wiki_dpr.py

import os

import pyarrow as pa
import pyarrow.parquet as pq

import datasets


logger = datasets.logging.get_logger(__name__)


_CITATION = """
@inproceedings{karpukhin-etal-2020-dense,
    title = "Dense Passage Retrieval for Open-Domain Question Answering",
    author = "Karpukhin, Vladimir and Oguz, Barlas and Min, Sewon and Lewis, Patrick and Wu, Ledell and Edunov, Sergey and Chen, Danqi and Yih, Wen-tau",
    booktitle = "Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP)",
    month = nov,
    year = "2020",
    address = "Online",
    publisher = "Association for Computational Linguistics",
    url = "https://www.aclweb.org/anthology/2020.emnlp-main.550",
    doi = "10.18653/v1/2020.emnlp-main.550",
    pages = "6769--6781",
}
"""

_DESCRIPTION = """
This is the wikipedia split used to evaluate the Dense Passage Retrieval (DPR) model.
It contains 21M passages from wikipedia along with their DPR embeddings.
The wikipedia articles were split into multiple, disjoint text blocks of 100 words as passages.
"""

_LICENSE = """DPR is CC-BY-NC 4.0 licensed."""

_DATA_URL = "https://dl.fbaipublicfiles.com/dpr/wikipedia_split/psgs_w100.tsv.gz"

_NQ_VECTORS_URL = "https://dl.fbaipublicfiles.com/dpr/data/wiki_encoded/single/nq/wiki_passages_{i}"

_MULTISET_VECTORS_URL = "https://dl.fbaipublicfiles.com/rag/rag_multiset_embeddings/wiki_passages_{i}"

_DATA_TO_NUM_SHARDS = {
    "nq": 157,
    "multiset": 157,
    "no_embeddings": 28,
    "dummy.nq": 1,
    "dummy.multiset": 1,
    "dummy.no_embeddings": 1,
}


class WikiDprConfig(datasets.BuilderConfig):
    """BuilderConfig for WikiDprConfig."""

    def __init__(
        self,
        with_embeddings=True,
        with_index=True,
        wiki_split="psgs_w100",
        embeddings_name="nq",
        index_name="compressed",
        index_train_size=262144,
        dummy=False,
        **kwargs,
    ):
        """BuilderConfig for WikiSnippets.
        Args:
            with_embeddings (`bool`, defaults to `True`): Load the 768-dimensional embeddings from DPR.
            with_index (`bool`, defaults to `True`): Load the faiss index trained on the embeddings.
            wiki_split (`str`, defaults to `psgs_w100`): name of the splitting method of wiki articles.
            embeddings_name (`str`, defaults to `nq`): "nq" or "multiset", depending on which dataset DPR was trained on.
            index_name (`str`, defaults to `compressed`): "compressed" or "exact", the configuration of the faiss index to use.
            index_train_size (`int`, defaults to `262144`): Size of the subset to use to train the index, if it is trainable.
            dummy (`bool`, defaults to `False`): Dummy uses only 10 000 examples for testing purposes.
          **kwargs: keyword arguments forwarded to super.
        """
        self.with_embeddings = with_embeddings
        self.with_index = with_index and index_name != "no_index"
        self.wiki_split = wiki_split
        self.embeddings_name = embeddings_name
        self.index_name = index_name if with_index else "no_index"
        self.index_train_size = index_train_size
        self.dummy = dummy
        name = [self.wiki_split, self.embeddings_name, self.index_name]
        if not self.with_embeddings:
            name.append("no_embeddings")
        if self.dummy:
            name = ["dummy"] + name
            assert (
                self.index_name != "compressed" or not self.with_index
            ), "Please use `index_name='exact' for dummy wiki_dpr`"
        assert wiki_split == "psgs_w100"
        assert embeddings_name in ("nq", "multiset")
        assert index_name in ("compressed", "exact", "no_index")
        kwargs["name"] = ".".join(name)
        super(WikiDprConfig, self).__init__(**kwargs)

        prefix = f"{wiki_split}.{embeddings_name}."
        if self.index_name == "exact":
            self.index_file = prefix + "HNSW128_SQ8-IP-{split}.faiss"
        else:
            self.index_file = prefix + "IVF4096_HNSW128_PQ128-IP-{split}.faiss"
        if self.dummy:
            self.index_file = "dummy." + self.index_file


class WikiDpr(datasets.ArrowBasedBuilder):
    BUILDER_CONFIG_CLASS = WikiDprConfig
    BUILDER_CONFIGS = [
        WikiDprConfig(
            embeddings_name=embeddings_name,
            with_embeddings=with_embeddings,
            index_name=index_name,
            version=datasets.Version("0.0.0"),
        )
        for with_embeddings in (True, False)
        for embeddings_name in ("nq", "multiset")
        for index_name in ("exact", "compressed", "no_index")
    ]

    def _info(self):
        return datasets.DatasetInfo(
            description=_DESCRIPTION,
            features=datasets.Features(
                {
                    "id": datasets.Value("string"),
                    "text": datasets.Value("string"),
                    "title": datasets.Value("string"),
                    "embeddings": datasets.Sequence(datasets.Value("float32")),
                }
            )
            if self.config.with_embeddings
            else datasets.Features(
                {"id": datasets.Value("string"), "text": datasets.Value("string"), "title": datasets.Value("string")}
            ),
            supervised_keys=None,
            homepage="https://github.com/facebookresearch/DPR",
            citation=_CITATION,
        )

    def _split_generators(self, dl_manager):
        data_dir = self.config.embeddings_name if self.config.with_embeddings else "no_embeddings"
        if self.config.dummy:
            data_dir = "dummy." + data_dir
        num_shards = _DATA_TO_NUM_SHARDS[data_dir]
        data_dir = os.path.join("data", self.config.wiki_split, data_dir)
        files = [os.path.join(data_dir, f"train-{i:05d}-of-{num_shards:05d}.parquet") for i in range(num_shards)]
        downloaded_files = dl_manager.download_and_extract(files)
        return [
            datasets.SplitGenerator(name=datasets.Split.TRAIN, gen_kwargs={"files": downloaded_files}),
        ]

    def _generate_tables(self, files):
        for file in files:
            with open(file, "rb") as f:
                f = pq.ParquetFile(f)  
                for batch_idx, batch in enumerate(f.iter_batches(batch_size=1000)):
                    yield batch_idx, pa.Table.from_batches([batch])
                    batch_idx += 1

    def _post_processing_resources(self, split):
        if self.config.with_index:
            return {"embeddings_index": self.config.index_file.format(split=split)}
        else:
            return {}

    def _download_post_processing_resources(self, split, resource_name, dl_manager):
        if resource_name == "embeddings_index":
            try:
                embeddings_name = self.config.embeddings_name
                if self.config.dummy:
                    embeddings_name = "dummy." + embeddings_name
                index_file = os.path.join(
                    "data",
                    self.config.wiki_split,
                    self.config.embeddings_name,
                    (
                        "HNSW128_SQ8-IP-{split}.faiss"
                        if self.config.index_name == "exaxt"
                        else "IVF4096_HNSW128_PQ128-IP-{split}.faiss"
                    ).format(split=split)
                )
                downloaded_resources = dl_manager.download_and_extract(
                    {"embeddings_index": index_file}
                )
                return downloaded_resources["embeddings_index"]
            except (FileNotFoundError, ConnectionError):  # index doesn't exist
                pass

    def _post_process(self, dataset, resources_paths):
        if self.config.with_index:
            index_file = resources_paths["embeddings_index"]
            if os.path.exists(index_file):
                dataset.load_faiss_index("embeddings", index_file)
            else:
                if "embeddings" not in dataset.column_names:
                    raise ValueError("Couldn't build the index because there are no embeddings.")
                import faiss

                d = 768
                train_size = self.config.index_train_size
                logger.info("Building wiki_dpr faiss index")
                if self.config.index_name == "exact":
                    index = faiss.IndexHNSWSQ(d, faiss.ScalarQuantizer.QT_8bit, 128, faiss.METRIC_INNER_PRODUCT)
                    index.hnsw.efConstruction = 200
                    index.hnsw.efSearch = 128
                    dataset.add_faiss_index("embeddings", custom_index=index, train_size=train_size)
                else:
                    quantizer = faiss.IndexHNSWFlat(d, 128, faiss.METRIC_INNER_PRODUCT)
                    quantizer.hnsw.efConstruction = 200
                    quantizer.hnsw.efSearch = 128
                    ivf_index = faiss.IndexIVFPQ(quantizer, d, 4096, 128, 8, faiss.METRIC_INNER_PRODUCT)
                    ivf_index.nprobe = 64
                    ivf_index.own_fields = True
                    quantizer.this.disown()
                    dataset.add_faiss_index(
                        "embeddings",
                        train_size=train_size,
                        custom_index=ivf_index,
                    )
                logger.info("Saving wiki_dpr faiss index")
                dataset.save_faiss_index("embeddings", index_file)
        return dataset