def restore_darknet19_variables(sess, imdb, net_name='darknet19', save_epoch=True):
"""Initialize or restore the varialbes in darknet19."""
sfiles = get_ordered_ckpts(sess, imdb, net_name, save_epoch=save_epoch)
lsf = len(sfiles)
if lsf == 0:
# TODO: this is not used because the current weight has deprecated scope names. Need to fix this
# if os.path.isfile(cfg.darknet_imagenet_weight_path + ".meta"):
# # Try to load from weight file first
# print 'Restorining model from weight file {:s}'.format(cfg.darknet_imagenet_weight_path)
# saver = tf.train.Saver()
# saver.restore(sess, cfg.darknet_pascal_weight_path)
# print 'Restored.'
# loading from imagenet trained model
# currently assume trined previously on imagenet
LOAD_CKPTS_DIR = cfg.get_ckpts_dir('darknet19', 'ilsvrc_2017_cls')
ckpt_vars = [t[0] for t in tf.contrib.framework.list_variables(LOAD_CKPTS_DIR)]
vars_to_restore = []
vars_to_init = []
for v in tf.global_variables():
if v.name[:-2] in ckpt_vars:
vars_to_restore.append(v)
else:
vars_to_init.append(v)
# print 'vars_to_restore:', len(vars_to_restore)
# print 'vars_to_init:', len(vars_to_init)
init_op = tf.variables_initializer(vars_to_init)
saver = tf.train.Saver(vars_to_restore)
print 'Initializing new variables to train from imagenet trained model'
sess.run(init_op)
imagnet_sfiles = get_ordered_ckpts_by_dbname(sess, 'ilsvrc_2017_cls', net_name, save_epoch=True)
print 'Restorining model snapshots from {:s}'.format(imagnet_sfiles[-1])
saver.restore(sess, str(imagnet_sfiles[-1]))
return 0
else:
print 'Restorining model snapshots from {:s}'.format(sfiles[-1])
saver = tf.train.Saver()
saver.restore(sess, str(sfiles[-1]))
print 'Restored.'
fnames = sfiles[-1].split('_')
return int(fnames[-1][:-5])
def get_ordered_ckpts_by_dbname(sess, imdb_name, net_name, save_epoch=True):
"""Get the ckpts for particular network on certain dataset (by name).
The ckpts is ordered in ascending order of time.
Returns: sorted list of ckpt names.
"""
# Find previous snapshots if there is any to restore from
ckpts_dir = cfg.get_ckpts_dir(net_name, imdb_name)
if save_epoch:
save_interval = 'epoch'
else:
save_interval = 'iter'
sfiles = os.path.join(ckpts_dir,
cfg.TRAIN_SNAPSHOT_PREFIX + '_' + save_interval + '_*.ckpt.meta')
sfiles = glob.glob(sfiles)
sfiles.sort(key=os.path.getmtime)
# Get the snapshot name in TensorFlow
sfiles = [ss.replace('.meta', '') for ss in sfiles]
return sfiles
queue_in.get()
images, labels = imdb.get()
queue_out.put([images, labels])
imdb = ilsvrc_cls('train', data_aug=True, multithread=cfg.MULTITHREAD)
val_imdb = ilsvrc_cls('val', batch_size=64)
# set up child process for getting validation data
queue_in = Queue()
queue_out = Queue()
val_data_process = Process(target=get_validation_process,
args=(val_imdb, queue_in, queue_out))
val_data_process.start()
queue_in.put(True) # start getting the first batch
CKPTS_DIR = cfg.get_ckpts_dir('darknet19', imdb.name)
TENSORBOARD_TRAIN_DIR, TENSORBOARD_VAL_DIR = cfg.get_output_tb_dir(
'darknet19', imdb.name)
input_data = tf.placeholder(tf.float32, [None, 224, 224, 3])
label_data = tf.placeholder(tf.int32, None)
is_training = tf.placeholder(tf.bool)
logits = darknet19(input_data, is_training=is_training)
loss = tf.nn.sparse_softmax_cross_entropy_with_logits(
labels=label_data, logits=logits)
loss = tf.reduce_mean(loss)
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(update_ops):
# NOTE: check fix the data imread and label
imdb = ilsvrc_cls('train',
multithread=cfg.MULTITHREAD,
batch_size=TRAIN_BATCH_SIZE,
image_size=299,
random_noise=True)
val_imdb = ilsvrc_cls('val', batch_size=18, image_size=299, random_noise=True)
# set up child process for getting validation data
queue_in = Queue()
queue_out = Queue()
val_data_process = Process(target=get_validation_process,
args=(val_imdb, queue_in, queue_out))
val_data_process.start()
queue_in.put(True) # start getting the first batch
CKPTS_DIR = cfg.get_ckpts_dir('inception_resnet', imdb.name)
TENSORBOARD_TRAIN_DIR, TENSORBOARD_VAL_DIR = cfg.get_output_tb_dir(
'inception_resnet', imdb.name)
TENSORBOARD_TRAIN_ADV_DIR = os.path.abspath(
os.path.join(cfg.ROOT_DIR, 'tensorboard', 'inception_resnet', imdb.name,
'train_adv'))
if not os.path.exists(TENSORBOARD_TRAIN_ADV_DIR):
os.makedirs(TENSORBOARD_TRAIN_ADV_DIR)
TENSORBOARD_VAL_ADV_DIR = os.path.abspath(
os.path.join(cfg.ROOT_DIR, 'tensorboard', 'inception_resnet', imdb.name,
'val_adv'))
if not os.path.exists(TENSORBOARD_VAL_ADV_DIR):
os.makedirs(TENSORBOARD_VAL_ADV_DIR)
g_inception_resnet = tf.Graph()
with g_inception_resnet.as_default():
from utils.timer import Timer
from yolo2_nets.net_utils import restore_resnet_tf_variables, show_yolo_detection
from yolo2_nets.tf_resnet import resnet_v1_50
slim = tf.contrib.slim
# TODO: make the image path to be user input
image_path = '/home/wenxi/Projects/tensorflow_yolo2/experiments/fig1.jpg'
IMAGE_SIZE = cfg.IMAGE_SIZE
S = cfg.S
B = cfg.B
# create database instance
imdb = pascal_voc('trainval')
NUM_CLASS = imdb.num_class
CKPTS_DIR = cfg.get_ckpts_dir('resnet50', imdb.name)
input_data = tf.placeholder(tf.float32, [None, 224, 224, 3])
# read in the test image
image = cv2.imread(image_path)
image = cv2.resize(image, (IMAGE_SIZE, IMAGE_SIZE))
image = image.astype(np.float32)
image = (image / 255.0) * 2.0 - 1.0
image = image.reshape((1, 224, 224, 3))
# get the right arg_scope in order to load weights
with slim.arg_scope(resnet_v1.resnet_arg_scope()):
# net is shape [batch_size, S, S, 2048] if input size is 244 x 244
net, end_points = resnet_v1_50(input_data, is_training=False)