NLVR2.py

"""NVLR2 loading script."""


import json
import os
import datasets


_CITATION = """\
@article{DBLP:journals/corr/abs-2202-01994,
  author    = {Yamini Bansal and
               Behrooz Ghorbani and
               Ankush Garg and
               Biao Zhang and
               Maxim Krikun and
               Colin Cherry and
               Behnam Neyshabur and
               Orhan Firat},
  title     = {Data Scaling Laws in {NMT:} The Effect of Noise and Architecture},
  journal   = {CoRR},
  volume    = {abs/2202.01994},
  year      = {2022},
  url       = {https://arxiv.org/abs/2202.01994},
  eprinttype = {arXiv},
  eprint    = {2202.01994},
  timestamp = {Mon, 24 Oct 2022 10:21:23 +0200},
  biburl    = {https://dblp.org/rec/journals/corr/abs-2202-01994.bib},
  bibsource = {dblp computer science bibliography, https://dblp.org}
}
"""

_DESCRIPTION = """\
The Natural Language for Visual Reasoning corpora are two language grounding datasets containing natural language sentences grounded in images. The task is to determine whether a sentence is true about a visual input. The data was collected through crowdsourcings, and solving the task requires reasoning about sets of objects, comparisons, and spatial relations. This includes two corpora: NLVR, with synthetically generated images, and NLVR2, which includes natural photographs.
"""

_HOMEPAGE = "https://lil.nlp.cornell.edu/nlvr/"

_LICENSE = "CC BY 4.0"  # TODO need to credit both ms coco and vqa authors!

_URL_BASE = "https://github.com/lil-lab/nlvr/tree/master/nlvr2/data"
_SPLITS = {
    "train": "train.json",
    "validation": "dev.json",
    "test1": "test1.json",
    "test2": "test2.json",
}


class NLVR2Dataset(datasets.GeneratorBasedBuilder):

    VERSION = datasets.Version("1.0.0")


    DEFAULT_CONFIG_NAME = "Default"

    def _info(self):
        return datasets.DatasetInfo(
            description=_DESCRIPTION,
            features=datasets.Features(
                {
                    "identifier": datasets.Value("string"),
                    "sentence": datasets.Value("string"),
                    "left_image": datasets.Image(),
                    "right_image": datasets.Image(),
                    "label": datasets.Value("string"),
                }
            ),
            homepage=_HOMEPAGE,
            license=_LICENSE,
            citation=_CITATION,
        )

    def _split_generators(self, dl_manager):
        urls = {
            "Default": {
                "train": os.path.join(_URL_BASE, _SPLITS["train"]),
                "validation": os.path.join(_URL_BASE, _SPLITS["validation"]),
                "test1": os.path.join(_URL_BASE, _SPLITS["test1"]),
                "test2": os.path.join(_URL_BASE, _SPLITS["test2"]),
            },
        }
        files_path = dl_manager.download_and_extract(urls)
        return [
            datasets.SplitGenerator(
                name=datasets.Split.TRAIN,
                gen_kwargs={"files_path": files_path[self.config.name]["train"]},
            ),
            datasets.SplitGenerator(
                name=datasets.Split.VALIDATION,
                gen_kwargs={"files_path": files_path[self.config.name]["validation"]},
            ),
            datasets.SplitGenerator(
                name=datasets.Split.TEST,
                gen_kwargs={"files_path": files_path[self.config.name]["test1"]},
            ),
            # datasets.SplitGenerator(
            #     name=datasets.Split.TEST2,
            #     gen_kwargs={"files_path": files_path[self.config.name]["test2"]},
            # ),
        ]

    def _generate_examples(self, files_path):
        json_file = json.load(open(files_path, "r"))

        for i, ex in enumerate(json_file):
            record = {
                "identifier": ex["identifier"],
                "sentence": ex["sentence"],
                "left_image": str(ex["left_url"]),
                "right_image": str(ex["right_url"]),
                "label": ex["label"],
            }
            yield i, record