risk_biased_dataset.py

import datasets
import json
import numpy
import torch

_DESCRIPTION = """\
    Dataset of pre-processed samples from a small portion of the \
    Waymo Open Motion Data for our risk-biased prediction task.
"""

_CITATION = """\
    @InProceedings{NiMe:2022,
    author = {Haruki Nishimura, Jean Mercat, Blake Wulfe, Rowan McAllister},
    title = {RAP: Risk-Aware Prediction for Robust Planning},
    booktitle = {Proceedings of the 2022 IEEE International Conference on Robot Learning (CoRL)},
    month = {December},
    year = {2022},
    address = {Grafton Road, Auckland CBD, Auckland 1010},
    url = {},
}
"""

_URL = "https://huggingface.co/datasets/jmercat/risk_biased_dataset/resolve/main/"
_URLS = {
    "test": _URL + "data.json",
}

class RiskBiasedDataset(datasets.GeneratorBasedBuilder):
    """Dataset of pre-processed samples from a portion of the 
    Waymo Open Motion Data for the risk-biased prediction task."""
    
    VERSION = datasets.Version("0.0.0")
    
    BUILDER_CONFIGS = [
        datasets.BuilderConfig(name="risk_biased_dataset", version=VERSION, description="Dataset of pre-processed samples from a portion of the Waymo Open Motion Data for the risk-biased prediction task."),
    ]
    
    DEFAULT_CONFIG_NAME = "risk_biased_dataset"
    
    def _info(self):
        return datasets.DatasetInfo(
            description= _DESCRIPTION,
            features=datasets.Features(
                {"x": datasets.Sequence(datasets.Sequence(datasets.Sequence(datasets.Sequence(datasets.Value("float32"))))),
                 "mask_x": datasets.Sequence(datasets.Sequence(datasets.Sequence(datasets.Value("bool")))),
                 "y": datasets.Sequence(datasets.Sequence(datasets.Sequence(datasets.Sequence(datasets.Value("float32"))))),
                 "mask_y": datasets.Sequence(datasets.Sequence(datasets.Sequence(datasets.Value("bool")))),
                 "mask_loss": datasets.Sequence(datasets.Sequence(datasets.Sequence(datasets.Value("bool")))),
                 "map_data": datasets.Sequence(datasets.Sequence(datasets.Sequence(datasets.Sequence(datasets.Value("float32"))))),
                 "mask_map": datasets.Sequence(datasets.Sequence(datasets.Sequence(datasets.Value("bool")))),
                 "offset": datasets.Sequence(datasets.Sequence(datasets.Sequence(datasets.Value("float32")))),
                 "x_ego": datasets.Sequence(datasets.Sequence(datasets.Sequence(datasets.Sequence(datasets.Value("float32"))))),
                 "y_ego": datasets.Sequence(datasets.Sequence(datasets.Sequence(datasets.Sequence(datasets.Value("float32"))))),
                 }
            ),
            supervised_keys=None,
            homepage="https://sites.google.com/view/corl-risk/home",
            citation=_CITATION,
        )
        
    def _split_generators(self, dl_manager):
        urls_to_download = _URLS
        downloaded_files = dl_manager.download_and_extract(urls_to_download)
        
        return [datasets.SplitGenerator(name=datasets.Split.TEST, gen_kwargs={"filepath": downloaded_files["test"], "split": "test"}),]
        
    def _generate_examples(self, filepath, split):
        """Yields examples."""
        assert split == "test"
        with open(filepath, "r") as f:
            data = json.load(f)
            
            x = torch.from_numpy(numpy.array(data["x"]).astype(numpy.float32))
            mask_x = torch.from_numpy(numpy.array(data["mask_x"]).astype(numpy.bool8))
            y = torch.from_numpy(numpy.array(data["y"]).astype(numpy.float32))
            mask_y = torch.from_numpy(numpy.array(data["mask_y"]).astype(numpy.bool8))
            mask_loss = torch.from_numpy( numpy.array(data["mask_loss"]).astype(numpy.bool8))
            map_data = torch.from_numpy(numpy.array(data["map_data"]).astype(numpy.float32))
            mask_map = torch.from_numpy(numpy.array(data["mask_map"]).astype(numpy.bool8))
            offset = torch.from_numpy(numpy.array(data["offset"]).astype(numpy.float32))
            x_ego = torch.from_numpy(numpy.array(data["x_ego"]).astype(numpy.float32))
            y_ego = torch.from_numpy(numpy.array(data["y_ego"]).astype(numpy.float32))
            
            batch_size = x.shape[0]
            
            for i in range(batch_size):
                # yield i, {"x": x[i], "mask_x": mask_x[i],
                #           "y": y[i], "mask_y": mask_y[i], "mask_loss": mask_loss[i],
                #           "map_data": map_data[i], "mask_map": mask_map[i],
                #           "offset": offset[i], 
                #           "x_ego": x_ego[i],
                #           "y_ego": y_ego[i]}
                yield i, {"x": x[i:i+1], "mask_x": mask_x[i:i+1],
                          "y": y[i:i+1], "mask_y": mask_y[i:i+1], "mask_loss": mask_loss[i:i+1],
                          "map_data": map_data[i:i+1], "mask_map": mask_map[i:i+1],
                          "offset": offset[i:i+1], 
                          "x_ego": x_ego[i:i+1],
                          "y_ego": y_ego[i:i+1]}