def load_flags(mode="") -> Flags:
parser = argparse.ArgumentParser(description='')
parser.add_argument('--yaml_filepath', '-y', type=str, default="./flags/20200905_sample_flag.yaml",
help='Flags yaml file path')
args = parser.parse_args()
# --- Default setting ---
flags = load_yaml(args.yaml_filepath)
# print("yaml flags", flags)
base_flags = Flags()
base_flags.update(flags)
return base_flags
help='Directory to output the result')
parser.add_argument('--debug',
'-d',
type=strtobool,
default='false',
help='Debug mode')
args = parser.parse_args()
return args
if __name__ == '__main__':
args = parse()
out_dir = Path(args.out)
debug = args.debug
flags_dict = load_yaml(out_dir / 'flags.yaml')
cfg = load_yaml(out_dir / 'cfg.yaml')
# flags = DotDict(flags_dict)
flags = Flags()
flags.update(flags_dict)
print(f"flags: {flags_dict}")
# set env variable for data
# Not use flags.l5kit_data_folder, but use fixed test data.
l5kit_data_folder = "../../input/lyft-motion-prediction-autonomous-vehicles"
os.environ["L5KIT_DATA_FOLDER"] = l5kit_data_folder
dm = LocalDataManager(None)
print("Load dataset...")
default_test_cfg = {
'key': 'scenes/test.zarr',
from lib.training.lr_scheduler import LRScheduler
from lib.training.snapshot_object_when_lr_increase import SnapshotObjectWhenLRIncrease
from lib.utils.resumable_distributed_sampler import ResumableDistributedSampler
from lib.functions.nll import pytorch_weighted_neg_multi_log_likelihood_batch, pytorch_neg_multi_log_likelihood_batch
if __name__ == '__main__':
# --- Distributed initial setup ---
is_mpi, rank, world_size, local_rank = setup_distributed()
print(f"is_mpi {is_mpi}, rank {rank}, world_size {world_size}, local_rank {local_rank}")
# --- Distributed initial setup done ---
mode = "distributed" if is_mpi else ""
flags: Flags = load_flags(mode=mode)
flags_dict = dataclasses.asdict(flags)
cfg = load_yaml(flags.cfg_filepath)
out_dir = Path(flags.out_dir)
if local_rank == 0:
print(f"cfg {cfg}")
os.makedirs(str(out_dir), exist_ok=True)
print(f"flags: {flags_dict}")
save_yaml(out_dir / 'flags.yaml', flags_dict)
save_yaml(out_dir / 'cfg.yaml', cfg)
debug = flags.debug
# set env variable for data
os.environ["L5KIT_DATA_FOLDER"] = flags.l5kit_data_folder
dm = LocalDataManager(None)
# ===== INIT DATASET
if local_rank == 0:
# h = torch.sum(h, dim=(-1, -2))
h = torch.mean(h, dim=(-1, -2))
else:
# [128, 2048, 4, 4] when input is (128, 128)
bs, ch, height, width = h.shape
h = h.view(bs, ch*height*width)
for layer in self.lin_layers:
h = layer(h)
return h
if __name__ == '__main__':
# --- test instantiation ---
from lib.utils.yaml_utils import load_yaml
cfg = load_yaml("../../../../modeling/configs/0905_cfg.yaml")
num_modes = 3
model = ResNeStMulti(cfg, num_modes=num_modes)
print(type(model))
print(model)
bs = 3
in_channels = model.in_channels
height, width = 224, 224
device = "cuda:0"
x = torch.rand((bs, in_channels, height, width), dtype=torch.float32).to(device)
model.to(device)
# pred, confidences = model(x)
# print("pred", pred.shape, "confidences", confidences.shape)
h = model(x)
return len(self.get_frame_arguments)
def __getitem__(self, index: int) -> dict:
track_id, scene_index, state_index = self.get_frame_arguments[index]
# return self.ego_dataset.get_frame(scene_index, state_index, track_id=track_id)
return get_frame_custom(self.ego_dataset, scene_index, state_index, track_id=track_id)
if __name__ == "__main__":
from l5kit.data import LocalDataManager
from l5kit.rasterization import build_rasterizer
from lib.utils.yaml_utils import load_yaml
repo_root = Path(__file__).parent.parent.parent.parent
dm = LocalDataManager(local_data_folder=str(repo_root / "input" / "lyft-motion-prediction-autonomous-vehicles"))
dataset = ChunkedDataset(dm.require("scenes/sample.zarr")).open(cached=False)
cfg = load_yaml(repo_root / "src" / "modeling" / "configs" / "0905_cfg.yaml")
rasterizer = build_rasterizer(cfg, dm)
faster_agent_dataset = FasterAgentDataset(cfg, dataset, rasterizer, None)
fast_agent_dataset = FastAgentDataset(cfg, dataset, rasterizer, None)
assert len(faster_agent_dataset) == len(fast_agent_dataset)
keys = ["image", "target_positions", "target_availabilities"]
for index in tqdm(range(min(1000, len(faster_agent_dataset)))):
actual = faster_agent_dataset[index]
expected = fast_agent_dataset[index]
for key in keys:
assert (actual[key] == expected[key]).all()