def main(args):
if not os.path.exists(args.pretrained_model):
print('invalid pretrained model path')
return
weights = np.load(args.pretrained_model)
pairs = test_utils.read_pairs('/exports_data/czj/data/lfw/files/pairs.txt')
imglist, labels = test_utils.get_paths(
'/exports_data/czj/data/lfw/lfw_aligned/', pairs, '_face_.jpg')
total_images = len(imglist)
# ---- build graph ---- #
input = tf.placeholder(tf.float32,
shape=[None, 160, 160, 3],
name='image_batch')
prelogits, _ = inception_resnet_v1.inference(input, 1, phase_train=False)
embeddings = tf.nn.l2_normalize(prelogits, 1, 1e-10)
# ---- extract ---- #
gpu_options = tf.GPUOptions(allow_growth=True)
sess = tf.Session(config=tf.ConfigProto(gpu_options=gpu_options,
log_device_placement=False,
allow_soft_placement=True))
with sess.as_default():
beg_time = time.time()
to_assign = [
v.assign(weights[()][v.name][0])
for v in tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES)
]
sess.run(to_assign)
print('restore parameters: %.2fsec' % (time.time() - beg_time))
beg_time = time.time()
images = load_data(imglist)
print('load images: %.2fsec' % (time.time() - beg_time))
beg_time = time.time()
batch_size = 32
beg = 0
end = 0
features = np.zeros((total_images, 128))
while end < total_images:
end = min(beg + batch_size, total_images)
features[beg:end] = sess.run(embeddings, {input: images[beg:end]})
beg = end
print('extract features: %.2fsec' % (time.time() - beg_time))
tpr, fpr, acc, vr, vr_std, far = test_utils.evaluate(features,
labels,
num_folds=10)
# display
auc = metrics.auc(fpr, tpr)
eer = brentq(lambda x: 1. - x - interpolate.interp1d(fpr, tpr)(x), 0., 1.)
print('Acc: %1.3f+-%1.3f' % (np.mean(acc), np.std(acc)))
print('VR@FAR=%2.5f: %2.5f+-%2.5f' % (far, vr, vr_std))
print('AUC: %1.3f' % auc)
print('EER: %1.3f' % eer)
sess.close()
def main(args):
pairs = test_utils.read_pairs(args.lfw_pairs)
model_list = test_utils.get_model_list(args.model_list)
for t, model in enumerate(model_list):
# get lfw pair filename
paths, labels = test_utils.get_paths(args.lfw_dir, pairs, model[1])
with tf.device('/gpu:%d' % (t + 1)):
gpu_options = tf.GPUOptions(allow_growth=True)
sess = tf.Session(config=tf.ConfigProto(gpu_options=gpu_options,
log_device_placement=False, allow_soft_placement=True))
with sess.as_default():
print("[%d] model: %s" % (t, model[1]))
# restore model
test_utils.load_model(sess, model[0])
# load data tensor
images_pl = tf.get_default_graph().get_tensor_by_name('image_batch:0')
embeddings = tf.get_default_graph().get_tensor_by_name('embeddings:0')
phase_train_pl = tf.get_default_graph().get_tensor_by_name('phase_train:0')
image_size = args.image_size
emb_size = embeddings.get_shape()[1]
# extract feature
batch_size = args.lfw_batch_size
num_images = len(paths)
num_batches = num_images // batch_size
emb_arr = np.zeros((num_images, emb_size))
for i in range(num_batches):
print('process %d/%d' % (i + 1, num_batches), end='\r')
beg_idx = i * batch_size
end_idx = min((i + 1) * batch_size, num_images)
images = test_utils.load_data(paths[beg_idx:end_idx], image_size)
emb = sess.run(embeddings, feed_dict={images_pl: images, phase_train_pl: False})
emb_arr[beg_idx:end_idx, :] = emb
# get lfw pair filename
print("\ndone.")
# concate feaure
if t == 0:
emb_ensemble = emb_arr * math.sqrt(float(model[2]))
else:
emb_ensemble = np.concatenate((emb_ensemble, emb_arr * math.sqrt(float(model[2]))), axis=1)
print("ensemble feature:", emb_ensemble.shape)
'''
norm = np.linalg.norm(emb_ensemble, axis=1)
for i in range(emb_ensemble.shape[0]):
emb_ensemble[i] = emb_ensemble[i] / norm[i]
'''
tpr, fpr, acc, vr, vr_std, far = test_utils.evaluate(emb_ensemble, labels, num_folds=args.num_folds)
# display
auc = metrics.auc(fpr, tpr)
eer = brentq(lambda x: 1. - x - interpolate.interp1d(fpr, tpr)(x), 0., 1.)
print('Acc: %1.3f+-%1.3f' % (np.mean(acc), np.std(acc)))
print('VR@FAR=%2.5f: %2.5f+-%2.5f' % (far, vr, vr_std))
print('AUC: %1.3f' % auc)
print('EER: %1.3f' % eer)
def main(args):
model_module = importlib.import_module(args.model_def)
subdir = datetime.strftime(datetime.now(), '%Y%m%d-%H%M%S')
log_dir = os.path.join(args.logs_base_dir, subdir)
model_dir = os.path.join(args.models_base_dir, subdir)
if not os.path.isdir(log_dir):
os.makedirs(log_dir)
if not os.path.isdir(model_dir):
os.makedirs(model_dir)
print('log dir: %s' % log_dir)
print('model dir: %s' % model_dir)
if args.lfw_dir:
print('LFW directory: %s' % args.lfw_dir)
pairs = test_utils.read_pairs(args.lfw_pairs)
lfw_paths, actual_issame = test_utils.get_paths(
args.lfw_dir, pairs, args.lfw_file_ext)
with tf.Graph().as_default():
# ---- data prepration ---- #
image_list, label_list, num_classes = train_utils.get_datasets(
args.data_dir, args.imglist_path)
range_size = len(image_list)
assert range_size > 0, 'The dataset should not be empty.'
# random indices producer
indices_que = tf.train.range_input_producer(range_size)
deque_op = indices_que.dequeue_many(args.batch_size * args.epoch_size,
'index_dequeue')
tf.set_random_seed(args.seed)
random.seed(args.seed)
np.random.seed(args.seed)
global_step = tf.Variable(0, trainable=False)
lr_pl = tf.placeholder(tf.float32, name='learning_rate')
batch_size_pl = tf.placeholder(tf.int32, name='batch_size')
phase_train_pl = tf.placeholder(tf.bool, name='phase_train')
imgpaths_pl = tf.placeholder(tf.string, name='image_paths')
labels_pl = tf.placeholder(tf.int64, name='labels')
# filename queue
input_queue = tf.FIFOQueue(
# [notice: capacity > bach_size*epoch_size]
capacity=100000,
dtypes=[tf.string, tf.int64],
shapes=[(1, ), (1, )],
shared_name=None,
name='input_que')
enque_op = input_queue.enqueue_many([imgpaths_pl, labels_pl],
name='enque_op')
# define 4 readers
num_threads = 4
threads_input_list = []
for _ in range(num_threads):
img_paths, label = input_queue.dequeue(
) # [notice: 'img_pathx' and 'label' are both tensors]
images = []
for img_path in tf.unstack(img_paths):
img_contents = tf.read_file(img_path)
img = tf.image.decode_jpeg(img_contents)
if args.random_crop:
img = tf.random_crop(img,
[args.image_size, args.image_size, 3])
else:
img = tf.image.resize_image_with_crop_or_pad(
img, args.image_size, args.image_size)
if args.random_flip:
img = tf.image.random_flip_left_right(img)
img.set_shape((args.image_size, args.image_size, 3))
images.append(
tf.image.per_image_standardization(img)) # prewhitened?
threads_input_list.append([images, label])
# define 4 buffer queue
image_batch, label_batch = tf.train.batch_join(
threads_input_list,
# [notice: here is 'batch_size_pl', not 'batch_size'!!]
batch_size=batch_size_pl,
shapes=[(args.image_size, args.image_size, 3), ()],
enqueue_many=True,
# [notice: how long the prefetching is allowed to fill the queue]
capacity=4 * num_threads * args.batch_size,
allow_smaller_final_batch=True)
image_batch = tf.identity(image_batch, 'image_batch')
image_batch = tf.identity(image_batch, 'input')
label_batch = tf.identity(label_batch, 'label_batch')
print('Total classes: %d' % num_classes)
print('Total images: %d' % range_size)
tf.summary.image('input_images', image_batch, 10)
# ---- build graph ---- #
with tf.device('/gpu:%d' % args.gpu_id):
# embeddings
prelogits, _ = model_module.inference(
image_batch,
args.keep_prob,
phase_train=phase_train_pl,
weight_decay=args.weight_decay)
# logits
logits = slim.fully_connected(
prelogits,
num_classes,
activation_fn=None,
weights_initializer=tf.truncated_normal_initializer(
stddev=0.1),
weights_regularizer=slim.l2_regularizer(args.weight_decay),
scope='Logits',
reuse=False)
# normalized features
# [notice: used in test stage]
embeddings = tf.nn.l2_normalize(prelogits,
1,
1e-10,
name='embeddings')
# ---- define loss & train op ---- #
# center loss
if args.center_loss_factor > 0.0:
prelogits_center_loss, _ = train_utils.center_loss(
prelogits, label_batch, args.center_loss_alpha,
num_classes)
tf.summary.scalar(
'center_loss',
prelogits_center_loss * args.center_loss_factor)
tf.add_to_collection(
tf.GraphKeys.REGULARIZATION_LOSSES,
prelogits_center_loss * args.center_loss_factor)
# cross-entropy
cross_entropy_mean = tf.reduce_mean(
tf.nn.sparse_softmax_cross_entropy_with_logits(
labels=label_batch, logits=logits),
name='cross_entropy')
tf.add_to_collection('losses', cross_entropy_mean)
# regularity: weight decay
reg_loss = tf.get_collection(tf.GraphKeys.REGULARIZATION_LOSSES)
# total loss
total_loss = tf.add_n([cross_entropy_mean] + reg_loss,
name='total_loss')
# [notice: here we decay manually]
lr = tf.train.exponential_decay(lr_pl,
global_step,
args.lr_decay_epochs *
args.epoch_size,
args.lr_decay_factor,
staircase=True)
tf.summary.scalar('learning_rate', lr)
train_op = train_utils.get_train_op(
total_loss,
global_step,
args.optimizer,
lr,
args.moving_average_decay,
# what is the usage of tf.global_variables()?
tf.trainable_variables())
# ---- training ---- #
# [notice: use 'allow_growth' instead of memory_fraction]
gpu_options = tf.GPUOptions(allow_growth=True)
# [notice: use 'allow_soft_placement' to solve the problem of 'no supported kernel...']
sess = tf.Session(config=tf.ConfigProto(gpu_options=gpu_options,
log_device_placement=False,
allow_soft_placement=True))
sess.run(tf.global_variables_initializer())
sess.run(tf.local_variables_initializer())
# [notice: 'max_to_keep': keep at most 'max_to_keep' checkpoint files]
saver = tf.train.Saver(tf.trainable_variables(), max_to_keep=5)
summary_op = tf.summary.merge_all()
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('Resume training: %s' % args.pretrained_model)
saver.restore(sess, args.pretrained_model)
print('Start training ...')
epoch = 0
while epoch < args.max_num_epochs:
step = sess.run(global_step,
feed_dict=None) # training counter
epoch = step // args.epoch_size
# run epoch
run_epoch(args, sess, epoch, image_list, label_list, deque_op,
enque_op, imgpaths_pl, labels_pl, lr_pl,
phase_train_pl, batch_size_pl, global_step,
total_loss, reg_loss, train_op, summary_op,
summary_writer)
# snapshot for currently learnt weights
snapshot(sess, saver, model_dir, subdir, step)
# evaluate on LFW
if args.lfw_dir:
evaluate(sess, enque_op, imgpaths_pl, labels_pl,
phase_train_pl, batch_size_pl, embeddings,
label_batch, lfw_paths, actual_issame,
args.lfw_batch_size, args.lfw_num_folds, log_dir,
step, summary_writer)
sess.close()
def main(args):
subdir = datetime.strftime(datetime.now(), '%Y%m%d-%H%M%S')
log_dir = os.path.join(args.logs_base_dir, subdir, 'logs')
model_dir = os.path.join(args.logs_base_dir, subdir, 'models')
if not os.path.isdir(log_dir):
os.makedirs(log_dir)
if not os.path.isdir(model_dir):
os.makedirs(model_dir)
print('log dir: %s' % log_dir)
print('model dir: %s' % model_dir)
# build the graph
# ---- load pretrained model ---- #
pretrained = {}
# Face model
pretrained['Face'] = np.load(args.face_model)[()]
# Nose model
pretrained['Nose'] = np.load(args.nose_model)[()]
# Lefteye model
pretrained['Lefteye'] = np.load(args.lefteye_model)[()]
# Rightmouth model
pretrained['Rightmouth'] = np.load(args.rightmouth_model)[()]
# ---- data preparation ---- #
image_list, label_list, num_classes = train_utils.get_datasets(
args.data_dir, args.imglist)
range_size = len(image_list)
if args.lfw_dir:
print('LFW directory: %s' % args.lfw_dir)
pairs = test_utils.read_pairs(args.lfw_pairs)
lfw_paths, actual_issame = test_utils.get_paths(
args.lfw_dir, pairs, args.lfw_file_ext)
with tf.Graph().as_default():
# random indices producer
indices_que = tf.train.range_input_producer(range_size)
dequeue_op = indices_que.dequeue_many(
args.batch_size * args.epoch_size, 'index_dequeue')
tf.set_random_seed(args.seed)
random.seed(args.seed)
np.random.seed(args.seed)
global_step = tf.Variable(0, trainable=False)
# lr_base_pl = tf.placeholder(tf.float32, name='base_learning_rate')
lr_fusion_pl = tf.placeholder(tf.float32, name='fusion_learning_rate')
batch_size_pl = tf.placeholder(tf.int32, name='batch_size')
phase_train_pl = tf.placeholder(tf.bool, name='phase_train')
face_pl = tf.placeholder(tf.string, name='image_paths1') # face images
nose_pl = tf.placeholder(tf.string, name='image_paths2') # nose images
lefteye_pl = tf.placeholder(tf.string,
name='image_paths3') # left eye images
rightmouth_pl = tf.placeholder(
tf.string, name='image_paths4') # right mouth images
labels_pl = tf.placeholder(tf.int64, name='labels')
# define a filename queue
input_queue = tf.FIFOQueue(
# [notice: capacity > bach_size*epoch_size]
capacity=100000,
dtypes=[tf.string, tf.string, tf.string, tf.string, tf.int64],
shapes=[(1, ), (1, ), (1, ), (1, ), (1, )],
shared_name=None,
name='input_que')
enque_op = input_queue.enqueue_many(
[face_pl, nose_pl, lefteye_pl, rightmouth_pl, labels_pl],
name='enque_op')
# define 4 readers
num_threads = 4
threads_input_list = []
for _ in range(num_threads):
imgpath1, imgpath2, imgpath3, imgpath4, label = input_queue.dequeue(
)
images1 = []
images2 = []
images3 = []
images4 = []
# face
for img_path in tf.unstack(imgpath1):
img_contents = tf.read_file(img_path)
img = tf.image.decode_jpeg(img_contents)
# [notice: random crop only used in face image]
if args.random_crop:
img = tf.random_crop(img, [160, 160, 3])
else:
img = tf.image.resize_image_with_crop_or_pad(img, 160, 160)
# [notice: flip only used in face image or nose patch]
if args.random_flip:
img = tf.image.random_flip_left_right(img)
img.set_shape((160, 160, 3))
images1.append(tf.image.per_image_standardization(img))
# Nose
for img_path in tf.unstack(imgpath2):
img_contents = tf.read_file(img_path)
img = tf.image.decode_jpeg(img_contents)
# [notice: flip only used in face image or nose patch]
if args.random_flip:
img = tf.image.random_flip_left_right(img)
img.set_shape((160, 160, 3))
images2.append(tf.image.per_image_standardization(img))
# Lefteye
for img_path in tf.unstack(imgpath3):
img_contents = tf.read_file(img_path)
img = tf.image.decode_jpeg(img_contents)
img.set_shape((160, 160, 3))
images3.append(tf.image.per_image_standardization(img))
# Rightmouth
for img_path in tf.unstack(imgpath4):
img_contents = tf.read_file(img_path)
img = tf.image.decode_jpeg(img_contents)
img.set_shape((160, 160, 3))
images4.append(tf.image.per_image_standardization(img))
threads_input_list.append(
[images1, images2, images3, images4, label])
# define 4 buffer queue
face_batch, nose_batch, lefteye_batch, rightmouth_batch, label_batch = tf.train.batch_join(
threads_input_list,
# [notice: here is 'batch_size_pl', not 'batch_size'!!]
batch_size=batch_size_pl,
shapes=[
# [notice: shape of each element should be assigned, otherwise it raises
# "tensorflow queue shapes must have the same length as dtype" exception]
(args.image_size, args.image_size, 3),
(args.image_size, args.image_size, 3),
(args.image_size, args.image_size, 3),
(args.image_size, args.image_size, 3),
()
],
enqueue_many=True,
# [notice: how long the prefetching is allowed to fill the queue]
capacity=4 * num_threads * args.batch_size,
allow_smaller_final_batch=True)
print('Total classes: %d' % num_classes)
print('Total images: %d' % range_size)
tf.summary.image('face_images', face_batch, 10)
tf.summary.image('nose_images', nose_batch, 10)
tf.summary.image('lefteye_images', lefteye_batch, 10)
tf.summary.image('rightmouth_images', rightmouth_batch, 10)
# ---- build graph ---- #
with tf.variable_scope('BaseModel'):
with tf.device('/gpu:%d' % args.gpu_id1):
# embeddings for face model
features1, _ = inception_resnet_v1.inference(
face_batch,
args.keep_prob,
phase_train=phase_train_pl,
weight_decay=args.weight_decay,
scope='Face')
with tf.device('/gpu:%d' % args.gpu_id2):
# embeddings for nose model
features2, _ = inception_resnet_v1.inference(
nose_batch,
args.keep_prob,
phase_train=phase_train_pl,
weight_decay=args.weight_decay,
scope='Nose')
with tf.device('/gpu:%d' % args.gpu_id3):
# embeddings for left eye model
features3, _ = inception_resnet_v1.inference(
lefteye_batch,
args.keep_prob,
phase_train=phase_train_pl,
weight_decay=args.weight_decay,
scope='Lefteye')
with tf.device('/gpu:%d' % args.gpu_id4):
# embeddings for right mouth model
features4, _ = inception_resnet_v1.inference(
rightmouth_batch,
args.keep_prob,
phase_train=phase_train_pl,
weight_decay=args.weight_decay,
scope='Rightmouth')
with tf.device('/gpu:%d' % args.gpu_id5):
with tf.variable_scope("Fusion"):
# ---- concatenate ---- #
concated_features = tf.concat(
[features1, features2, features3, features4], 1)
# prelogits
prelogits = slim.fully_connected(
concated_features,
args.fusion_dim,
activation_fn=None,
weights_initializer=tf.truncated_normal_initializer(
stddev=0.1),
weights_regularizer=slim.l2_regularizer(args.weight_decay),
scope='prelogits',
reuse=False)
# logits
logits = slim.fully_connected(
prelogits,
num_classes,
activation_fn=None,
weights_initializer=tf.truncated_normal_initializer(
stddev=0.1),
weights_regularizer=slim.l2_regularizer(args.weight_decay),
scope='logits',
reuse=False)
# normalized feaures
# [notice: used in test stage]
embeddings = tf.nn.l2_normalize(prelogits,
1,
1e-10,
name='embeddings')
# ---- define loss & train op ---- #
cross_entropy = -tf.reduce_sum(tf.one_hot(
indices=tf.cast(label_batch, tf.int32),
depth=num_classes) * tf.log(tf.nn.softmax(logits) + 1e-10),
reduction_indices=[1])
cross_entropy_mean = tf.reduce_mean(cross_entropy)
tf.add_to_collection('losses', cross_entropy_mean)
# weight decay
reg_loss = tf.get_collection(tf.GraphKeys.REGULARIZATION_LOSSES)
# total loss: cross_entropy + weight_decay
total_loss = tf.add_n([cross_entropy_mean] + reg_loss,
name='total_loss')
'''
lr_base = tf.train.exponential_decay(lr_base_pl,
global_step,
args.lr_decay_epochs * args.epoch_size,
args.lr_decay_factor,
staircase = True)
'''
lr_fusion = tf.train.exponential_decay(lr_fusion_pl,
global_step,
args.lr_decay_epochs *
args.epoch_size,
args.lr_decay_factor,
staircase=True)
# tf.summary.scalar('base_learning_rate', lr_base)
tf.summary.scalar('fusion_learning_rate', lr_fusion)
var_list1 = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES,
scope='BaseModel')
var_list2 = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES,
scope='Fusion')
'''
train_op = train_utils.get_fusion_train_op(
total_loss, global_step, args.optimizer,
lr_base, var_list1, lr_fusion, var_list2,
args.moving_average_decay)
'''
train_op = train_utils.get_train_op(total_loss, global_step,
args.optimizer, lr_fusion,
args.moving_average_decay,
var_list2)
# ---- training ---- #
gpu_options = tf.GPUOptions(allow_growth=True)
sess = tf.Session(config=tf.ConfigProto(
gpu_options=gpu_options,
log_device_placement=False,
# [notice: 'allow_soft_placement' will switch to cpu automatically
# when some operations are not supported by GPU]
allow_soft_placement=True))
saver = tf.train.Saver(var_list1 + var_list2)
summary_op = tf.summary.merge_all()
summary_writer = tf.summary.FileWriter(log_dir, sess.graph)
sess.run(tf.global_variables_initializer())
sess.run(tf.local_variables_initializer())
coord = tf.train.Coordinator()
tf.train.start_queue_runners(coord=coord, sess=sess)
with sess.as_default():
# ---- restore pre-trained parameters ---- #
to_assign = []
print("restore pretrained parameters...")
print("total:", len(var_list1))
for v in var_list1:
v_name = v.name # 'BaseModel/Face/xxx'
v_name = v_name[v_name.find('/') + 1:] # 'Face/xxx'
v_name_1 = v_name[:v_name.find('/')] # 'Face'
v_name_2 = v_name[v_name.find('/'):] # '/xxx'
print("precess: %s" % v_name, end=" ")
if v_name_1 in pretrained:
to_assign.append(
v.assign(pretrained[v_name_1][v_name_2][0]))
print("[ok]")
else:
print("[no found]")
v.assign(pretrained[v_name_1][v_name_2][0])
print("done")
sess.run(to_assign)
print("start training ...")
epoch = 0
while epoch < args.max_num_epochs:
step = sess.run(global_step, feed_dict=None)
epoch = step // args.epoch_size
# run one epoch
run_epoch(args, sess, epoch, image_list, label_list,
dequeue_op, enque_op, face_pl, nose_pl, lefteye_pl,
rightmouth_pl, labels_pl, lr_fusion_pl,
phase_train_pl, batch_size_pl, global_step,
total_loss, reg_loss, train_op, summary_op,
summary_writer)
# snapshot for currently learnt weights
snapshot(sess, saver, model_dir, subdir, step)
# evaluate on LFW
if args.lfw_dir:
evaluate(sess, enque_op, face_pl, nose_pl, lefteye_pl,
rightmouth_pl, labels_pl, phase_train_pl,
batch_size_pl, embeddings, label_batch, lfw_paths,
actual_issame, args.lfw_batch_size,
args.lfw_num_folds, log_dir, step, summary_writer)
sess.close()
def main(args):
subdir = datetime.strftime(datetime.now(), '%Y%m%d-%H%M%S')
log_dir = os.path.join(args.logs_base_dir, 'logs', subdir)
model_dir = os.path.join(args.logs_base_dir, 'models', subdir)
if not os.path.isdir(log_dir):
os.makedirs(log_dir)
if not os.path.isdir(model_dir):
os.makedirs(model_dir)
print('log dir: %s' % log_dir)
print('model dir: %s' % model_dir)
if args.lfw_dir:
print('lfw directory: %s' % args.lfw_dir)
pairs = test_utils.read_pairs(args.lfw_pairs)
lfw_paths, lfw_label = test_utils.get_paths(args.lfw_dir, pairs,
args.lfw_file_ext)
with tf.Graph().as_default():
# ------------ data preparation ------------ #
image_list, label_list, num_classes = train_utils.get_datasets(
args.data_dir, args.imglist_path)
range_size = len(image_list)
assert range_size > 0, 'The data set should not be empty.'
# random indices producer
indices_que = tf.train.range_input_producer(range_size)
deque_op = indices_que.dequeue_many(args.batch_size * args.epoch_size,
'index_dequeue')
# [notice: how to set random seed?]
tf.set_random_seed(args.seed)
random.seed(args.seed)
np.random.seed(args.seed)
# filename queue
imgpaths_pl = tf.placeholder(tf.string, name='image_paths')
labels_pl = tf.placeholder(tf.int64, name='labels')
input_queue = tf.FIFOQueue(
# [notice: capacity > bach_size*epoch_size]
capacity=100000,
dtypes=[tf.string, tf.int64],
shapes=[(1, ), (1, )],
shared_name=None,
name='input_que')
enque_op = input_queue.enqueue_many([imgpaths_pl, labels_pl],
name='enque_op')
# define 4 readers
num_threads = 4
threads_input_list = []
for _ in range(num_threads):
img_paths, label = input_queue.dequeue(
) # [notice: 'img_pathx' and 'label' are both tensors]
images = []
for img_path in tf.unstack(img_paths):
img_contents = tf.read_file(img_path)
img = tf.image.decode_jpeg(img_contents)
if args.random_crop:
img = tf.random_crop(img,
[args.image_size, args.image_size, 3])
else:
img = tf.image.resize_image_with_crop_or_pad(
img, args.image_size, args.image_size)
if args.random_flip:
img = tf.image.random_flip_left_right(img)
img.set_shape((args.image_size, args.image_size, 3))
images.append(
tf.image.per_image_standardization(img)) # pre-whitened?
threads_input_list.append([images, label])
# define 4 buffer queue
batch_size_pl = tf.placeholder(tf.int32, name='batch_size')
image_batch, label_batch = tf.train.batch_join(
threads_input_list,
# [notice: here is 'batch_size_pl', not 'batch_size'!!]
batch_size=batch_size_pl,
shapes=[(args.image_size, args.image_size, 3), ()],
enqueue_many=True,
# [notice: how long the pre-fetching is allowed to fill the queue]
capacity=4 * num_threads * args.batch_size,
allow_smaller_final_batch=True)
image_batch = tf.identity(image_batch, 'image_batch')
label_batch = tf.identity(label_batch, 'label_batch')
print('Total classes: %d' % num_classes)
print('Total images: %d' % range_size)
tf.summary.image('input_images', image_batch, 10)
# ------------ build graph ------------ #
hps_train = resnet.HParams(batch_size=batch_size_pl,
num_residual_units=5,
use_bottleneck=True,
relu_leakiness=0.1)
global_step = tf.Variable(0, trainable=False)
phase_train_pl = tf.placeholder(tf.bool, name='phase_train')
resnet_model = resnet(hps_train, phase_train_pl)
with tf.device('/gpu:%d' % args.gpu_id):
# ---- base graph ---- #
with tf.variable_scope('ResNet'):
# prelogits
prelogits = resnet_model.inference(image_batch)
# prelogits -> embeddings [notice: used in test stage]
embeddings = tf.nn.l2_normalize(prelogits,
1,
1e-10,
name='embeddings')
# prelogits -> logits
with tf.variable_scope('Logits'):
logits = resnet.fully_connected(prelogits, num_classes)
# predictions = tf.nn.softmax(logits)
# ---- losses ---- #
# cross entropy
with tf.variable_scope('cross_entropy'):
cross_entropy = tf.reduce_sum(
tf.one_hot(indices=tf.cast(label_batch, tf.int32),
depth=num_classes) *
tf.log(tf.nn.softmax(logits) + 1e-10),
reduction_indices=[1])
cross_entropy_mean = tf.reduce_mean(cross_entropy)
tf.summary.scalar('cross_entropy', cross_entropy_mean)
# l2 loss
reg_loss = resnet.decay(args.weight_decay)
tf.summary.scalar('reg_loss', reg_loss)
# total loss
total_loss = tf.add_n([cross_entropy_mean] + reg_loss,
name='total_loss')
train_op = resnet_model.get_train_op(total_loss, global_step,
args.lr)
# ------------ training ------------ #
# [notice: use 'allow_growth' instead of memory_fraction]
gpu_options = tf.GPUOptions(allow_growth=True)
# [notice: use 'allow_soft_placement' to solve the problem of 'no supported kernel...']
sess = tf.Session(config=tf.ConfigProto(gpu_options=gpu_options,
log_device_placement=False,
allow_soft_placement=True))
sess.run(tf.global_variables_initializer())
sess.run(tf.local_variables_initializer())
# [notice: 'max_to_keep': keep at most 'max_to_keep' checkpoint files]
saver = tf.train.Saver(tf.trainable_variables(), max_to_keep=5)
summary_op = tf.summary.merge_all()
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('Resume training: %s' % args.pretrained_model)
saver.restore(sess, args.pretrained_model)
print('Start training ...')
epoch = 0
while epoch < args.max_num_epochs:
step = sess.run(global_step,
feed_dict=None) # training counter
epoch = step // args.epoch_size
# run epoch
run_epoch(args, sess, epoch, image_list, label_list, deque_op,
enque_op, imgpaths_pl, labels_pl, phase_train_pl,
batch_size_pl, global_step, total_loss, reg_loss,
train_op, summary_op, summary_writer)
# snapshot for currently learnt weights
snapshot(sess, saver, model_dir, subdir, step)
# evaluate on LFW
if args.lfw_dir:
evaluate(sess, enque_op, imgpaths_pl, labels_pl,
phase_train_pl, batch_size_pl, embeddings,
label_batch, lfw_paths, lfw_label,
args.lfw_batch_size, args.lfw_num_folds, log_dir,
step, summary_writer)
sess.close()