odinsynth_dataset.py

import csv
import json
import os
from collections import defaultdict

import datasets
from tqdm import tqdm

_DESCRIPTION = """\
Supervised training data for odinsynth
"""



class OdinsynthDatasetBuilder(datasets.GeneratorBasedBuilder):

    VERSION = datasets.Version("1.0.0")

    def _info(self):

        features = datasets.Features(
            {
                "rule_id": datasets.Value("int32"),
                "parent": datasets.Value("string"),
                "child": datasets.Value("string"),
                "negative_child": datasets.Value("string"),
                "spec": datasets.Sequence(datasets.Value("string")),
                "matches": datasets.Sequence(datasets.Sequence(datasets.Value("int16"))),
                "step": datasets.Value("int8"),
                "length": datasets.Value("int8")
            }
        )

        return datasets.DatasetInfo(
            # This is the description that will appear on the datasets page.
            description=_DESCRIPTION,
            # This defines the different columns of the dataset and their types
            features=features,  # Here we define them above because they are different between the two configurations
            # If there's a common (input, target) tuple from the features, uncomment supervised_keys line below and
            # specify them. They'll be used if as_supervised=True in builder.as_dataset.
            # supervised_keys=("sentence", "label"),
            # Homepage of the dataset for documentation
            # homepage=_HOMEPAGE,
            # # License for the dataset if available
            # license=_LICENSE,
            # # Citation for the dataset
            # citation=_CITATION,
        )

    def _build_specs(self, path:str):
        id_to_rule = {}
        specs = defaultdict(list)
        matches = defaultdict(list)
        with open(path) as f:
            for l in tqdm(f, desc="Pre-computing specs"):
                try:
                    instance = json.loads(l)
                    if instance['match']:
                        rule_id = int(instance['id'])
                        rule = instance['question']
                        sent  = instance['context']
                        if sent not in specs[rule]:
                            specs[rule].append(sent)
                            matches[rule].append([instance['match_start'], instance['match_end']])
                            id_to_rule[rule_id] = rule
                except:
                    # TODO log
                    pass

        return {rule_id:(specs[rule], matches[rule]) for rule_id, rule in id_to_rule.items()}



    def _split_generators(self, dl_manager):
        # If several configurations are possible (listed in BUILDER_CONFIGS), the configuration selected by the user is in self.config.name

        JSON_PATH = dl_manager.download_and_extract('merged_train_split_train.jsonl.gz')
        TRAIN_ARCHIVE_PATH = dl_manager.download('train.tar.bz2')
        VAL_ARCHIVE_PATH = dl_manager.download('val.tar.bz2')
        TEST_ARCHIVE_PATH = dl_manager.download('test.tar.bz2')

        specs = self._build_specs(JSON_PATH)
        return [
            datasets.SplitGenerator(
                name=datasets.Split.TRAIN,
                # These kwargs will be passed to _generate_examples
                gen_kwargs={
                    "archive_iter": dl_manager.iter_archive(TRAIN_ARCHIVE_PATH),
                    "specs": specs,
                    "split": "train",
                },
            ),
            datasets.SplitGenerator(
                name=datasets.Split.TEST,
                # These kwargs will be passed to _generate_examples
                gen_kwargs={
                    "archive_iter": dl_manager.iter_archive(TEST_ARCHIVE_PATH),
                    "specs": specs,
                    "split": "test",
                },
            ),
            datasets.SplitGenerator(
                name=datasets.Split.VALIDATION,
                # These kwargs will be passed to _generate_examples
                gen_kwargs={
                    "archive_iter": dl_manager.iter_archive(VAL_ARCHIVE_PATH),
                    "specs": specs,
                    "split": "val",
                },
            ),
        ]

    # method parameters are unpacked from `gen_kwargs` as given in `_split_generators`
    def _generate_examples(self, archive_iter, specs, split):
        # The `key` is for legacy reasons (tfds) and is not important in itself, but must be unique for each example.

        key = 0
        for tsv_path, file in archive_iter:
            if tsv_path.endswith(".tsv"):
                # Read the lines
                reader = csv.reader((l.decode() for l in  file), delimiter='\t')
                for row in reader:
                    rule_id = int(row[0])
                    if rule_id in specs:
                        spec, matches = specs[rule_id]
                        assert len(spec) == len(matches), f"Rule id {id} has different number of sentences and matches"

                        yield key, {
                            "rule_id": rule_id,
                            "parent": row[1],
                            "child": row[2],
                            "negative_child": row[3],
                            "spec": spec,
                            "matches": matches,
                            "step": int(row[4]),
                            "length": int(row[5]),
                        }
                        # Increase the key after yielding the instacne
                        key += 1


if __name__  == "__main__":
    ds = OdinsynthDatasetBuilder()
    ds.download_and_prepare()
    print(ds.cache_dir)