default=None,
help='test steps')
parser.add_argument("--corr_type",
dest="corr_type",
type=str,
default="tf",
help="correlation layer realization - 'tf' or 'cuda'")
args = parser.parse_args()
ft3d_dataset = ft3d_filenames(args.dataset_path)
tf.logging.set_verbosity(tf.logging.ERROR)
dispnet = DispNet(mode="test",
ckpt_path=args.checkpoint_path,
dataset=ft3d_dataset,
input_size=INPUT_SIZE,
batch_size=args.batch_size,
corr_type=args.corr_type)
ckpt = tf.train.latest_checkpoint(args.checkpoint_path)
if not ckpt:
logging.error("no checkpoint in provided path found!")
sys.exit()
init_logger(args.checkpoint_path)
log_step = args.log_step
if args.n_steps is None:
N_test = len(ft3d_dataset["TEST"])
else:
N_test = args.n_steps
gpu_options = tf.GPUOptions(allow_growth=True)
right_placeholder = tf.placeholder(dtype=tf.float32, shape=[None, None, 3])
left_input = tf.expand_dims(left_placeholder, axis=0)
right_input = tf.expand_dims(right_placeholder, axis=0)
#left_img = tf.placeholder(dtype=tf.float32,shape=[1,Npne,None,3])
#right_img = tf.placeholder(dtype=tf.float32,)
# build input batch
#left_img_batch, right_img_batch, name_batch, resolution_batch = tf.train.batch([left_img, right_img, left_fn, original_resolution], args.batch_size, num_threads=4, capacity=args.batch_size * 100, allow_smaller_final_batch=True)
# build model
is_corr = args.corr_type != 'none'
dispnet = DispNet(mode="inference",
ckpt_path=args.checkpoint_path,
batch_size=1,
is_corr=is_corr,
corr_type=args.corr_type,
image_ops=[left_input, right_input])
raw_prediction = dispnet.predictions_test[0]
rescaled_prediction = tf.image.resize_images(
raw_prediction,
tf.shape(left_placeholder)[0:2],
method=tf.image.ResizeMethod.NEAREST_NEIGHBOR)
cropped_prediction = tf.image.resize_image_with_crop_or_pad(
rescaled_prediction, target_shape[0], target_shape[1])
gpu_options = tf.GPUOptions(allow_growth=True)
with tf.Session(config=tf.ConfigProto(gpu_options=gpu_options)) as sess:
sess.run(dispnet.init)
print("initialized")
help='batch size')
parser.add_argument("-l", "--log_step", dest="log_step", type=int, default=100,
help='log step size')
parser.add_argument("-s", "--save_step", dest="save_step", type=int, default=5000,
help='save checkpoint step size')
parser.add_argument("-n", "--n_steps", dest="n_steps", type=int, default=500000,
help='test steps')
parser.add_argument('--use_corr', action='store_true', default=False)
parser.add_argument('--weight_schedule', action='store_true', default=False)
args = parser.parse_args()
ft3d_dataset = ft3d_filenames(args.dataset_path)
tf.logging.set_verbosity(tf.logging.ERROR)
dispnet = DispNet(mode="traintest", ckpt_path=args.checkpoint_path, dataset=ft3d_dataset,
batch_size=args.batch_size, is_corr=args.use_corr)
ckpt = tf.train.latest_checkpoint(args.checkpoint_path)
if not ckpt:
if not os.path.exists(args.checkpoint_path):
os.makedirs(args.checkpoint_path)
model_name = "DispNet"
if args.use_corr:
model_name += "Corr1D"
init_logger(args.checkpoint_path, name=model_name)
writer = tf.summary.FileWriter(args.checkpoint_path)
schedule_step = 50000
if args.weight_schedule is True:
weights_schedule = [[0., 0., 0., 0., .2, 1.],
type=str,
metavar="FILE",
help='model checkpoint path')
parser.add_argument("-o",
"--output",
dest="output_path",
required=True,
type=str,
metavar="FILE",
help='path to output frozen model')
parser.add_argument('--use_corr', action='store_true', default=False)
args = parser.parse_args()
dispnet = DispNet(mode="inference",
ckpt_path=args.checkpoint_path,
input_size=INPUT_SIZE,
is_corr=args.use_corr)
ckpt = tf.train.latest_checkpoint(args.checkpoint_path)
gpu_options = tf.GPUOptions(allow_growth=True)
with tf.Session(graph=dispnet.graph,
config=tf.ConfigProto(gpu_options=gpu_options)) as sess:
sess.run(dispnet.init)
dispnet.saver.restore(sess=sess, save_path=ckpt)
print("Restoring from %s" % ckpt)
freeze_graph_def = graph_util.convert_variables_to_constants(
sess, sess.graph.as_graph_def(), ['prediction/conv/BiasAdd'])
if not os.path.exists(args.output_path):
os.makedirs(args.output_path)
help='batch size')
parser.add_argument("-l", "--log_step", dest="log_step", type=int, default=100,
help='log step size')
parser.add_argument("-s", "--save_step", dest="save_step", type=int, default=5000,
help='save checkpoint step size')
parser.add_argument("-n", "--n_steps", dest="n_steps", type=int, default=None,
help='test steps')
parser.add_argument("--corr_type", dest="corr_type", type=str, default="tf",
help="correlation layer realization - 'tf' or 'cuda'")
args = parser.parse_args()
ft3d_dataset = ft3d_filenames(args.dataset_path)
tf.logging.set_verbosity(tf.logging.ERROR)
dispnet = DispNet(mode="traintest", ckpt_path=args.checkpoint_path, dataset=ft3d_dataset,
batch_size=args.batch_size, is_corr=CORR, corr_type="cuda")
ckpt = tf.train.latest_checkpoint(args.checkpoint_path)
if not ckpt:
if not os.path.exists(args.checkpoint_path):
os.mkdir(args.checkpoint_path)
model_name = "DispNet"
if CORR:
model_name += "Corr1D"
init_logger(args.checkpoint_path, name=model_name)
writer = tf.summary.FileWriter(args.checkpoint_path)
schedule_step = 50000
weights_schedule = [[0., 0., 0., 0., .2, 1.],
[0., 0., 0., .2, 1., .5],
[0., 0., .2, 1., .5, 0.],
help="flag to read confidence as a 16 bit png",
action='store_true')
args = parser.parse_args()
dataset = trainingLists_conf(args.training,
args.testing,
kittiGt=args.kittigt,
doublePrecisionConf=args.doubleConf)
tf.logging.set_verbosity(tf.logging.ERROR)
is_corr = args.corr_type != 'none'
dispnet = DispNet(mode="traintest",
ckpt_path=args.checkpoint_path,
dataset=dataset,
batch_size=args.batch_size,
is_corr=is_corr,
corr_type=args.corr_type,
smoothness_lambda=args.smooth,
confidence_th=args.confidence_th)
if not os.path.exists(args.checkpoint_path):
os.mkdir(args.checkpoint_path)
init_logger(args.checkpoint_path)
writer = tf.summary.FileWriter(args.checkpoint_path)
#Flying Things train
# schedule_step = 100000 # ORIGINAL
# weights_schedule = [[0., 0., 0., 0., .2, 1.],
# [0., 0., 0., .2, 1., .5],
# [0., 0., .2, 1., .5, 0.],
# [0., .2, 1., .5, 0., 0.],