def test_unsupervised_depth_problem_cpu(self):
device = "cpu"
sequence_8 = Downloader('08')
if not os.path.exists("./dataset/poses"):
print("Download dataset")
sequence_8.download_sequence()
lengths = (200, 30, 30)
dataset = pykitti.odometry(sequence_8.main_dir,
sequence_8.sequence_id,
frames=range(0, 260, 1))
dataset_manager = DatasetManagerMock(dataset,
lenghts=lengths,
num_workers=WORKERS_COUNT)
model = UnDeepVO(max_depth=2., min_depth=1.0).to(device)
optimizer_manger = OptimizerManager()
criterion = UnsupervisedCriterion(
dataset_manager.get_cameras_calibration(device), 0.1, 1, 0.85)
handler = TrainingProcessHandler(mlflow_tags={"name": "test"})
problem = UnsupervisedDepthProblem(model,
criterion,
optimizer_manger,
dataset_manager,
handler,
batch_size=1,
device=device)
problem.train(1)
def test_get_cameras_calibration(self):
sequence_8 = Downloader('08')
if not os.path.exists("./dataset/poses"):
print("Download dataset")
sequence_8.download_sequence()
lengths = (1, 1, 1)
dataset = pykitti.odometry(sequence_8.main_dir,
sequence_8.sequence_id,
frames=range(0, 3, 1))
dataset_manager = UnsupervisedDatasetManager(dataset,
lenghts=lengths,
num_workers=WORKERS_COUNT)
camera_calibration = dataset_manager.get_cameras_calibration()
self.assertEqual(camera_calibration.left_camera_matrix.shape,
torch.Size([1, 3, 3]))
self.assertEqual(camera_calibration.right_camera_matrix.shape,
torch.Size([1, 3, 3]))
def test_pose_loss(self):
sequence_8 = Downloader('08')
if not os.path.exists("./dataset/poses"):
print("Download dataset")
sequence_8.download_sequence()
dataset = pykitti.odometry(sequence_8.main_dir,
sequence_8.sequence_id,
frames=range(0, 3, 1))
dataset_manager = UnsupervisedDatasetManager(dataset,
lengths=(1, 1, 1),
num_workers=WORKERS_COUNT)
angles = torch.tensor([[1., 1., 1.]]).cuda()
translation = torch.tensor([[0.1, 0.2, 0.3]]).cuda()
pose_loss = PoseLoss(
1, 1,
dataset_manager.get_cameras_calibration().
transform_from_left_to_right)
out = pose_loss(translation, translation, angles, angles)
self.assertEqual(out.shape, torch.Size([]))
self.assertGreaterEqual(out, 0)
self.assertLessEqual(out, 0.01)
def test_dataset_manager(self):
sequence_8 = Downloader('08')
if not os.path.exists("./dataset/poses"):
print("Download dataset")
sequence_8.download_sequence()
lengths = (200, 30, 30)
dataset = pykitti.odometry(sequence_8.main_dir,
sequence_8.sequence_id,
frames=range(0, 260, 1))
dataset_manager = UnsupervisedDatasetManager(dataset,
lenghts=lengths,
num_workers=WORKERS_COUNT)
self.assertEqual(len(dataset_manager.get_train_dataset()), lengths[0])
self.assertEqual(len(dataset_manager.get_test_dataset()), lengths[1])
self.assertEqual(len(dataset_manager.get_validation_dataset()),
lengths[2])
batches = dataset_manager.get_train_batches(20)
for batch in batches:
self.assertEqual(batch["left_current_image"].shape,
torch.Size([20, 3, 128, 384]))
self.assertEqual(batch["right_current_image"].shape,
torch.Size([20, 3, 128, 384]))
self.assertEqual(batch["left_next_image"].shape,
torch.Size([20, 3, 128, 384]))
self.assertEqual(batch["right_next_image"].shape,
torch.Size([20, 3, 128, 384]))
self.assertEqual(batch["current_position"].shape,
torch.Size([20, 3]))
self.assertEqual(batch["current_angle"].shape, torch.Size([20, 3]))
self.assertEqual(batch["next_position"].shape, torch.Size([20, 3]))
self.assertEqual(batch["next_angle"].shape, torch.Size([20, 3]))
self.assertEqual(batch["delta_position"].shape, torch.Size([20,
3]))
self.assertEqual(batch["delta_angle"].shape, torch.Size([20, 3]))
self.assertEqual(batch["right_next_image"].dtype, torch.float32)
break
batches = dataset_manager.get_validation_batches(20)
for batch in batches:
self.assertEqual(batch["left_current_image"].shape,
torch.Size([20, 3, 128, 384]))
self.assertEqual(batch["right_current_image"].shape,
torch.Size([20, 3, 128, 384]))
self.assertEqual(batch["left_next_image"].shape,
torch.Size([20, 3, 128, 384]))
self.assertEqual(batch["right_next_image"].shape,
torch.Size([20, 3, 128, 384]))
self.assertEqual(batch["current_position"].shape,
torch.Size([20, 3]))
self.assertEqual(batch["current_angle"].shape, torch.Size([20, 3]))
self.assertEqual(batch["next_position"].shape, torch.Size([20, 3]))
self.assertEqual(batch["next_angle"].shape, torch.Size([20, 3]))
self.assertEqual(batch["delta_position"].shape, torch.Size([20,
3]))
self.assertEqual(batch["delta_angle"].shape, torch.Size([20, 3]))
break
batches = dataset_manager.get_test_batches(20)
for batch in batches:
self.assertEqual(batch["left_current_image"].shape,
torch.Size([20, 3, 128, 384]))
self.assertEqual(batch["right_current_image"].shape,
torch.Size([20, 3, 128, 384]))
self.assertEqual(batch["left_next_image"].shape,
torch.Size([20, 3, 128, 384]))
self.assertEqual(batch["right_next_image"].shape,
torch.Size([20, 3, 128, 384]))
self.assertEqual(batch["current_position"].shape,
torch.Size([20, 3]))
self.assertEqual(batch["current_angle"].shape, torch.Size([20, 3]))
self.assertEqual(batch["next_position"].shape, torch.Size([20, 3]))
self.assertEqual(batch["next_angle"].shape, torch.Size([20, 3]))
self.assertEqual(batch["delta_position"].shape, torch.Size([20,
3]))
self.assertEqual(batch["delta_angle"].shape, torch.Size([20, 3]))
break
default="cuda:0",
type=str,
help='whether to use resnet or not')
parser.add_argument('-model_path',
default="",
type=str,
help='whether to use resnet or not')
args = parser.parse_args()
MAIN_DIR = args.main_dir
lengths = args.split
problem = None
if args.method == "unsupervised":
sequence_8 = Downloader('08')
if not os.path.exists("./dataset/poses"):
print("Download dataset")
sequence_8.download_sequence()
dataset = pykitti.odometry(MAIN_DIR, '08', frames=range(*args.frames_range))
dataset_manager = UnsupervisedDatasetManager(dataset, lengths=lengths)
model = UnDeepVO(args.max_depth, args.min_depth, args.resnet).to(args.device)
if args.model_path != "":
model.load_state_dict(torch.load(args.model_path, map_location=args.device))
criterion = UnsupervisedCriterion(dataset_manager.get_cameras_calibration(args.device),
args.lambda_position,
args.lambda_rotation,
args.lambda_s,
args.lambda_disparity,