gqa-lxmert.py

# coding=utf-8
# Copyright 2020 The HuggingFace Datasets Authors and the current dataset script contributor.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""The GQA dataset preprocessed for LXMERT."""

import base64
import csv
import json
import os
import sys

import datasets
import numpy as np


csv.field_size_limit(sys.maxsize)


_CITATION = """\
@inproceedings{hudson2019gqa,
  title={Gqa: A new dataset for real-world visual reasoning and compositional question answering},
  author={Hudson, Drew A and Manning, Christopher D},
  booktitle={Proceedings of the IEEE/CVF conference on computer vision and pattern recognition},
  pages={6700--6709},
  year={2019}
}
"""

_DESCRIPTION = """\
GQA is a new dataset for real-world visual reasoning and compositional question answering,
seeking to address key shortcomings of previous visual question answering (VQA) datasets.
"""

_URLS = {
    "train": "https://nlp.cs.unc.edu/data/lxmert_data/gqa/train.json",
    "dev": "https://nlp.cs.unc.edu/data/lxmert_data/gqa/valid.json",
    "feat": "https://nlp.cs.unc.edu/data/lxmert_data/vg_gqa_imgfeat/vg_gqa_obj36.zip",
    "ans2label": "https://raw.githubusercontent.com/airsplay/lxmert/master/data/gqa/trainval_ans2label.json",
}

_FEAT_PATH = "vg_gqa_imgfeat/vg_gqa_obj36.tsv"

FIELDNAMES = [
    "img_id", "img_h", "img_w", "objects_id", "objects_conf", "attrs_id", "attrs_conf", "num_boxes", "boxes", "features"
]

_SHAPE_FEATURES = (36, 2048)
_SHAPE_BOXES = (36, 4)


class GqaLxmert(datasets.GeneratorBasedBuilder):
    """The GQA dataset preprocessed for LXMERT, with the objects features detected by a Faster RCNN replacing the
    raw images."""

    BUILDER_CONFIGS = [
        datasets.BuilderConfig(name="gqa", version=datasets.Version("1.0.0"), description="GQA dataset."),
    ]

    def _info(self):
        features = datasets.Features(
            {
                "question": datasets.Value("string"),
                "question_id": datasets.Value("int32"),
                "image_id": datasets.Value("string"),
                "features": datasets.Array2D(_SHAPE_FEATURES, dtype="float32"),
                "normalized_boxes": datasets.Array2D(_SHAPE_BOXES, dtype="float32"),
                "label": datasets.Value("int32"),
            }
        )
        return datasets.DatasetInfo(
            description=_DESCRIPTION,
            features=features,
            supervised_keys=None,
            citation=_CITATION,
        )

    def _split_generators(self, dl_manager):
        """Returns SplitGenerators."""
        dl_dir = dl_manager.download_and_extract(_URLS)
        features_path = os.path.join(dl_dir["feat"], _FEAT_PATH)
        return [
            datasets.SplitGenerator(
                name=datasets.Split.TRAIN,
                gen_kwargs={
                    "filepath": dl_dir["train"],
                    "ans2label_path": dl_dir["ans2label"],
                    "features_path": features_path,
                },
            ),
            datasets.SplitGenerator(
                name=datasets.Split.VALIDATION,
                gen_kwargs={
                    "filepath": dl_dir["dev"],
                    "ans2label_path": dl_dir["ans2label"],
                    "features_path": features_path,
                },
            ),
        ]

    def _generate_examples(self, filepath, ans2label_path, features_path):
        """ Yields examples as (key, example) tuples."""
        if not hasattr(self, "ans2label"):
            with open(ans2label_path, encoding="utf-8") as f:
                self.ans2label = json.load(f)
        if not hasattr(self, "id2features"):
            self.id2features = self._load_features(features_path)
        with open(filepath, encoding="utf-8") as f:
            gqa = json.load(f)
            for id_, d in enumerate(gqa):
                img_features = self.id2features[d["img_id"]]
                label = self.ans2label[next(iter(d["label"]))]
                yield id_, {
                    "question": d["sent"],
                    "question_id": d["question_id"],
                    "image_id": d["img_id"],
                    "features": img_features["features"],
                    "normalized_boxes": img_features["normalized_boxes"],
                    "label": label,
                }

    def _load_features(self, filepath):
        """Returns a dictionary mapping an image id to the corresponding image's objects features."""
        id2features = {}
        with open(filepath) as f:
            reader = csv.DictReader(f, FIELDNAMES, delimiter="\t")
            for i, item in enumerate(reader):
                features = {}
                img_h = int(item["img_h"])
                img_w = int(item["img_w"])
                num_boxes = int(item["num_boxes"])
                features["features"] = np.frombuffer(base64.b64decode(item["features"]), dtype=np.float32).reshape(
                    (num_boxes, -1)
                )
                boxes = np.frombuffer(base64.b64decode(item["boxes"]), dtype=np.float32).reshape((num_boxes, 4))
                features["normalized_boxes"] = self._normalize_boxes(boxes, img_h, img_w)
                id2features[item["img_id"]] = features
        return id2features

    def _normalize_boxes(self, boxes, img_h, img_w):
        """ Normalizes the input boxes given the original image size."""
        normalized_boxes = boxes.copy()
        normalized_boxes[:, (0, 2)] /= img_w
        normalized_boxes[:, (1, 3)] /= img_h
        return normalized_boxes