def validate_network(sess):
with tf.name_scope('validation'):
images, masks = list_files(FLAGS.valid)
dataset = create_dataset(image_files=images,
mask_files=masks,
batch_size=32)
iterator = dataset.make_initializable_iterator()
images, masks = iterator.get_next()
tensors = build_network(name=FLAGS.name,
reuse=True,
train=False,
images=images,
masks=masks,
iou=True)
def validate_fn():
sess.run(iterator.initializer)
while True:
try:
sess.run(tensors.update_iou)
except tf.errors.OutOfRangeError:
break
iou_val = sess.run(tensors.mean_iou)
return iou_val
return validate_fn
def score_networks(sess):
with tf.name_scope('score'):
follow_dataset = score_dataset('following_images')
target_dataset = score_dataset('patrol_with_targ')
image_files = os.path.join(FLAGS.evaluate, '*/images/*.jpeg')
mask_files = os.path.join(FLAGS.evaluate, '*/masks/*.png')
image_files, mask_files = glob_list_dir(image_files, mask_files)
concat_dataset = create_dataset(image_files,
mask_files,
batch_size=32,
mask_fmt='png')
follow_fn = score_network('follow', sess, follow_dataset)
target_fn = score_network('target', sess, target_dataset)
concat_fn = score_network('concat', sess, concat_dataset)
def score_fn():
target_iou = target_fn()
follow_iou = follow_fn()
print('iou:', follow_iou, target_iou)
final_iou = (follow_fn() + target_fn()) / 2
weight = concat_fn()
return final_iou, weight
return score_fn
def score_dataset(name):
path = os.path.join(FLAGS.evaluate, name)
images, masks = list_files(path)
dataset = create_dataset(image_files=images,
mask_files=masks,
is_train=False,
batch_size=32)
return dataset
def score_datasets():
follow_dataset = score_dataset('following_images')
target_dataset = score_dataset('patrol_with_targ')
image_files = os.path.join(FLAGS.evaluate, '*/images/*.jpeg')
mask_files = os.path.join(FLAGS.evaluate, '*/masks/*.png')
image_files, mask_files = glob_list_dir(image_files, mask_files)
concat_dataset = create_dataset(image_files,
mask_files,
batch_size=32,
mask_fmt='png')
return follow_dataset, target_dataset, concat_dataset
def train_network(sess):
images, masks = list_files(FLAGS.train)
dataset = create_dataset(image_files=images,
mask_files=masks,
is_train=True)
iterator = dataset.make_initializable_iterator()
images, masks = iterator.get_next()
with tf.name_scope('train'):
tensors = build_network(FLAGS.name,
images=images,
masks=masks,
train=True,
iou=True)
def train_fun(epoch):
total_loss, counter = 0, 0
sess.run(iterator.initializer)
while True:
try:
loss, _, _ = sess.run(
[tensors.loss, tensors.train_op, tensors.update_iou])
total_loss += loss
counter += 1
if counter % 10 == 0:
train_iou = sess.run(tensors.mean_iou)
log_info('[%s] %d samples processed, loss [%f] iou [%f]' %
(timestamp(), counter * FLAGS.batch_size,
total_loss / counter, train_iou))
except tf.errors.OutOfRangeError:
break
train_iou = sess.run(tensors.mean_iou)
average_loss = total_loss / counter
log_info('[%s] epoch [%d] training completed' % (timestamp(), epoch))
return train_iou, average_loss
return train_fun