def _get_simple_dataset(self) -> ChunkedDataset:
# build a simple dataset with 3 frames
# frame 0:
# agent 0
# agent 1
# agent 2
# frame 1:
# agent 0
# agent 1
# frame 2:
# agent 0
dataset = ChunkedDataset("")
dataset.scenes = np.zeros(1, dtype=SCENE_DTYPE)
dataset.frames = np.zeros(3, dtype=FRAME_DTYPE)
dataset.agents = np.zeros(6, dtype=AGENT_DTYPE)
dataset.scenes[0]["frame_index_interval"] = (0, 3)
dataset.frames["agent_index_interval"] = [(0, 3), (3, 5), (5, 6)]
dataset.agents["track_id"] = [0, 1, 2, 0, 1, 0]
# set properties to something different than 0
dataset.agents["centroid"] = np.random.rand(*dataset.agents["centroid"].shape)
dataset.agents["yaw"] = np.random.rand(*dataset.agents["yaw"].shape)
dataset.agents["extent"] = np.random.rand(*dataset.agents["extent"].shape)
return dataset
def test_compute_mse_error(tmp_path: Path) -> None:
data = ChunkedDataset(path="./l5kit/tests/artefacts/single_scene.zarr")
data.open()
export_zarr_to_ground_truth_csv(data, str(tmp_path / "gt1.csv"), 0, 50, 0.5)
data.open() # avoid double select_agents
export_zarr_to_ground_truth_csv(data, str(tmp_path / "gt2.csv"), 0, 50, 0.5)
err = compute_mse_error_csv(str(tmp_path / "gt1.csv"), str(tmp_path / "gt2.csv"))
assert np.all(err == 0.0)
data_fake = ChunkedDataset(str(tmp_path))
data_fake.scenes = np.asarray(data.scenes).copy()
data_fake.frames = np.asarray(data.frames).copy()
data_fake.agents = np.asarray(data.agents).copy()
data_fake.root = data.root
data_fake.agents["centroid"] += np.random.rand(*data_fake.agents["centroid"].shape)
export_zarr_to_ground_truth_csv(data_fake, str(tmp_path / "gt3.csv"), 0, 50, 0.5)
err = compute_mse_error_csv(str(tmp_path / "gt1.csv"), str(tmp_path / "gt3.csv"))
assert np.any(err > 0.0)
# test invalid conf by removing lines in gt1
with open(str(tmp_path / "gt4.csv"), "w") as fp:
lines = open(str(tmp_path / "gt1.csv")).readlines()
fp.writelines(lines[:-10])
with pytest.raises(ValueError):
compute_mse_error_csv(str(tmp_path / "gt1.csv"), str(tmp_path / "gt4.csv"))
def test_compute_mse_error(tmp_path: Path,
zarr_dataset: ChunkedDataset) -> None:
export_zarr_to_ground_truth_csv(zarr_dataset, str(tmp_path / "gt1.csv"),
10, 50, 0.5)
export_zarr_to_ground_truth_csv(zarr_dataset, str(tmp_path / "gt2.csv"),
10, 50, 0.5)
err = compute_mse_error_csv(str(tmp_path / "gt1.csv"),
str(tmp_path / "gt2.csv"))
assert np.all(err == 0.0)
data_fake = ChunkedDataset(str(tmp_path))
data_fake.scenes = np.asarray(zarr_dataset.scenes).copy()
data_fake.frames = np.asarray(zarr_dataset.frames).copy()
data_fake.agents = np.asarray(zarr_dataset.agents).copy()
data_fake.agents["centroid"] += np.random.rand(
*data_fake.agents["centroid"].shape) * 1e-2
export_zarr_to_ground_truth_csv(data_fake, str(tmp_path / "gt3.csv"), 10,
50, 0.5)
err = compute_mse_error_csv(str(tmp_path / "gt1.csv"),
str(tmp_path / "gt3.csv"))
assert np.any(err > 0.0)
# test invalid conf by removing lines in gt1
with open(str(tmp_path / "gt4.csv"), "w") as fp:
lines = open(str(tmp_path / "gt1.csv")).readlines()
fp.writelines(lines[:-10])
with pytest.raises(ValueError):
compute_mse_error_csv(str(tmp_path / "gt1.csv"),
str(tmp_path / "gt4.csv"))
def get_frames_subset(dataset: ChunkedDataset, frame_start_idx: int,
frame_end_idx: int) -> ChunkedDataset:
"""Get a new dataset with frames between start (included) and end (excluded).
Assumptions:
- the dataset has only 1 scene
- the dataset is in numpy format and not zarr anymore
:param dataset: the single-scene dataset.
:param frame_start_idx: first frame to keep.
:param frame_end_idx: where to stop taking frames (excluded).
"""
if not len(dataset.scenes) == 1:
raise ValueError(
f"dataset should have a single scene, got {len(dataset.scenes)}")
if not isinstance(dataset.agents, np.ndarray):
raise ValueError("dataset agents should be an editable np array")
if not isinstance(dataset.tl_faces, np.ndarray):
raise ValueError("dataset tls should be an editable np array")
if not isinstance(dataset.frames, np.ndarray):
raise ValueError("dataset frames should be an editable np array")
if frame_start_idx >= len(dataset.frames):
raise ValueError(
f"frame start {frame_start_idx} is over the length of the dataset")
if frame_end_idx > len(dataset.frames):
raise ValueError(
f"frame end {frame_end_idx} is over the length of the dataset")
if frame_start_idx >= frame_end_idx:
raise ValueError(
f"end frame {frame_end_idx} should be higher than start {frame_start_idx}"
)
if frame_start_idx < 0:
raise ValueError(f"start frame {frame_start_idx} should be positive")
new_dataset = ChunkedDataset("")
new_dataset.scenes = dataset.scenes.copy()
new_dataset.scenes[0]["start_time"] = dataset.frames[frame_start_idx][
"timestamp"]
new_dataset.scenes[0]["end_time"] = dataset.frames[frame_end_idx -
1]["timestamp"]
new_dataset.frames = dataset.frames[frame_start_idx:frame_end_idx].copy()
new_dataset.scenes[0]["frame_index_interval"] = (0,
len(new_dataset.frames))
agent_slice = get_agents_slice_from_frames(
*dataset.frames[[frame_start_idx, frame_end_idx - 1]])
tls_slice = get_tl_faces_slice_from_frames(
*dataset.frames[[frame_start_idx, frame_end_idx - 1]])
new_dataset.frames["agent_index_interval"] -= new_dataset.frames[
"agent_index_interval"][0, 0]
new_dataset.frames[
"traffic_light_faces_index_interval"] -= new_dataset.frames[
"traffic_light_faces_index_interval"][0, 0]
new_dataset.agents = dataset.agents[agent_slice].copy()
new_dataset.tl_faces = dataset.tl_faces[tls_slice].copy()
return new_dataset
def dataset(tmp_path: Path) -> ChunkedDataset:
dataset = ChunkedDataset(str(tmp_path))
dataset.scenes = np.zeros(1, dtype=dataset.scenes.dtype)
dataset.frames = np.zeros(SCENE_LENGTH, dtype=dataset.frames.dtype)
dataset.agents = np.zeros(SCENE_LENGTH, dtype=dataset.agents.dtype)
dataset.scenes[0]["frame_index_interval"] = (0, SCENE_LENGTH)
for idx in range(len(dataset.frames)):
dataset.frames[idx]["agent_index_interval"] = (idx, idx + 1)
dataset.frames[idx]["timestamp"] = idx
for idx in range(len(dataset.agents)):
# we don't check moving anymore, so the agent can stay still
dataset.agents[idx]["extent"] = (5, 5, 5)
dataset.agents[idx]["yaw"] = 0
dataset.agents[idx]["track_id"] = 1
dataset.agents[idx]["label_probabilities"][3] = 1.0
return dataset
def _mock_dataset() -> ChunkedDataset:
zarr_dt = ChunkedDataset("")
zarr_dt.scenes = np.zeros(1, dtype=SCENE_DTYPE)
zarr_dt.scenes["frame_index_interval"][0] = (0, 4)
zarr_dt.frames = np.zeros(4, dtype=FRAME_DTYPE)
zarr_dt.frames["agent_index_interval"][0] = (0, 3)
zarr_dt.frames["agent_index_interval"][1] = (3, 5)
zarr_dt.frames["agent_index_interval"][2] = (5, 6)
zarr_dt.frames["agent_index_interval"][3] = (6, 6)
zarr_dt.agents = np.zeros(6, dtype=AGENT_DTYPE)
# all agents except the first one are valid
zarr_dt.agents["label_probabilities"][1:, 3] = 1
# FRAME 0
# second agent is close to ego and has id 1
zarr_dt.agents["track_id"][1] = 1
zarr_dt.agents["centroid"][1] = (1, 1)
# third agent is too far and has id 2
zarr_dt.agents["track_id"][2] = 2
zarr_dt.agents["centroid"][2] = (100, 100)
# FRAME 1
# track 1 agent is still close to ego
zarr_dt.agents["track_id"][3] = 1
zarr_dt.agents["centroid"][3] = (1, 2)
# track 2 is now close enough
zarr_dt.agents["track_id"][4] = 2
zarr_dt.agents["centroid"][4] = (1, 1)
# FRAME 2
# track 1 agent is far
zarr_dt.agents["track_id"][5] = 1
zarr_dt.agents["centroid"][5] = (100, 100)
# FRAME 3 is empty
zarr_dt.tl_faces = np.zeros(0, dtype=TL_FACE_DTYPE)
return zarr_dt
def _get_simple_dataset(self) -> ChunkedDataset:
# build a simple dataset with 3 frames
# frame 0:
# agent 0
# agent 1
# agent 2
# frame 1:
# agent 0
# agent 1
# frame 2:
# agent 0
dataset = ChunkedDataset("")
dataset.scenes = np.zeros(1, dtype=SCENE_DTYPE)
dataset.frames = np.zeros(3, dtype=FRAME_DTYPE)
dataset.agents = np.zeros(6, dtype=AGENT_DTYPE)
dataset.scenes[0]["frame_index_interval"] = (0, 3)
dataset.frames["agent_index_interval"] = [(0, 3), (3, 5), (5, 6)]
dataset.agents["track_id"] = [0, 1, 2, 0, 1, 0]
return dataset
def test_mock_dataset_frames_subset() -> None:
zarr_dataset = ChunkedDataset("")
zarr_dataset.scenes = np.zeros(1, dtype=SCENE_DTYPE)
zarr_dataset.scenes[0]["frame_index_interval"] = (0, 4)
zarr_dataset.frames = np.zeros(4, dtype=FRAME_DTYPE)
zarr_dataset.frames["agent_index_interval"] = [(0, 1), (1, 2), (2, 3), (3, 4)]
zarr_dataset.agents = np.zeros(4, dtype=AGENT_DTYPE)
zarr_dataset.agents["track_id"] = np.arange(4)
zarr_dataset.tl_faces = np.zeros(0, dtype=TL_FACE_DTYPE)
frame_start = 1
frame_end = 3
zarr_cut = get_frames_subset(zarr_dataset, frame_start, frame_end)
assert np.all(zarr_cut.agents["track_id"] == [1, 2])
frame_start = 0
frame_end = 3
zarr_cut = get_frames_subset(zarr_dataset, frame_start, frame_end)
assert np.all(zarr_cut.agents["track_id"] == [0, 1, 2])
frame_start = 2
frame_end = 4
zarr_cut = get_frames_subset(zarr_dataset, frame_start, frame_end)
assert np.all(zarr_cut.agents["track_id"] == [2, 3])
def build_dataloader(
cfg: Dict,
split: str,
data_manager: DataManager,
dataset_class: Callable,
rasterizer: Rasterizer,
perturbation: Optional[Perturbation] = None,
combine_scenes: bool = False,
) -> DataLoader:
"""
Function to build a dataloader from a dataset of dataset_class. Note we have to pass rasterizer and
perturbation as the factory functions for those are likely to change between repos.
Args:
cfg (dict): configuration dict
split (str): this will be used to index the cfg to get the correct datasets (train or val currently)
data_manager (DataManager): manager for resolving paths
dataset_class (Callable): a class object (EgoDataset or AgentDataset currently) to build the dataset
rasterizer (Rasterizer): the rasterizer for the dataset
perturbation (Optional[Perturbation]): an optional perturbation object
combine_scenes (bool): if to combine scenes that follow up each other perfectly
Returns:
DataLoader: pytorch Dataloader object built with Concat and Sub datasets
"""
data_loader_cfg = cfg[f"{split}_data_loader"]
datasets = []
for dataset_param in data_loader_cfg["datasets"]:
zarr_dataset_path = data_manager.require(key=dataset_param["key"])
zarr_dataset = ChunkedDataset(path=zarr_dataset_path)
zarr_dataset.open()
if combine_scenes: # possible future deprecation
zarr_dataset.scenes = get_combined_scenes(zarr_dataset.scenes)
# Let's load the zarr dataset with our dataset.
dataset = dataset_class(cfg,
zarr_dataset,
rasterizer,
perturbation=perturbation)
scene_indices = dataset_param["scene_indices"]
scene_subsets = []
if dataset_param["scene_indices"][0] == -1: # TODO replace with empty
scene_subset = Subset(dataset, np.arange(0, len(dataset)))
scene_subsets.append(scene_subset)
else:
for scene_idx in scene_indices:
valid_indices = dataset.get_scene_indices(scene_idx)
scene_subset = Subset(dataset, valid_indices)
scene_subsets.append(scene_subset)
datasets.extend(scene_subsets)
# Let's concatenate the training scenes into one dataset for the data loader to load from.
concat_dataset: ConcatDataset = ConcatDataset(datasets)
# Initialize the data loader that our training loop will iterate on.
batch_size = data_loader_cfg["batch_size"]
shuffle = data_loader_cfg["shuffle"]
num_workers = data_loader_cfg["num_workers"]
dataloader = DataLoader(dataset=concat_dataset,
batch_size=batch_size,
shuffle=shuffle,
num_workers=num_workers)
return dataloader
