def read_datasets(args):
imdb, roidb = read_db(args.imdb_name)
#print('{:d} roidb entries'.format(len(roidb)))
# output directory where the models are saved
#output_dir = get_output_dir(imdb, args.tag)
#output_dir = args.model
#print('Output will be saved to `{:s}`'.format(output_dir))
# tensorboard directory where the summaries are saved during training
tb_dir = get_output_tb_dir(imdb, args.tag)
print('TensorFlow summaries will be saved to `{:s}`'.format(tb_dir))
# also add the validation set, but with no flipping images
orgflip = cfg.TRAIN.USE_FLIPPED
cfg.TRAIN.USE_FLIPPED = False
#imdb, valroidb = read_db(args.imdbval_name)
print('{:d} validation roidb entries'.format(len(valroidb)))
cfg.TRAIN.USE_FLIPPED = orgflip
return imdb, roidb
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):
# train_op = tf.train.AdamOptimizer(0.0005).minimize(loss)
train_op = tf.train.MomentumOptimizer(0.001, 0.9).minimize(loss)
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():
input_data = tf.placeholder(tf.float32, [None, 299, 299, 15])
label_data = tf.placeholder(tf.int32, None)
# Initialize Session #
######################
tfconfig = tf.ConfigProto(allow_soft_placement=True)
tfconfig.gpu_options.allow_growth = True
sess = tf.Session(config=tfconfig)
sess.run(init_op)
# last_iter_num = get_tf_variables(sess)
last_iter_num = 0
cur_saver = tf.train.Saver()
# generate summary on tensorboard
merged = tf.summary.merge_all()
tb_dir = cfg.get_output_tb_dir('darknet19', "imagenet")
train_writer = tf.summary.FileWriter(tb_dir, sess.graph)
# TOTAL_ITER = ADD_ITER + last_iter_num
# T = Timer()
# T.tic()
# for i in range(last_iter_num + 1, TOTAL_ITER + 1):
# print("here")
# summary, loss_value = \
# sess.run([merged, loss])
# print("and here")
# # if i>10:
# train_writer.add_summary(summary, i)
# if i % 1 == 0:
# _time = T.toc(average=False)
# print('iter {:d}/{:d}, total loss: {:.3}, take {:.2}s'.format(i, TOTAL_ITER, loss_value, _time))
img_size = resnet_v1.resnet_v1.default_image_size
# train set
print("Setting up image reader...")
data_reader = deepscores_classification_datareader.class_dataset_reader(
cfg.DATA_DIR + "/DeepScores_2017/DeepScores_classification",
pad_to=[img_size, img_size])
imdb = imdb("DeepScores_2017")
# output directory where the models are saved
output_dir = get_output_dir(imdb, args.tag)
print('Output will be saved to `{:s}`'.format(output_dir))
# tensorboard directory where the summaries are saved during training
tb_dir = get_output_tb_dir(imdb, args.tag)
print('TensorFlow summaries will be saved to `{:s}`'.format(tb_dir))
num_classes = 124
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
# Get the selected model.
# Some of they require pre-trained ResNet
print("Preparing the model ...")
input = tf.placeholder(tf.float32, shape=[None, img_size, img_size, 1])
label = tf.placeholder(tf.int32, shape=[None, num_classes])
network = None
# print 'Initializing new variables to train from imagenet trained model'
# sess.run(init_op)
# saver.restore(sess, os.path.join(LOAD_CKPTS_DIR, 'train_epoch_98.ckpt'))
last_iter_num = restore_darknet19_variables(sess,
imdb,
net_name='darknet19',
save_epoch=False)
####################################################################################
cur_saver = tf.train.Saver()
# generate summary on tensorboard
merged = tf.summary.merge_all()
tb_dir, _ = cfg.get_output_tb_dir('darknet19', imdb.name, val=False)
train_writer = tf.summary.FileWriter(tb_dir, sess.graph)
TOTAL_ITER = ADD_ITER + last_iter_num
T = Timer()
T.tic()
for i in range(last_iter_num + 1, TOTAL_ITER + 1):
image, gt_labels = imdb.get()
summary, loss_value, _, ious_value, object_mask_value = \
sess.run([merged, loss, train_op, ious, object_mask],
{input_data: image, label_data: gt_labels, is_training: 1})
# if i>10:
train_writer.add_summary(summary, i)
if i % 10 == 0:
images, labels = imdb.get()
queue_out.put([images, labels])
imdb = ilsvrc_cls('train', data_aug=True, multithread=cfg.MULTITHREAD, batch_size=32)
val_imdb = ilsvrc_cls('val', batch_size=32)
# 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('resnet50', imdb.name)
TENSORBOARD_TRAIN_DIR, TENSORBOARD_VAL_DIR = cfg.get_output_tb_dir(
'resnet50', 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)
with slim.arg_scope(resnet_v1.resnet_arg_scope()):
logits, end_points = resnet_v1_50(input_data, num_classes=imdb.num_class,
is_training=is_training, global_pool=True)
loss = tf.nn.sparse_softmax_cross_entropy_with_logits(
labels=label_data, logits=logits)
loss = tf.reduce_mean(loss)
vars_to_train = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope='resnet_v1_50/logits')
assert len(vars_to_train) != 0
print "###vars to train###:", vars_to_train
######################
# Initialize Session #
######################
tfconfig = tf.ConfigProto(allow_soft_placement=True)
tfconfig.gpu_options.allow_growth = True
sess = tf.Session(config=tfconfig)
last_iter_num = restore_resnet_tf_variables(
sess, imdb, 'resnet50', save_epoch=False)
cur_saver = tf.train.Saver()
# generate summary on tensorboard
merged = tf.summary.merge_all()
tb_dir, _ = cfg.get_output_tb_dir('resnet50', imdb.name, val=False)
train_writer = tf.summary.FileWriter(tb_dir, sess.graph)
TOTAL_ITER = ADD_ITER + last_iter_num
T = Timer()
T.tic()
for i in range(last_iter_num + 1, TOTAL_ITER + 1):
image, gt_labels = imdb.get()
summary, loss_value, _, ious_value, object_mask_value = \
sess.run([merged, loss, train_op, ious, object_mask],
{input_data: image, label_data: gt_labels})
# if i>10:
train_writer.add_summary(summary, i)
if i % 10 == 0: