def _run_inference():
"""Runs all images through depth model and saves depth maps."""
ckpt_basename = os.path.basename(FLAGS.model_ckpt)
ckpt_modelname = os.path.basename(os.path.dirname(FLAGS.model_ckpt))
output_dir = os.path.join(
FLAGS.output_dir,
FLAGS.kitti_video.replace('/', '_') + '_' + ckpt_modelname + '_' +
ckpt_basename)
if not gfile.exists(output_dir):
gfile.makedirs(output_dir)
inference_model = model.Model(is_training=False,
seq_length=FLAGS.seq_length,
batch_size=FLAGS.batch_size,
img_height=FLAGS.img_height,
img_width=FLAGS.img_width)
vars_to_restore = util.get_vars_to_restore(FLAGS.model_ckpt)
saver = tf.compat.v1.train.Saver(vars_to_restore)
sv = tf.compat.v1.train.Supervisor(logdir='/tmp/', saver=None)
with sv.managed_session() as sess:
saver.restore(sess, FLAGS.model_ckpt)
if FLAGS.kitti_video == 'test_files_eigen':
im_files = util.read_text_lines(
util.get_resource_path('dataset/kitti/test_files_eigen.txt'))
im_files = [os.path.join(FLAGS.kitti_dir, f) for f in im_files]
else:
video_path = os.path.join(FLAGS.kitti_dir, FLAGS.kitti_video)
im_files = gfile.glob(
os.path.join(video_path, 'image_02/data', '*.png'))
im_files = [f for f in im_files if 'disp' not in f]
im_files = sorted(im_files)
for i in range(0, len(im_files), FLAGS.batch_size):
if i % 100 == 0:
logging.info('Generating from %s: %d/%d', ckpt_basename, i,
len(im_files))
inputs = np.zeros(
(FLAGS.batch_size, FLAGS.img_height, FLAGS.img_width, 3),
dtype=np.uint8)
for b in range(FLAGS.batch_size):
idx = i + b
if idx >= len(im_files):
break
im = imageio.imread(im_files[idx])
inputs[b] = np.array(
Image.fromarray(im).resize(
(FLAGS.img_width, FLAGS.img_height)))
results = inference_model.inference(inputs, sess, mode='depth')
for b in range(FLAGS.batch_size):
idx = i + b
if idx >= len(im_files):
break
if FLAGS.kitti_video == 'test_files_eigen':
depth_path = os.path.join(output_dir, '%03d.png' % idx)
else:
depth_path = os.path.join(output_dir, '%04d.png' % idx)
depth_map = results['depth'][b]
depth_map = np.squeeze(depth_map)
colored_map = _normalize_depth_for_display(depth_map,
cmap=CMAP)
imageio.imsave(depth_path, colored_map)
def _run_inference():
"""Runs all images through depth model and saves depth maps."""
ckpt_basename = os.path.basename(FLAGS.model_ckpt)
ckpt_modelname = os.path.basename(os.path.dirname(FLAGS.model_ckpt))
output_dir = os.path.join(FLAGS.output_dir,
FLAGS.kitti_video.replace('/', '_') + '_' +
ckpt_modelname + '_' + ckpt_basename)
if not gfile.Exists(output_dir):
gfile.MakeDirs(output_dir)
inference_model = model.Model(is_training=False,
seq_length=FLAGS.seq_length,
batch_size=FLAGS.batch_size,
img_height=FLAGS.img_height,
img_width=FLAGS.img_width)
vars_to_restore = util.get_vars_to_restore(FLAGS.model_ckpt)
saver = tf.train.Saver(vars_to_restore)
sv = tf.train.Supervisor(logdir='/tmp/', saver=None)
with sv.managed_session() as sess:
saver.restore(sess, FLAGS.model_ckpt)
if FLAGS.kitti_video == 'test_files_eigen':
im_files = util.read_text_lines(
util.get_resource_path('dataset/kitti/test_files_eigen.txt'))
im_files = [os.path.join(FLAGS.kitti_dir, f) for f in im_files]
else:
video_path = os.path.join(FLAGS.kitti_dir, FLAGS.kitti_video)
im_files = gfile.Glob(os.path.join(video_path, 'image_02/data', '*.png'))
im_files = [f for f in im_files if 'disp' not in f]
im_files = sorted(im_files)
for i in range(0, len(im_files), FLAGS.batch_size):
if i % 100 == 0:
logging.info('Generating from %s: %d/%d', ckpt_basename, i,
len(im_files))
inputs = np.zeros(
(FLAGS.batch_size, FLAGS.img_height, FLAGS.img_width, 3),
dtype=np.uint8)
for b in range(FLAGS.batch_size):
idx = i + b
if idx >= len(im_files):
break
im = scipy.misc.imread(im_files[idx])
inputs[b] = scipy.misc.imresize(im, (FLAGS.img_height, FLAGS.img_width))
results = inference_model.inference(inputs, sess, mode='depth')
for b in range(FLAGS.batch_size):
idx = i + b
if idx >= len(im_files):
break
if FLAGS.kitti_video == 'test_files_eigen':
depth_path = os.path.join(output_dir, '%03d.png' % idx)
else:
depth_path = os.path.join(output_dir, '%04d.png' % idx)
depth_map = results['depth'][b]
depth_map = np.squeeze(depth_map)
colored_map = _normalize_depth_for_display(depth_map, cmap=CMAP)
input_float = inputs[b].astype(np.float32) / 255.0
vertical_stack = np.concatenate((input_float, colored_map), axis=0)
scipy.misc.imsave(depth_path, vertical_stack)
def train(train_model, pretrained_ckpt, checkpoint_dir, train_steps,
summary_freq):
"""Train model."""
if pretrained_ckpt is not None:
vars_to_restore = util.get_vars_to_restore(pretrained_ckpt)
pretrain_restorer = tf.train.Saver(vars_to_restore)
vars_to_save = util.get_vars_to_restore()
saver = tf.train.Saver(vars_to_save + [train_model.global_step],
max_to_keep=MAX_TO_KEEP)
sv = tf.train.Supervisor(logdir=checkpoint_dir,
save_summaries_secs=0,
saver=None)
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
with sv.managed_session(config=config) as sess:
if pretrained_ckpt is not None:
logging.info('Restoring pretrained weights from %s',
pretrained_ckpt)
pretrain_restorer.restore(sess, pretrained_ckpt)
logging.info('Attempting to resume training from %s...',
checkpoint_dir)
checkpoint = tf.train.latest_checkpoint(checkpoint_dir)
logging.info('Last checkpoint found: %s', checkpoint)
if checkpoint:
saver.restore(sess, checkpoint)
logging.info('Training...')
start_time = time.time()
last_summary_time = time.time()
steps_per_epoch = train_model.reader.steps_per_epoch
step = 1
while step <= train_steps:
fetches = {
'train': train_model.train_op,
'global_step': train_model.global_step,
'incr_global_step': train_model.incr_global_step
}
if step % summary_freq == 0:
fetches['loss'] = train_model.total_loss
fetches['summary'] = sv.summary_op
results = sess.run(fetches)
global_step = results['global_step']
if step % summary_freq == 0:
sv.summary_writer.add_summary(results['summary'], global_step)
train_epoch = math.ceil(global_step / steps_per_epoch)
train_step = global_step - (train_epoch - 1) * steps_per_epoch
this_cycle = time.time() - last_summary_time
last_summary_time += this_cycle
logging.info(
'Epoch: [%2d] [%5d/%5d] time: %4.2fs (%ds total) loss: %.3f',
train_epoch, train_step, steps_per_epoch, this_cycle,
time.time() - start_time, results['loss'])
if step % steps_per_epoch == 0:
logging.info('[*] Saving checkpoint to %s...', checkpoint_dir)
saver.save(sess,
os.path.join(checkpoint_dir, 'model'),
global_step=global_step)
# Setting step to global_step allows for training for a total of
# train_steps even if the program is restarted during training.
step = global_step + 1
def train(train_model, pretrained_ckpt, checkpoint_dir, train_steps,
summary_freq):
"""Train model."""
if pretrained_ckpt is not None:
vars_to_restore = util.get_vars_to_restore(pretrained_ckpt)
pretrain_restorer = tf.train.Saver(vars_to_restore)
vars_to_save = util.get_vars_to_restore()
saver = tf.train.Saver(vars_to_save + [train_model.global_step],
max_to_keep=MAX_TO_KEEP)
sv = tf.train.Supervisor(logdir=checkpoint_dir, save_summaries_secs=0,
saver=None)
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
with sv.managed_session(config=config) as sess:
if pretrained_ckpt is not None:
logging.info('Restoring pretrained weights from %s', pretrained_ckpt)
pretrain_restorer.restore(sess, pretrained_ckpt)
logging.info('Attempting to resume training from %s...', checkpoint_dir)
checkpoint = tf.train.latest_checkpoint(checkpoint_dir)
logging.info('Last checkpoint found: %s', checkpoint)
if checkpoint:
saver.restore(sess, checkpoint)
logging.info('Training...')
start_time = time.time()
last_summary_time = time.time()
steps_per_epoch = train_model.reader.steps_per_epoch
step = 1
while step <= train_steps:
fetches = {
'train': train_model.train_op,
'global_step': train_model.global_step,
'incr_global_step': train_model.incr_global_step
}
if step % summary_freq == 0:
fetches['loss'] = train_model.total_loss
fetches['summary'] = sv.summary_op
results = sess.run(fetches)
global_step = results['global_step']
if step % summary_freq == 0:
sv.summary_writer.add_summary(results['summary'], global_step)
train_epoch = math.ceil(global_step / steps_per_epoch)
train_step = global_step - (train_epoch - 1) * steps_per_epoch
this_cycle = time.time() - last_summary_time
last_summary_time += this_cycle
logging.info(
'Epoch: [%2d] [%5d/%5d] time: %4.2fs (%ds total) loss: %.3f',
train_epoch, train_step, steps_per_epoch, this_cycle,
time.time() - start_time, results['loss'])
if step % steps_per_epoch == 0:
logging.info('[*] Saving checkpoint to %s...', checkpoint_dir)
saver.save(sess, os.path.join(checkpoint_dir, 'model'),
global_step=global_step)
# Setting step to global_step allows for training for a total of
# train_steps even if the program is restarted during training.
step = global_step + 1
def _run_inference():
"""Runs all images through depth model and saves depth maps."""
ckpt_basename = os.path.basename(FLAGS.model_ckpt)
ckpt_modelname = os.path.basename(os.path.dirname(FLAGS.model_ckpt))
og_dir = FLAGS.output_dir
output_dir = os.path.join(
FLAGS.output_dir,
FLAGS.kitti_video.replace('/', '_') + '_' + ckpt_modelname + '_' +
ckpt_basename)
if not gfile.Exists(output_dir):
gfile.MakeDirs(output_dir)
inference_model = model.Model(is_training=False,
seq_length=FLAGS.seq_length,
batch_size=FLAGS.batch_size,
img_height=FLAGS.img_height,
img_width=FLAGS.img_width)
vars_to_restore = util.get_vars_to_restore(FLAGS.model_ckpt)
saver = tf.train.Saver(vars_to_restore)
sv = tf.train.Supervisor(logdir='/tmp/', saver=None)
with sv.managed_session() as sess:
saver.restore(sess, FLAGS.model_ckpt)
if FLAGS.kitti_video == 'test_files_eigen':
im_files = util.read_text_lines(
util.get_resource_path('dataset/kitti/test_files_eigen.txt'))
im_files = [os.path.join(FLAGS.kitti_dir, f) for f in im_files]
else:
video_path = os.path.join(FLAGS.kitti_dir, FLAGS.kitti_video)
#edit
#im_files = gfile.Glob(os.path.join(video_path, 'image_02/data', '*.png'))
im_files = gfile.Glob(os.path.join(video_path, '*.png')) #delete
#end edit
im_files = [f for f in im_files if 'disp' not in f]
im_files = sorted(im_files)
# regularPictures(im_files, output_dir, inference_model)
for i in range(0, len(im_files), FLAGS.batch_size):
if i % 100 == 0:
logging.info('Generating from %s: %d/%d', ckpt_basename, i,
len(im_files))
inputs = np.zeros(
(FLAGS.batch_size, FLAGS.img_height, FLAGS.img_width, 3),
dtype=np.uint8)
for b in range(FLAGS.batch_size):
idx = i + b
if idx >= len(im_files):
break
im = scipy.misc.imread(im_files[idx],
mode='RGB') #added 2nd arg
inputs[b] = scipy.misc.imresize(
im, (FLAGS.img_height, FLAGS.img_width))
results = inference_model.inference(inputs, sess, mode='depth')
for b in range(FLAGS.batch_size):
idx = i + b
if idx >= len(im_files):
break
if FLAGS.kitti_video == 'test_files_eigen':
depth_path = os.path.join(output_dir, '%03d.png' % idx)
else:
depth_path = os.path.join(output_dir, '%04d.png' % idx)
depth_map = results['depth'][b]
depth_map = np.squeeze(depth_map)
colored_map = _normalize_depth_for_display(depth_map,
cmap=CMAP)
input_float = inputs[b].astype(np.float32) / 255.0
vertical_stack = np.concatenate((input_float, colored_map),
axis=0)
scipy.misc.imsave(depth_path, vertical_stack)
#edit
if FLAGS.kitti_video == 'test_files_eigen':
outPath = os.path.join(og_dir, 'reg_pics',
'%03d.png' % idx)
else:
outPath = os.path.join(og_dir, 'reg_pics',
'%04d.png' % idx)
tempImg = scipy.misc.imread(im_files[0], mode='RGB')
resized_colored_map = scipy.misc.imresize(
colored_map, (len(tempImg), len(tempImg[0])))
scipy.misc.imsave(outPath, resized_colored_map)