import data_flow_ops
from tensorflow.python.ops import math_ops
from tensorflow.python.ops import variables as tf_variables
from tensorflow.python.platform import tf_logging as logging
from tensorflow.python.summary import summary
from tensorflow.python.training import optimizer as tf_optimizer
from tensorflow.python.training import saver as tf_saver
from tensorflow.python.training import supervisor
from tensorflow.python.training import sync_replicas_optimizer
from tensorflow.python.training import training_util
slim = tf.contrib.slim
cuda_devices = os.environ['CUDA_VISIBLE_DEVICES']
NUM_GPUS = len(cuda_devices.split(','))
prt("NUM_GPUS %d" % NUM_GPUS)
NUM_CLASSES = 120
NUM_ATTRIBS = 10654
BATCH_PER_GPU = 16
# assert BATCH_SIZE%NUM_GPUS==0
SAVE_EVERY_N_EPOCH = 2
DEFAULT_IMAGE_SIZE = 448
IMAGE_SIZE = DEFAULT_IMAGE_SIZE
if len(sys.argv) > 1:
IMAGE_SIZE = int(sys.argv[1])
STN_OUT_SIZE = 224
prt("IMAGE_SIZE %d" % IMAGE_SIZE)
INIT_LR = 0.01
LOC_LR = 0.00001
def train_one_epoch(args, sess, dataset, image_paths_placeholder,
labels_placeholder, is_training_placeholder, enqueue_op,
input_queue, clones, loss, train_op, summary_op,
summary_writer):
global_step = variables.get_or_create_global_step()
step = sess.run(global_step, feed_dict=None)
epoch = step // args.epoch_size
batch_number = 0
lr = args.learning_rate
batch_size = args.batch_size * args.num_gpus
while batch_number < args.epoch_size:
# Sample people randomly from the dataset
prt('start to sample entities')
image_paths, num_per_class = sample_entities(args, dataset)
#print(num_per_class[0:5])
#prt(len(image_paths))
#print(num_per_class)
#print(image_paths[0:10])
#print('Running forward pass on sampled images: ', end='')
start_time = time.time()
nrof_examples = len(image_paths)
assert (nrof_examples % batch_size == 0)
labels_array = np.reshape(np.arange(nrof_examples), (-1, 3))
image_paths_array = np.reshape(
np.expand_dims(np.array(image_paths), 1), (-1, 3))
#print(image_paths_array.shape)
#print(labels_array.shape)
sess.run(
enqueue_op, {
image_paths_placeholder: image_paths_array,
labels_placeholder: labels_array
})
emb_array = np.zeros((nrof_examples, args.embedding_size))
nrof_batches = int(np.ceil(nrof_examples / args.batch_size))
embeddings = clones[0].outputs[0]
label_batch = clones[0].outputs[1]
#print(nrof_batches)
for i in xrange(nrof_batches):
emb, lab = sess.run([embeddings, label_batch],
feed_dict={is_training_placeholder: True})
emb_array[lab, :] = emb
print('time for fetching all embedding %.3f' %
(time.time() - start_time))
#print(emb_array[0:5,0:5])
# Select triplets based on the embeddings
print('Selecting suitable triplets for training')
triplets, triplets_info = select_triplets(args, emb_array,
num_per_class, image_paths)
selection_time = time.time() - start_time
print(
'(nrof_random_negs, nrof_triplets) = (%d, %d): time=%.3f seconds' %
(0, len(triplets), selection_time))
assert len(triplets) > 0
#post-processing
assert (args.batch_size % 3 == 0)
triplet_size = batch_size // 3
_a = len(triplets) // triplet_size
nrof_triplets = _a * triplet_size
triplets = triplets[0:nrof_triplets]
#post-processing finish
# Perform training on the selected triplets
triplet_paths = list(itertools.chain(*triplets))
nrof_batches = int(np.ceil(nrof_triplets * 3 / batch_size))
labels_array = np.reshape(np.arange(len(triplet_paths)), (-1, 3))
triplet_paths_array = np.reshape(
np.expand_dims(np.array(triplet_paths), 1), (-1, 3))
sess.run(
enqueue_op, {
image_paths_placeholder: triplet_paths_array,
labels_placeholder: labels_array
})
nrof_examples = len(triplet_paths)
train_time = 0
i = 0
emb_array = np.zeros((nrof_examples, args.embedding_size))
#loss_array = np.zeros((nrof_triplets,))
prt('nrof_batches: %d' % nrof_batches)
while i < nrof_batches:
start_time = time.time()
#err, _, step, emb, lab = sess.run([loss, train_op, global_step, embeddings, labels_batch], feed_dict={is_training_placeholder:True})
#emb_array[lab,:] = emb
#loss_array[i] = err
err, _, step = sess.run([loss, train_op, global_step],
feed_dict={is_training_placeholder: True})
duration = time.time() - start_time
prt('Epoch: [%d][%d@%d/%d]\tTime %.3f\tLoss %2.3f' %
(epoch, i, batch_number + 1, args.epoch_size, duration, err))
batch_number += 1
i += 1
train_time += duration
prt('one sample finish')
# Add validation loss and accuracy to summary
summary = tf.Summary()
#pylint: disable=maybe-no-member
summary.value.add(tag='time/selection', simple_value=selection_time)
summary_writer.add_summary(summary, step)
return step
def main():
print(args)
prt('')
subdir = datetime.strftime(datetime.now(), '%Y%m%d-%H%M%S')
log_dir = os.path.join(os.path.expanduser(args.logs_base_dir), subdir)
if not os.path.isdir(
log_dir): # Create the log directory if it doesn't exist
os.makedirs(log_dir)
model_dir = os.path.join(os.path.expanduser(args.models_base_dir), subdir)
if not os.path.isdir(
model_dir): # Create the model directory if it doesn't exist
os.makedirs(model_dir)
# Store some git revision info in a text file in the log directory
src_path, _ = os.path.split(os.path.realpath(__file__))
np.random.seed(seed=args.seed)
print('Model directory: %s' % model_dir)
print('Log directory: %s' % log_dir)
if args.pretrained_model:
print('Pre-trained model: %s' %
os.path.expanduser(args.pretrained_model))
with tf.Graph().as_default():
deploy_config = model_deploy.DeploymentConfig(num_clones=args.num_gpus,
clone_on_cpu=False)
tf.set_random_seed(args.seed)
#global_step = tf.Variable(0, trainable=False)
global_step = variables.get_or_create_global_step()
# Placeholder for the learning rate
#learning_rate_placeholder = tf.placeholder(tf.float32, name='learning_rate')
#batch_size_placeholder = tf.placeholder(tf.int32, name='batch_size')
with tf.device('/cpu:0'):
is_training_placeholder = tf.placeholder(tf.bool,
name='is_training')
image_paths_placeholder = tf.placeholder(tf.string,
shape=(None, 3),
name='image_paths')
labels_placeholder = tf.placeholder(tf.int64,
shape=(None, 3),
name='labels')
input_queue = data_flow_ops.FIFOQueue(capacity=100000,
dtypes=[tf.string, tf.int64],
shapes=[(3, ), (3, )],
shared_name=None,
name=None)
enqueue_op = input_queue.enqueue_many(
[image_paths_placeholder, labels_placeholder])
nrof_preprocess_threads = 8
images_and_labels = []
for _ in range(nrof_preprocess_threads):
filenames, label = input_queue.dequeue()
#filenames = tf.Print(filenames, [tf.shape(filenames)], 'filenames shape:')
images = []
for filename in tf.unstack(filenames):
#filename = tf.Print(filename, [filename], 'filename = ')
file_contents = tf.read_file(filename)
image = tf.image.decode_jpeg(file_contents)
#image = tf.Print(image, [tf.shape(image)], 'data count = ')
if image.dtype != tf.float32:
image = tf.image.convert_image_dtype(image,
dtype=tf.float32)
if args.random_crop:
#image = tf.random_crop(image, [args.image_size, args.image_size, 3])
bbox = tf.constant([0.0, 0.0, 1.0, 1.0],
dtype=tf.float32,
shape=[1, 1, 4])
sample_distorted_bounding_box = tf.image.sample_distorted_bounding_box(
tf.shape(image),
bounding_boxes=bbox,
area_range=(0.7, 1.0),
use_image_if_no_bounding_boxes=True)
bbox_begin, bbox_size, distort_bbox = sample_distorted_bounding_box
image = tf.slice(image, bbox_begin, bbox_size)
#else:
# image = tf.image.resize_image_with_crop_or_pad(image, args.image_size, args.image_size)
image = tf.expand_dims(image, 0)
image = tf.image.resize_bilinear(
image, [args.image_size, args.image_size],
align_corners=False)
image = tf.squeeze(image, [0])
if args.random_flip:
image = tf.image.random_flip_left_right(image)
image.set_shape((args.image_size, args.image_size, 3))
##pylint: disable=no-member
image = tf.subtract(image, 0.5)
image = tf.multiply(image, 2.0)
#image = tf.Print(image, [tf.shape(image)], 'data count = ')
images.append(image)
#images.append(tf.image.per_image_standardization(image))
images_and_labels.append([images, label])
learning_rate = get_learning_rate(args)
opt = get_optimizer(args, learning_rate)
image_batch, label_batch = tf.train.batch_join(
images_and_labels,
batch_size=args.batch_size,
shapes=[(args.image_size, args.image_size, 3), ()],
enqueue_many=True,
capacity=4 * nrof_preprocess_threads * args.batch_size,
allow_smaller_final_batch=False)
batch_queue = slim.prefetch_queue.prefetch_queue(
[image_batch, label_batch], capacity=9000)
def clone_fn(_batch_queue):
_image_batch, _label_batch = _batch_queue.dequeue()
embeddings = image_to_embedding(_image_batch,
is_training_placeholder, args)
# Split embeddings into anchor, positive and negative and calculate triplet loss
anchor, positive, negative = tf.unstack(
tf.reshape(embeddings, [-1, 3, args.embedding_size]), 3, 1)
triplet_loss = triplet_loss_fn(anchor, positive, negative,
args.alpha)
tf.losses.add_loss(triplet_loss)
#tf.summary.scalar('learning_rate', learning_rate)
return embeddings, _label_batch, triplet_loss
clones = model_deploy.create_clones(deploy_config, clone_fn,
[batch_queue])
first_clone = clones[0]
triplet_loss = first_clone.outputs[2]
embeddings = first_clone.outputs[0]
_label_batch = first_clone.outputs[1]
#embedding_clones = model_deploy.create_clones(deploy_config, embedding_fn, [batch_queue])
#first_clone_scope = deploy_config.clone_scope(0)
#update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS, first_clone_scope)
update_ops = []
vdic = [
v for v in tf.trainable_variables() if v.name.find("Logits/") < 0
]
pretrained_saver = tf.train.Saver(vdic)
saver = tf.train.Saver(max_to_keep=3)
# Build the summary operation based on the TF collection of Summaries.
summary_op = tf.summary.merge_all()
with tf.device(deploy_config.optimizer_device()):
learning_rate = get_learning_rate(args)
opt = get_optimizer(args, learning_rate)
total_loss, clones_gradients = model_deploy.optimize_clones(
clones, opt, var_list=tf.trainable_variables())
grad_updates = opt.apply_gradients(clones_gradients,
global_step=global_step)
update_ops.append(grad_updates)
update_op = tf.group(*update_ops)
train_op = control_flow_ops.with_dependencies([update_op],
total_loss,
name='train_op')
vdic = [
v for v in tf.trainable_variables() if v.name.find("Logits/") < 0
]
pretrained_saver = tf.train.Saver(vdic)
saver = tf.train.Saver(max_to_keep=3)
# Build the summary operation based on the TF collection of Summaries.
summary_op = tf.summary.merge_all()
# Start running operations on the Graph.
#gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=args.gpu_memory_fraction)
#sess = tf.Session(config=tf.ConfigProto(gpu_options=gpu_options))
sess = tf.Session()
# Initialize variables
sess.run(tf.global_variables_initializer(),
feed_dict={is_training_placeholder: True})
sess.run(tf.local_variables_initializer(),
feed_dict={is_training_placeholder: True})
summary_writer = tf.summary.FileWriter(log_dir, sess.graph)
coord = tf.train.Coordinator()
tf.train.start_queue_runners(coord=coord, sess=sess)
with sess.as_default():
if args.pretrained_model:
print('Restoring pretrained model: %s' % args.pretrained_model)
pretrained_saver.restore(
sess, os.path.expanduser(args.pretrained_model))
# Training and validation loop
epoch = 0
while epoch < args.max_nrof_epochs:
eval_one_epoch(args, sess, dataset, image_paths_placeholder,
labels_placeholder, is_training_placeholder,
enqueue_op, clones)
# Train for one epoch
train_one_epoch(args, sess, dataset, image_paths_placeholder,
labels_placeholder, is_training_placeholder,
enqueue_op, input_queue, clones, total_loss,
train_op, summary_op, summary_writer)
# Save variables and the metagraph if it doesn't exist already
global_step = variables.get_or_create_global_step()
step = sess.run(global_step, feed_dict=None)
print('one epoch finish', step)
save_variables_and_metagraph(sess, saver, summary_writer,
model_dir, subdir, step)
print('saver finish')
sess.close()
return model_dir
self.val_key_list = self.key_list[train_count:]
#def train_key_list():
# return self.train_key_list
#def val_key_list():
# return self.val_key_list
def get_contents(self, car_id):
return self.data_map[car_id]
NUM_GPUS = 0
dataset = AutoDataset()
dataset.split_train_val(0.7, SEED)
prt("dataset loaded %d %d" %
(len(dataset.train_key_list), len(dataset.val_key_list)))
BASE_EPOCH_SIZE = 10000
def parse_arguments(argv):
parser = argparse.ArgumentParser()
parser.add_argument('--logs_base_dir',
type=str,
help='Directory where to write event logs.',
default='./logs')
parser.add_argument(
'--models_base_dir',
type=str,
help='Directory where to write trained models and checkpoints.',
default='./models')