def test_simulation_dataset_build(zarr_cat_dataset: ChunkedDataset,
dmg: LocalDataManager, cfg: dict,
tmp_path: Path) -> None:
# modify one frame to ensure everything works also when scenes are different
zarr_cat_dataset.frames = np.asarray(zarr_cat_dataset.frames)
for scene_idx in range(len(zarr_cat_dataset.scenes)):
frame_slice = get_frames_slice_from_scenes(zarr_cat_dataset.scenes)
zarr_cat_dataset.frames[
frame_slice.start]["ego_translation"] += np.random.randn(3)
rasterizer = build_rasterizer(cfg, dmg)
ego_dataset = EgoDataset(cfg, zarr_cat_dataset, rasterizer)
sim_cfg = SimulationConfig(use_ego_gt=True,
use_agents_gt=True,
disable_new_agents=False,
distance_th_far=30,
distance_th_close=10)
# we should be able to create the same object by using both constructor and factory
scene_indices = list(range(len(zarr_cat_dataset.scenes)))
scene_dataset_batch: Dict[int, EgoDataset] = {}
for scene_idx in scene_indices:
scene_dataset = ego_dataset.get_scene_dataset(scene_idx)
scene_dataset_batch[scene_idx] = scene_dataset
sim_1 = SimulationDataset(scene_dataset_batch, sim_cfg)
sim_2 = SimulationDataset.from_dataset_indices(ego_dataset, scene_indices,
sim_cfg)
for (k_1, v_1), (k_2, v_2) in zip(sim_1.scene_dataset_batch.items(),
sim_2.scene_dataset_batch.items()):
assert k_1 == k_2
assert np.allclose(v_1.dataset.frames["ego_translation"],
v_2.dataset.frames["ego_translation"])
def test_get_frame_indices_ego(frame_idx: int, zarr_dataset: ChunkedDataset,
dmg: LocalDataManager, cfg: dict) -> None:
cfg["raster_params"]["map_type"] = "box_debug"
rasterizer = build_rasterizer(cfg, dmg)
dataset = EgoDataset(cfg, zarr_dataset, rasterizer)
frame_indices = dataset.get_frame_indices(frame_idx)
# this should be only one and match the index of the frame (i.e. it should be frame_idx)
assert frame_indices[0] == frame_idx
def test_get_scene_indices_ego(scene_idx: int, zarr_dataset: ChunkedDataset,
dmg: LocalDataManager, cfg: dict) -> None:
cfg["raster_params"]["map_type"] = "box_debug"
rasterizer = build_rasterizer(cfg, dmg)
dataset = EgoDataset(cfg, zarr_dataset, rasterizer)
scene_indices = dataset.get_scene_indices(scene_idx)
frame_slice = get_frames_slice_from_scenes(zarr_dataset.scenes[scene_idx])
assert scene_indices[0] == frame_slice.start
assert scene_indices[-1] == frame_slice.stop - 1
def test_same_displacement(
cfg: dict,
zarr_dataset: ChunkedDataset,
base_displacement: np.ndarray,
raster_size: tuple,
ego_center: tuple,
pixel_size: tuple,
) -> None:
cfg["raster_params"]["raster_size"] = raster_size
cfg["raster_params"]["pixel_size"] = np.asarray(pixel_size)
cfg["raster_params"]["ego_center"] = np.asarray(ego_center)
render_context = RenderContext(
np.asarray(raster_size),
np.asarray(pixel_size),
np.asarray(ego_center),
set_origin_to_bottom=cfg["raster_params"]["set_origin_to_bottom"],
)
dataset = EgoDataset(
cfg,
zarr_dataset,
StubRasterizer(render_context),
)
data = dataset[0]
assert np.allclose(data["target_positions"], base_displacement)
def test_simulation_agents_mock_insert(dmg: LocalDataManager, cfg: dict,
tmp_path: Path) -> None:
zarr_dataset = _mock_dataset()
rasterizer = build_rasterizer(cfg, dmg)
ego_dataset = EgoDataset(cfg, zarr_dataset, rasterizer)
sim_cfg = SimulationConfig(use_ego_gt=True,
use_agents_gt=True,
disable_new_agents=True,
distance_th_far=100,
distance_th_close=10)
dataset = SimulationDataset.from_dataset_indices(ego_dataset, [0], sim_cfg)
_ = dataset.rasterise_agents_frame_batch(0)
# insert (0, 1) in following frames
next_agent = np.zeros(1, dtype=AGENT_DTYPE)
next_agent["centroid"] = (-1, -1)
next_agent["yaw"] = -0.5
next_agent["track_id"] = 1
next_agent["extent"] = (1, 1, 1)
next_agent["label_probabilities"][:, 3] = 1
for frame_idx in [1, 2, 3]:
dataset.set_agents(frame_idx, {(0, 1): next_agent})
agents_dict = dataset.rasterise_agents_frame_batch(frame_idx)
assert len(agents_dict) == 1 and (0, 1) in agents_dict
assert np.allclose(agents_dict[(0, 1)]["centroid"], (-1, -1))
assert np.allclose(agents_dict[(0, 1)]["yaw"], -0.5)
def test_simulation_agents_mock_disable(dmg: LocalDataManager, cfg: dict,
tmp_path: Path) -> None:
zarr_dataset = _mock_dataset()
rasterizer = build_rasterizer(cfg, dmg)
ego_dataset = EgoDataset(cfg, zarr_dataset, rasterizer)
sim_cfg = SimulationConfig(use_ego_gt=True,
use_agents_gt=True,
disable_new_agents=True,
distance_th_far=100,
distance_th_close=10)
dataset = SimulationDataset.from_dataset_indices(ego_dataset, [0], sim_cfg)
# nothing should be tracked
assert len(dataset._agents_tracked) == 0
agents_dict = dataset.rasterise_agents_frame_batch(0)
# only (0, 1) should be in
assert len(agents_dict) == 1 and (0, 1) in agents_dict
assert len(dataset._agents_tracked) == 1
agents_dict = dataset.rasterise_agents_frame_batch(1)
# again, only (0, 1) should be in
assert len(agents_dict) == 1
assert (0, 1) in agents_dict
assert len(dataset._agents_tracked) == 1
agents_dict = dataset.rasterise_agents_frame_batch(2)
assert len(agents_dict) == 0
assert len(dataset._agents_tracked) == 0
def test_simulation_agents(zarr_cat_dataset: ChunkedDataset,
dmg: LocalDataManager, cfg: dict,
tmp_path: Path) -> None:
rasterizer = build_rasterizer(cfg, dmg)
scene_indices = list(range(len(zarr_cat_dataset.scenes)))
ego_dataset = EgoDataset(cfg, zarr_cat_dataset, rasterizer)
sim_cfg = SimulationConfig(use_ego_gt=True,
use_agents_gt=True,
disable_new_agents=False,
distance_th_far=100,
distance_th_close=30)
dataset = SimulationDataset.from_dataset_indices(ego_dataset,
scene_indices, sim_cfg)
# nothing should be tracked
assert len(dataset._agents_tracked) == 0
agents_dict = dataset.rasterise_agents_frame_batch(0)
# we should have the same agents in each scene
for k in agents_dict:
assert (0, k[1]) in agents_dict
# now everything should be tracked
assert len(dataset._agents_tracked) == len(agents_dict)
def test_perturbation_is_applied(perturb_prob: float) -> None:
cfg = load_config_data("./l5kit/tests/artefacts/config.yaml")
zarr_dataset = ChunkedDataset(path="./l5kit/tests/artefacts/single_scene.zarr")
zarr_dataset.open()
dm = LocalDataManager("./l5kit/tests/artefacts/")
rasterizer = build_rasterizer(cfg, dm)
dataset = EgoDataset(cfg, zarr_dataset, rasterizer, None) # no perturb
data_no_perturb = dataset[0]
# note we cannot change the object we already have as a partial is built at init time
perturb = AckermanPerturbation(ReplayRandomGenerator(np.asarray([[4.0, 0.33]])), perturb_prob=perturb_prob)
dataset = EgoDataset(cfg, zarr_dataset, rasterizer, perturb) # perturb
data_perturb = dataset[0]
assert np.linalg.norm(data_no_perturb["target_positions"] - data_perturb["target_positions"]) > 0
assert np.linalg.norm(data_no_perturb["target_yaws"] - data_perturb["target_yaws"]) > 0
def test_perturbation_is_applied(perturb_prob: float, dmg: LocalDataManager,
cfg: dict,
zarr_dataset: ChunkedDataset) -> None:
rasterizer = build_rasterizer(cfg, dmg)
dataset = EgoDataset(cfg, zarr_dataset, rasterizer, None) # no perturb
data_no_perturb = dataset[0]
# note we cannot change the object we already have as a partial is built at init time
perturb = AckermanPerturbation(ReplayRandomGenerator(
np.asarray([[4.0, 0.33]])),
perturb_prob=perturb_prob)
dataset = EgoDataset(cfg, zarr_dataset, rasterizer, perturb) # perturb
data_perturb = dataset[0]
assert np.linalg.norm(data_no_perturb["target_positions"] -
data_perturb["target_positions"]) > 0
assert np.linalg.norm(data_no_perturb["target_yaws"] -
data_perturb["target_yaws"]) > 0
def get_train_dataloaders(cfg, dm):
"""Modified from L5Kit"""
train_cfg = cfg["train_data_loader"]
rasterizer = build_rasterizer(cfg, dm)
train_zarr = ChunkedDataset(dm.require(train_cfg["key"])).open()
train_dataset = AgentDataset(cfg, train_zarr, rasterizer)
train_dataloader = DataLoader(train_dataset, shuffle=train_cfg["shuffle"], batch_size=train_cfg["batch_size"],
num_workers=train_cfg["num_workers"])
train_dataset_ego = EgoDataset(cfg, train_zarr, rasterizer)
train_dataloader_ego = DataLoader(train_dataset_ego, shuffle=train_cfg["shuffle"], batch_size=train_cfg["batch_size"],
num_workers=train_cfg["num_workers"])
return train_dataset, train_dataset_ego, train_dataloader, train_dataloader_ego
def __init__(self):
print("Visualization Class initialized.")
# get config
self.cfg = load_config_data("/mnt/extra/kaggle/competitions/2020lyft/ProjectLyft/Modules/visualisation_config.yaml")
print(self.cfg)
dm = LocalDataManager()
self.dataset_path = dm.require(self.cfg["val_data_loader"]["key"])
self.zarr_dataset = ChunkedDataset(self.dataset_path)
self.zarr_dataset.open()
# Dataset package
self.rast = build_rasterizer(self.cfg, dm)
self.dataset = EgoDataset(self.cfg, self.zarr_dataset, self.rast)
def test_invalid_simulation_dataset(zarr_cat_dataset: ChunkedDataset,
dmg: LocalDataManager, cfg: dict,
tmp_path: Path) -> None:
rasterizer = build_rasterizer(cfg, dmg)
scene_indices = [0, len(zarr_cat_dataset.scenes)]
ego_dataset = EgoDataset(cfg, zarr_cat_dataset, rasterizer)
sim_cfg = SimulationConfig(use_ego_gt=True,
use_agents_gt=True,
disable_new_agents=False,
distance_th_far=30,
distance_th_close=10)
with pytest.raises(ValueError):
SimulationDataset.from_dataset_indices(ego_dataset, scene_indices,
sim_cfg)
def base_displacement(zarr_dataset: ChunkedStateDataset) -> np.ndarray:
cfg = load_config_data("./l5kit/tests/artefacts/config.yaml")
cfg["raster_params"]["raster_size"] = (100, 100)
cfg["raster_params"]["ego_center"] = np.asarray((0.5, 0.5))
cfg["raster_params"]["pixel_size"] = np.asarray((0.25, 0.25))
dataset = EgoDataset(
cfg,
zarr_dataset,
StubRasterizer(
cfg["raster_params"]["raster_size"],
cfg["raster_params"]["pixel_size"],
cfg["raster_params"]["ego_center"],
0.5,
),
)
data = dataset[0]
return data["target_positions"]
def test_simulation_ego(zarr_cat_dataset: ChunkedDataset,
dmg: LocalDataManager, cfg: dict,
tmp_path: Path) -> None:
rasterizer = build_rasterizer(cfg, dmg)
scene_indices = list(range(len(zarr_cat_dataset.scenes)))
ego_dataset = EgoDataset(cfg, zarr_cat_dataset, rasterizer)
sim_cfg = SimulationConfig(use_ego_gt=True,
use_agents_gt=True,
disable_new_agents=False,
distance_th_far=30,
distance_th_close=10)
dataset = SimulationDataset.from_dataset_indices(ego_dataset,
scene_indices, sim_cfg)
# this also ensure order is checked
assert list(dataset.scene_dataset_batch.keys()) == scene_indices
# ensure we can call the aggregated get frame
out_0 = dataset.rasterise_frame_batch(0)
assert len(out_0) == len(scene_indices)
out_last = dataset.rasterise_frame_batch(len(dataset) - 1)
assert len(out_last) == len(scene_indices)
with pytest.raises(IndexError):
_ = dataset.rasterise_frame_batch(len(dataset))
# ensure we can set the ego in multiple frames for all scenes
frame_indices = np.random.randint(0, len(dataset), 10)
for frame_idx in frame_indices:
mock_tr = np.random.rand(len(scene_indices), 12, 2)
mock_yaw = np.random.rand(len(scene_indices), 12)
dataset.set_ego(frame_idx, 0, mock_tr, mock_yaw)
for scene_idx in scene_indices:
scene_zarr = dataset.scene_dataset_batch[scene_idx].dataset
ego_tr = scene_zarr.frames["ego_translation"][frame_idx]
ego_yaw = rotation33_as_yaw(
scene_zarr.frames["ego_rotation"][frame_idx])
assert np.allclose(mock_tr[scene_idx, 0], ego_tr[:2])
assert np.allclose(mock_yaw[scene_idx, 0], ego_yaw)
def test_coordinates_straight_road(zarr_dataset: ChunkedDataset,
cfg: dict) -> None:
# on a straight road `target_positions` should increase on x only
render_context = RenderContext(
np.asarray(cfg["raster_params"]["raster_size"]),
np.asarray(cfg["raster_params"]["pixel_size"]),
np.asarray(cfg["raster_params"]["ego_center"]),
)
dataset = EgoDataset(
cfg,
zarr_dataset,
StubRasterizer(
render_context,
0.5,
),
)
# get first prediction and first 50 centroids
centroids = []
preds = []
preds_world = []
for idx in range(50):
data = dataset[idx]
if idx == 0:
preds = data["target_positions"]
preds_world = transform_points(
preds, np.linalg.inv(data["agent_from_world"]))
centroids.append(data["centroid"][:2])
centroids = np.stack(centroids)
# compute XY variances for preds and centroids
var_preds = np.var(preds, 0, ddof=1)
var_centroids = np.var(centroids, 0, ddof=1)
assert var_preds[1] / var_preds[
0] < 0.001 # variance on Y is way lower than on X
assert var_centroids[1] / var_centroids[
0] > 0.9 # variance on Y is similar to X
# check similarity between coordinates
assert np.allclose(preds_world[:-1], centroids[1:])
def generate_eval_dataset(cfg, dm, rasterizer):
eval_cfg = cfg["test_data_loader"]
eval_dir = shutil.copytree(dm.require(eval_cfg["key"]), '/tmp/lyft/test.zarr')
eval_cfg = cfg["test_data_loader"]
num_frames_to_chop = 50
eval_base_path = create_chopped_dataset(eval_dir, cfg["raster_params"]["filter_agents_threshold"],
num_frames_to_chop, cfg["model_params"]["future_num_frames"], MIN_FUTURE_STEPS)
eval_zarr_path = str(Path(eval_base_path) / "test.zarr")
eval_mask_path = str(Path(eval_base_path) / "mask.npz")
eval_gt_path = str(Path(eval_base_path) / "gt.csv")
eval_zarr = ChunkedDataset(eval_zarr_path).open()
eval_mask = np.load(eval_mask_path)["arr_0"]
# ===== INIT DATASET AND LOAD MASK
eval_dataset = AgentDataset(cfg, eval_zarr, rasterizer, agents_mask=eval_mask)
eval_dataloader = DataLoader(eval_dataset, shuffle=eval_cfg["shuffle"], batch_size=eval_cfg["batch_size"],
num_workers=eval_cfg["num_workers"])
eval_dataset_ego = EgoDataset(cfg, eval_zarr, rasterizer)
return eval_dataset, eval_dataloader, eval_dataset_ego, eval_gt_path
def from_dataset_indices(dataset: EgoDataset, scene_indices: List[int],
sim_cfg: SimulationConfig) -> "SimulationDataset":
"""Create a SimulationDataset by picking indices from the provided dataset
:param dataset: the EgoDataset
:param scene_indices: scenes from the EgoDataset to pick
:param sim_cfg: a simulation config
:return: the new SimulationDataset
"""
if len(np.unique(scene_indices)) != len(scene_indices):
raise ValueError(f"can't simulate repeated scenes: {scene_indices}")
if np.any(np.asarray(scene_indices) >= len(dataset.dataset.scenes)):
raise ValueError(
f"can't pick indices {scene_indices} from dataset with length: {len(dataset.dataset.scenes)}")
scene_dataset_batch: Dict[int, EgoDataset] = {} # dicts preserve insertion order
for scene_idx in scene_indices:
scene_dataset = dataset.get_scene_dataset(scene_idx)
scene_dataset_batch[scene_idx] = scene_dataset
return SimulationDataset(scene_dataset_batch, sim_cfg)
def __init__(
self,
cfg: dict,
zarr_dataset: ChunkedDataset,
rasterizer: Rasterizer,
perturbation: Optional[Perturbation] = None,
agents_mask: Optional[np.ndarray] = None,
min_frame_history: int = MIN_FRAME_HISTORY,
min_frame_future: int = MIN_FRAME_FUTURE,
override_sample_function_name: str = "",
):
assert perturbation is None, "AgentDataset does not support perturbation (yet)"
self.cfg = cfg
self.ego_dataset = EgoDataset(cfg, zarr_dataset, rasterizer, perturbation)
self.get_frame_arguments = self.load_get_frame_arguments(agents_mask, min_frame_history, min_frame_future)
if override_sample_function_name != "":
print("override_sample_function_name", override_sample_function_name)
if override_sample_function_name == "generate_agent_sample_tl_history":
self.ego_dataset.sample_function = create_generate_agent_sample_tl_history_partial(cfg, rasterizer)
elif override_sample_function_name == "generate_agent_sample_fixing_yaw":
self.ego_dataset.sample_function = create_generate_agent_sample_fixing_yaw_partial(cfg, rasterizer)
def test_same_displacement(
zarr_dataset: ChunkedStateDataset,
base_displacement: np.ndarray,
raster_size: tuple,
ego_center: tuple,
pixel_size: tuple,
) -> None:
cfg = load_config_data("./l5kit/tests/artefacts/config.yaml")
cfg["raster_params"]["raster_size"] = raster_size
cfg["raster_params"]["ego_center"] = np.asarray(ego_center)
cfg["raster_params"]["pixel_size"] = np.asarray(pixel_size)
dataset = EgoDataset(
cfg,
zarr_dataset,
StubRasterizer(
cfg["raster_params"]["raster_size"],
cfg["raster_params"]["pixel_size"],
cfg["raster_params"]["ego_center"],
0.5,
),
)
data = dataset[0]
assert np.allclose(data["target_positions"], base_displacement)
def test_same_displacement(
cfg: dict,
zarr_dataset: ChunkedDataset,
base_displacement: np.ndarray,
raster_size: tuple,
ego_center: tuple,
pixel_size: tuple,
) -> None:
cfg["raster_params"]["raster_size"] = raster_size
cfg["raster_params"]["ego_center"] = np.asarray(ego_center)
cfg["raster_params"]["pixel_size"] = np.asarray(pixel_size)
dataset = EgoDataset(
cfg,
zarr_dataset,
StubRasterizer(
cfg["raster_params"]["raster_size"],
cfg["raster_params"]["pixel_size"],
cfg["raster_params"]["ego_center"],
0.5,
),
)
data = dataset[0]
assert np.allclose(data["target_positions"], base_displacement)
# Prepare all rasterizer and EgoDataset for each rasterizer
rasterizer_dict = {}
dataset_dict = {}
rasterizer_type_list = [
"py_satellite", "satellite_debug", "py_semantic", "semantic_debug",
"box_debug", "stub_debug"
]
for i, key in enumerate(rasterizer_type_list):
# print("key", key)
cfg["raster_params"]["map_type"] = key
rasterizer_dict[key] = build_rasterizer(cfg, dm)
dataset_dict[key] = EgoDataset(cfg, zarr_dataset, rasterizer_dict[key])
# default lane color is "light yellow" (255, 217, 82).
# green, yellow, red color on lane is to show trafic light condition.
# orange box represents crosswalk
fig, axes = plt.subplots(2, 3, figsize=(15, 10))
axes = axes.flatten()
for i, key in enumerate([
"stub_debug", "satellite_debug", "semantic_debug", "box_debug",
"py_satellite", "py_semantic"
]):
visualize_rgb_image(dataset_dict[key],
index=0,
title=f"{key}: {type(rasterizer_dict[key]).__name__}",
ax=axes[i])
# Save Metric
np.save(metric_path, metrics)
######################## Plot Prediction Tractories ##################################
model.eval()
torch.set_grad_enabled(False)
# Uncomment to choose satelliter or semantic rasterizer
# validate_cfg["raster_params"]["map_type"] = "py_satellite"
validate_cfg["raster_params"]["map_type"] = "py_semantic"
rast = build_rasterizer(validate_cfg, dm)
eval_ego_dataset = EgoDataset(validate_cfg, valid_dataset.dataset, rast)
num_frames = 2 # randomly pick _ frames
random_frames = np.random.randint(0, len(eval_ego_dataset) - 1, (num_frames, ))
for frame_number in random_frames:
agent_indices = valid_dataset.get_frame_indices(frame_number)
if not len(agent_indices):
continue
# get AV point-of-view frame
data_ego = eval_ego_dataset[frame_number]
im_ego = rasterizer.to_rgb(data_ego["image"].transpose(1, 2, 0))
center = np.asarray(validate_cfg["raster_params"]["ego_center"]
) * validate_cfg["raster_params"]["raster_size"]
predicted_positions = []
pred_path = 'submission1.csv'
write_pred_csv(pred_path,
timestamps=np.concatenate(timestamps),
track_ids=np.concatenate(agent_ids),
coords=np.concatenate(future_coords_offsets_pd),
confs=np.concatenate(confidences_list))
metrics = compute_metrics_csv(
eval_gt_path, pred_path, [neg_multi_log_likelihood, time_displace])
for metric_name, metric_mean in metrics.items():
print(metric_name, metric_mean)
gt_rows = {}
for row in read_gt_csv(eval_gt_path):
gt_rows[row["track_id"] + row["timestamp"]] = row["coord"]
eval_ego_dataset = EgoDataset(cfg, eval_dataset.dataset, rasterizer)
for frame_number in range(
99, len(eval_zarr.frames),
100): # start from last frame of scene_0 and increase by 100
agent_indices = eval_dataset.get_frame_indices(frame_number)
if not len(agent_indices):
continue
# get AV point-of-view frame
data_ego = eval_ego_dataset[frame_number]
im_ego = rasterizer.to_rgb(data_ego["image"].transpose(1, 2, 0))
center = np.asarray(cfg["raster_params"]["ego_center"]
) * cfg["raster_params"]["raster_size"]
predicted_positions = []
zarr_dataset.open()
print(zarr_dataset)
frames = zarr_dataset.frames
# This is much faster!
coords = frames["ego_translation"][:, :2]
plt.scatter(coords[:, 0], coords[:, 1], marker='.')
axes = plt.gca()
axes.set_xlim([-2500, 1600])
axes.set_ylim([-2500, 1600])
plt.title("ego_translation of frames")
semantic_rasterizer = build_rasterizer(cfg, dm)
semantic_dataset = EgoDataset(cfg, zarr_dataset, semantic_rasterizer)
def visualize_trajectory(
dataset,
index,
title="target_positions movement with draw_trajectory"):
data = dataset[index]
im = data["image"].transpose(1, 2, 0)
im = dataset.rasterizer.to_rgb(im)
target_positions_pixels = transform_points(
data["target_positions"] + data["centroid"][:2],
data["world_to_image"])
draw_trajectory(im, target_positions_pixels, data["target_yaws"],
TARGET_POINTS_COLOR)
dataset_path = dm.require(cfg["val_data_loader"]["key"])
zarr_dataset = ChunkedDataset(dataset_path)
zarr_dataset.open()
print(zarr_dataset)
# %% [markdown]
# Now, however it's time for us to look at the scenes and analyze them in depth. Theoretically, we could create a nifty little data-loader to do some heavy lifting for us.
# %% [code] {"_kg_hide-input":true}
import numpy as np
from IPython.display import display, clear_output
import PIL
cfg["raster_params"]["map_type"] = "py_semantic"
rast = build_rasterizer(cfg, dm)
dataset = EgoDataset(cfg, zarr_dataset, rast)
scene_idx = 2
indexes = dataset.get_scene_indices(scene_idx)
images = []
for idx in indexes:
data = dataset[idx]
im = data["image"].transpose(1, 2, 0)
im = dataset.rasterizer.to_rgb(im)
target_positions_pixels = transform_points(data["target_positions"] + data["centroid"][:2], data["world_to_image"])
center_in_pixels = np.asarray(cfg["raster_params"]["ego_center"]) * cfg["raster_params"]["raster_size"]
draw_trajectory(im, target_positions_pixels, data["target_yaws"], TARGET_POINTS_COLOR)
clear_output(wait=True)
#display(PIL.Image.fromarray(im[::-1]))
def ego_cat_dataset(cfg: dict, dmg: LocalDataManager,
zarr_cat_dataset: ChunkedDataset) -> EgoDataset:
rasterizer = build_rasterizer(cfg, dmg)
return EgoDataset(cfg, zarr_cat_dataset, rasterizer)
def __init__(self,
env_config_path: Optional[str] = None,
dmg: Optional[LocalDataManager] = None,
sim_cfg: Optional[SimulationConfig] = None,
train: bool = True,
reward: Optional[Reward] = None,
cle: bool = True,
rescale_action: bool = True,
use_kinematic: bool = False,
kin_model: Optional[KinematicModel] = None,
reset_scene_id: Optional[int] = None,
return_info: bool = False,
randomize_start: bool = True) -> None:
"""Constructor method
"""
super(L5Env, self).__init__()
# Required to register environment
if env_config_path is None:
return
# env config
dm = dmg if dmg is not None else LocalDataManager(None)
cfg = load_config_data(env_config_path)
self.step_time = cfg["model_params"]["step_time"]
# rasterisation
rasterizer = build_rasterizer(cfg, dm)
raster_size = cfg["raster_params"]["raster_size"][0]
n_channels = rasterizer.num_channels()
# load dataset of environment
self.train = train
self.overfit = cfg["gym_params"]["overfit"]
self.randomize_start_frame = randomize_start
if self.train or self.overfit:
loader_key = cfg["train_data_loader"]["key"]
else:
loader_key = cfg["val_data_loader"]["key"]
dataset_zarr = ChunkedDataset(dm.require(loader_key)).open()
self.dataset = EgoDataset(cfg, dataset_zarr, rasterizer)
# Define action and observation space
# Continuous Action Space: gym.spaces.Box (X, Y, Yaw * number of future states)
self.action_space = spaces.Box(low=-1, high=1, shape=(3, ))
# Observation Space: gym.spaces.Dict (image: [n_channels, raster_size, raster_size])
obs_shape = (n_channels, raster_size, raster_size)
self.observation_space = spaces.Dict({
'image':
spaces.Box(low=0, high=1, shape=obs_shape, dtype=np.float32)
})
# Simulator Config within Gym
self.sim_cfg = sim_cfg if sim_cfg is not None else SimulationConfigGym(
)
self.simulator = ClosedLoopSimulator(self.sim_cfg,
self.dataset,
device=torch.device("cpu"),
mode=ClosedLoopSimulatorModes.GYM)
self.reward = reward if reward is not None else L2DisplacementYawReward(
)
self.max_scene_id = cfg["gym_params"]["max_scene_id"]
if not self.train:
self.max_scene_id = cfg["gym_params"]["max_val_scene_id"]
self.randomize_start_frame = False
if self.overfit:
self.overfit_scene_id = cfg["gym_params"]["overfit_id"]
self.randomize_start_frame = False
self.cle = cle
self.rescale_action = rescale_action
self.use_kinematic = use_kinematic
if self.use_kinematic:
self.kin_model = kin_model if kin_model is not None else UnicycleModel(
)
self.kin_rescale = self._get_kin_rescale_params()
else:
self.non_kin_rescale = self._get_non_kin_rescale_params()
# If not None, reset_scene_id is the scene_id that will be rolled out when reset is called
self.reset_scene_id = reset_scene_id
if self.overfit:
self.reset_scene_id = self.overfit_scene_id
# flag to decide whether to return any info at end of episode
# helps to limit the IPC
self.return_info = return_info
self.seed()
class L5Env(gym.Env):
"""Custom Environment of L5 Kit that can be registered in OpenAI Gym.
:param env_config_path: path to the L5Kit environment configuration file
:param dmg: local data manager object
:param simulation_cfg: configuration of the L5Kit closed loop simulator
:param train: flag to determine whether to use train or validation dataset
:param reward: calculates the reward for the gym environment
:param cle: flag to enable close loop environment updates
:param rescale_action: flag to rescale the model action back to the un-normalized action space
:param use_kinematic: flag to use the kinematic model
:param kin_model: the kinematic model
:param return_info: flag to return info when a episode ends
:param randomize_start: flag to randomize the start frame of episode
"""
def __init__(self,
env_config_path: Optional[str] = None,
dmg: Optional[LocalDataManager] = None,
sim_cfg: Optional[SimulationConfig] = None,
train: bool = True,
reward: Optional[Reward] = None,
cle: bool = True,
rescale_action: bool = True,
use_kinematic: bool = False,
kin_model: Optional[KinematicModel] = None,
reset_scene_id: Optional[int] = None,
return_info: bool = False,
randomize_start: bool = True) -> None:
"""Constructor method
"""
super(L5Env, self).__init__()
# Required to register environment
if env_config_path is None:
return
# env config
dm = dmg if dmg is not None else LocalDataManager(None)
cfg = load_config_data(env_config_path)
self.step_time = cfg["model_params"]["step_time"]
# rasterisation
rasterizer = build_rasterizer(cfg, dm)
raster_size = cfg["raster_params"]["raster_size"][0]
n_channels = rasterizer.num_channels()
# load dataset of environment
self.train = train
self.overfit = cfg["gym_params"]["overfit"]
self.randomize_start_frame = randomize_start
if self.train or self.overfit:
loader_key = cfg["train_data_loader"]["key"]
else:
loader_key = cfg["val_data_loader"]["key"]
dataset_zarr = ChunkedDataset(dm.require(loader_key)).open()
self.dataset = EgoDataset(cfg, dataset_zarr, rasterizer)
# Define action and observation space
# Continuous Action Space: gym.spaces.Box (X, Y, Yaw * number of future states)
self.action_space = spaces.Box(low=-1, high=1, shape=(3, ))
# Observation Space: gym.spaces.Dict (image: [n_channels, raster_size, raster_size])
obs_shape = (n_channels, raster_size, raster_size)
self.observation_space = spaces.Dict({
'image':
spaces.Box(low=0, high=1, shape=obs_shape, dtype=np.float32)
})
# Simulator Config within Gym
self.sim_cfg = sim_cfg if sim_cfg is not None else SimulationConfigGym(
)
self.simulator = ClosedLoopSimulator(self.sim_cfg,
self.dataset,
device=torch.device("cpu"),
mode=ClosedLoopSimulatorModes.GYM)
self.reward = reward if reward is not None else L2DisplacementYawReward(
)
self.max_scene_id = cfg["gym_params"]["max_scene_id"]
if not self.train:
self.max_scene_id = cfg["gym_params"]["max_val_scene_id"]
self.randomize_start_frame = False
if self.overfit:
self.overfit_scene_id = cfg["gym_params"]["overfit_id"]
self.randomize_start_frame = False
self.cle = cle
self.rescale_action = rescale_action
self.use_kinematic = use_kinematic
if self.use_kinematic:
self.kin_model = kin_model if kin_model is not None else UnicycleModel(
)
self.kin_rescale = self._get_kin_rescale_params()
else:
self.non_kin_rescale = self._get_non_kin_rescale_params()
# If not None, reset_scene_id is the scene_id that will be rolled out when reset is called
self.reset_scene_id = reset_scene_id
if self.overfit:
self.reset_scene_id = self.overfit_scene_id
# flag to decide whether to return any info at end of episode
# helps to limit the IPC
self.return_info = return_info
self.seed()
def seed(self, seed: int = None) -> List[int]:
"""Generate the random seed.
:param seed: the seed integer
:return: the output random seed
"""
self.np_random, seed = seeding.np_random(seed)
# TODO : add a torch seed for future
return [seed]
def set_reset_id(self, reset_id: int = None) -> None:
"""Set the reset_id to unroll from specific scene_id.
Useful during deterministic evaluation.
:param reset_id: the scene_id to unroll
"""
self.reset_scene_id = reset_id
def reset(self) -> Dict[str, np.ndarray]:
""" Resets the environment and outputs first frame of a new scene sample.
:return: the observation of first frame of sampled scene index
"""
# Define in / outs for new episode scene
self.agents_ins_outs: DefaultDict[
int, List[List[UnrollInputOutput]]] = defaultdict(list)
self.ego_ins_outs: DefaultDict[
int, List[UnrollInputOutput]] = defaultdict(list)
# Select Scene ID
self.scene_index = self.np_random.randint(0, self.max_scene_id)
if self.reset_scene_id is not None:
self.scene_index = min(self.reset_scene_id, self.max_scene_id - 1)
self.reset_scene_id += 1
# Select Frame ID (within bounds of the scene)
if self.randomize_start_frame:
scene_length = len(self.dataset.get_scene_indices(
self.scene_index))
self.eps_length = self.sim_cfg.num_simulation_steps or scene_length
end_frame = scene_length - self.eps_length
self.sim_cfg.start_frame_index = self.np_random.randint(
0, end_frame + 1)
# Prepare episode scene
self.sim_dataset = SimulationDataset.from_dataset_indices(
self.dataset, [self.scene_index], self.sim_cfg)
# Reset CLE evaluator
self.reward.reset()
# Output first observation
self.frame_index = 1 # Frame_index 1 has access to the true ego speed
ego_input = self.sim_dataset.rasterise_frame_batch(self.frame_index)
self.ego_input_dict = {
k: np.expand_dims(v, axis=0)
for k, v in ego_input[0].items()
}
# Reset Kinematic model
if self.use_kinematic:
init_kin_state = np.array(
[0.0, 0.0, 0.0, self.step_time * ego_input[0]['curr_speed']])
self.kin_model.reset(init_kin_state)
# Only output the image attribute
obs = {'image': ego_input[0]["image"]}
return obs
def step(self, action: np.ndarray) -> GymStepOutput:
"""Inputs the action, updates the environment state and outputs the next frame.
:param action: the action to perform on current state/frame
:return: the namedTuple comprising the (next observation, reward, done, info)
based on the current action
"""
frame_index = self.frame_index
next_frame_index = frame_index + 1
episode_over = next_frame_index == (len(self.sim_dataset) - 1)
# EGO
if not self.sim_cfg.use_ego_gt:
action = self._rescale_action(action)
ego_output = self._convert_action_to_ego_output(action)
self.ego_output_dict = ego_output
if self.cle:
# In closed loop training, the raster is updated according to predicted ego positions.
self.simulator.update_ego(self.sim_dataset, next_frame_index,
self.ego_input_dict,
self.ego_output_dict)
ego_frame_in_out = self.simulator.get_ego_in_out(
self.ego_input_dict, self.ego_output_dict,
self.simulator.keys_to_exclude)
self.ego_ins_outs[self.scene_index].append(
ego_frame_in_out[self.scene_index])
# generate simulated_outputs
simulated_outputs = SimulationOutputCLE(self.scene_index,
self.sim_dataset,
self.ego_ins_outs,
self.agents_ins_outs)
# reward calculation
reward = self.reward.get_reward(self.frame_index, [simulated_outputs])
# done is True when episode ends
done = episode_over
# Optionally we can pass additional info
# We are using "info" to output rewards and simulated outputs (during evaluation)
info: Dict[str, Any]
info = {
'reward_tot': reward["total"],
'reward_dist': reward["distance"],
'reward_yaw': reward["yaw"]
}
if done and self.return_info:
info = {
"sim_outs": self.get_episode_outputs(),
"reward_tot": reward["total"],
"reward_dist": reward["distance"],
"reward_yaw": reward["yaw"]
}
# Get next obs
self.frame_index += 1
obs = self._get_obs(self.frame_index, episode_over)
# return obs, reward, done, info
return GymStepOutput(obs, reward["total"], done, info)
def get_episode_outputs(self) -> List[EpisodeOutputGym]:
"""Generate and return the outputs at the end of the episode.
:return: List of episode outputs
"""
episode_outputs = [
EpisodeOutputGym(self.scene_index, self.sim_dataset,
self.ego_ins_outs, self.agents_ins_outs)
]
return episode_outputs
def render(self) -> None:
"""Render a frame during the simulation
"""
raise NotImplementedError
def _get_obs(self, frame_index: int,
episode_over: bool) -> Dict[str, np.ndarray]:
"""Get the observation corresponding to a given frame index in the scene.
:param frame_index: the index of the frame which provides the observation
:param episode_over: flag to determine if the episode is over
:return: the observation corresponding to the frame index
"""
if episode_over:
frame_index = 0 # Dummy final obs (when episode_over)
ego_input = self.sim_dataset.rasterise_frame_batch(frame_index)
self.ego_input_dict = {
k: np.expand_dims(v, axis=0)
for k, v in ego_input[0].items()
}
obs = {"image": ego_input[0]["image"]}
return obs
def _rescale_action(self, action: np.ndarray) -> np.ndarray:
"""Rescale the input action back to the un-normalized action space. PPO and related algorithms work well
with normalized action spaces. The environment receives a normalized action and we un-normalize it back to
the original action space for environment updates.
:param action: the normalized action
:return: the unnormalized action
"""
if self.rescale_action:
if self.use_kinematic:
action[0] = self.kin_rescale.steer_scale * action[0]
action[1] = self.kin_rescale.acc_scale * action[1]
else:
action[
0] = self.non_kin_rescale.x_mu + self.non_kin_rescale.x_scale * action[
0]
action[
1] = self.non_kin_rescale.y_mu + self.non_kin_rescale.y_scale * action[
1]
action[
2] = self.non_kin_rescale.yaw_mu + self.non_kin_rescale.yaw_scale * action[
2]
return action
def _get_kin_rescale_params(self) -> KinematicActionRescaleParams:
"""Determine the action un-normalization parameters for the kinematic model
from the current dataset in the L5Kit environment.
:return: Tuple of the action un-normalization parameters for kinematic model
"""
global MAX_ACC, MAX_STEER
return KinematicActionRescaleParams(MAX_STEER * self.step_time,
MAX_ACC * self.step_time)
def _get_non_kin_rescale_params(self,
max_num_scenes: int = 10
) -> NonKinematicActionRescaleParams:
"""Determine the action un-normalization parameters for the non-kinematic model
from the current dataset in the L5Kit environment.
:param max_num_scenes: maximum number of scenes to consider to determine parameters
:return: Tuple of the action un-normalization parameters for non-kinematic model
"""
scene_ids = list(range(self.max_scene_id)) if not self.overfit else [
self.overfit_scene_id
]
if len(scene_ids) > max_num_scenes: # If too many scenes, CPU crashes
scene_ids = scene_ids[:max_num_scenes]
sim_dataset = SimulationDataset.from_dataset_indices(
self.dataset, scene_ids, self.sim_cfg)
return calculate_non_kinematic_rescale_params(sim_dataset)
def _convert_action_to_ego_output(
self, action: np.ndarray) -> Dict[str, np.ndarray]:
"""Convert the input action into ego output format.
:param action: the input action provided by policy
:return: action in ego output format, a numpy dict with keys 'positions' and 'yaws'
"""
if self.use_kinematic:
data_dict = self.kin_model.update(action[:2])
else:
# [batch_size=1, num_steps, (X, Y, yaw)]
data = action.reshape(1, 1, 3)
pred_positions = data[:, :, :2]
# [batch_size, num_steps, 1->(yaw)]
pred_yaws = data[:, :, 2:3]
data_dict = {"positions": pred_positions, "yaws": pred_yaws}
return data_dict
def test_unroll(zarr_cat_dataset: ChunkedDataset, dmg: LocalDataManager,
cfg: dict) -> None:
rasterizer = build_rasterizer(cfg, dmg)
# change the first yaw of scene 1
# this will break if some broadcasting happens
zarr_cat_dataset.frames = np.asarray(zarr_cat_dataset.frames)
slice_frames = get_frames_slice_from_scenes(zarr_cat_dataset.scenes[1])
rot = zarr_cat_dataset.frames[slice_frames.start]["ego_rotation"].copy()
zarr_cat_dataset.frames[
slice_frames.start]["ego_rotation"] = yaw_as_rotation33(
rotation33_as_yaw(rot + 0.75))
scene_indices = list(range(len(zarr_cat_dataset.scenes)))
ego_dataset = EgoDataset(cfg, zarr_cat_dataset, rasterizer)
# control only agents at T0, control them forever
sim_cfg = SimulationConfig(use_ego_gt=False,
use_agents_gt=False,
disable_new_agents=True,
distance_th_close=1000,
distance_th_far=1000,
num_simulation_steps=10)
# ego will move by 1 each time
ego_model = MockModel(advance_x=1.0)
# agents will move by 0.5 each time
agents_model = MockModel(advance_x=0.5)
sim = ClosedLoopSimulator(sim_cfg, ego_dataset, torch.device("cpu"),
ego_model, agents_model)
sim_outputs = sim.unroll(scene_indices)
# check ego movement
for sim_output in sim_outputs:
ego_tr = sim_output.simulated_ego[
"ego_translation"][:sim_cfg.num_simulation_steps, :2]
ego_dist = np.linalg.norm(np.diff(ego_tr, axis=0), axis=-1)
assert np.allclose(ego_dist, 1.0)
ego_tr = sim_output.simulated_ego_states[:sim_cfg.num_simulation_steps,
TrajectoryStateIndices.
X:TrajectoryStateIndices.Y +
1]
ego_dist = np.linalg.norm(np.diff(ego_tr.numpy(), axis=0), axis=-1)
assert np.allclose(ego_dist, 1.0, atol=1e-3)
# all rotations should be the same as the first one as the MockModel outputs 0 for that
rots_sim = sim_output.simulated_ego[
"ego_rotation"][:sim_cfg.num_simulation_steps]
r_rep = sim_output.recorded_ego["ego_rotation"][0]
for r_sim in rots_sim:
assert np.allclose(rotation33_as_yaw(r_sim),
rotation33_as_yaw(r_rep),
atol=1e-2)
# all rotations should be the same as the first one as the MockModel outputs 0 for that
rots_sim = sim_output.simulated_ego_states[:sim_cfg.
num_simulation_steps,
TrajectoryStateIndices.
THETA]
r_rep = sim_output.recorded_ego_states[0, TrajectoryStateIndices.THETA]
for r_sim in rots_sim:
assert np.allclose(r_sim, r_rep, atol=1e-2)
# check agents movements
for sim_output in sim_outputs:
# we need to know which agents were controlled during simulation
# TODO: this is not ideal, we should keep track of them through the simulation
sim_dataset = SimulationDataset.from_dataset_indices(
ego_dataset, [sim_output.scene_id], sim_cfg)
sim_dataset.rasterise_agents_frame_batch(
0) # this will fill agents_tracked
agents_tracks = [el[1] for el in sim_dataset._agents_tracked]
for track_id in agents_tracks:
states = sim_output.simulated_agents
agents = filter_agents_by_track_id(
states, track_id)[:sim_cfg.num_simulation_steps]
agent_dist = np.linalg.norm(np.diff(agents["centroid"], axis=0),
axis=-1)
assert np.allclose(agent_dist, 0.5)
def test_unroll_subset(zarr_cat_dataset: ChunkedDataset, dmg: LocalDataManager,
cfg: dict, frame_range: Tuple[int, int]) -> None:
rasterizer = build_rasterizer(cfg, dmg)
scene_indices = list(range(len(zarr_cat_dataset.scenes)))
ego_dataset = EgoDataset(cfg, zarr_cat_dataset, rasterizer)
# control only agents at T0, control them forever
sim_cfg = SimulationConfig(use_ego_gt=False,
use_agents_gt=False,
disable_new_agents=True,
distance_th_close=1000,
distance_th_far=1000,
num_simulation_steps=frame_range[1],
start_frame_index=frame_range[0])
# ego will move by 1 each time
ego_model = MockModel(advance_x=1.0)
# agents will move by 0.5 each time
agents_model = MockModel(advance_x=0.5)
sim = ClosedLoopSimulator(sim_cfg, ego_dataset, torch.device("cpu"),
ego_model, agents_model)
sim_outputs = sim.unroll(scene_indices)
for sim_out in sim_outputs:
assert zarr_cat_dataset.frames[
0] != sim_out.recorded_dataset.dataset.frames[0]
assert zarr_cat_dataset.frames[
0] != sim_out.simulated_dataset.dataset.frames[0]
for ego_in_out in sim_out.ego_ins_outs:
assert "positions" in ego_in_out.outputs and "yaws" in ego_in_out.outputs
assert np.allclose(ego_in_out.outputs["positions"][:, 0], 1.0)
assert np.allclose(ego_in_out.outputs["positions"][:, 1], 0.0)
for agents_in_out in sim_out.agents_ins_outs:
for agent_in_out in agents_in_out:
assert "positions" in agent_in_out.outputs and "yaws" in agent_in_out.outputs
assert np.allclose(agent_in_out.outputs["positions"][:, 0],
0.5)
assert np.allclose(agent_in_out.outputs["positions"][:, 1],
0.0)
if None not in frame_range:
assert len(
sim_out.recorded_dataset.dataset.frames) == frame_range[1]
assert len(
sim_out.simulated_dataset.dataset.frames) == frame_range[1]
assert len(sim_out.simulated_ego_states) == frame_range[1]
assert len(sim_out.recorded_ego_states) == frame_range[1]
assert len(sim_out.recorded_ego) == frame_range[1]
assert len(sim_out.simulated_ego) == frame_range[1]
assert len(sim_out.ego_ins_outs) == len(
sim_out.agents_ins_outs) == frame_range[1]
ego_tr = sim_out.simulated_ego[
"ego_translation"][:sim_cfg.num_simulation_steps, :2]
ego_dist = np.linalg.norm(np.diff(ego_tr, axis=0), axis=-1)
assert np.allclose(ego_dist, 1.0)
