unsplash-image-text.py

import os
import urllib.request
import requests
import datasets
import json
import pandas as pd
from PIL import Image

_CITATION = """\
@InProceedings{huggingface:dataset,
title = {Unsplash Lite Dataset Images},
author={Unsplash},
year={2022}
}
"""

_DESCRIPTION = """\
This is a dataset that streams photos data from the Unsplash 25K servers.
"""
_HOMEPAGE = "https://github.com/unsplash/datasets/"

_LICENSE = ""

_REPO = "https://huggingface.co/datasets/jamescalam/unsplash-image-text"

_URL = "https://unsplash.com/data/lite/latest"

URL_JSON = "https://huggingface.co/datasets/jamescalam/unsplash-image-text/raw/main/url.json"

_IMAGE_TGZ = [f"images_{i}.tgz" for i in range(25)]

_COLS = [
    'photo_id',
    'photo_url',
    'photo_width',
    'photo_height',
    'photo_aspect_ratio',
    'photo_description',
    'ai_description',
    'image'
]

class Unsplash(datasets.GeneratorBasedBuilder):
    """The Unsplash 25K dataset for photos"""

    def _info(self):
        return datasets.DatasetInfo(
            description=_DESCRIPTION,
            features=datasets.Features(
                {
                    'photo_id': datasets.Value("string"),
                    'photo_url': datasets.Value("string"),
                    'photo_width': datasets.Value("int32"),
                    'photo_height': datasets.Value("int32"),
                    'photo_aspect_ratio': datasets.Value("float32"),
                    'photo_description': datasets.Value("string"),
                    'ai_description': datasets.Value("string"),
                    'image': datasets.Image(),
                }
            ),
            supervised_keys=None,
            homepage="https://github.com/unsplash/datasets/",
            citation=_CITATION,
        )

    def _split_generators(self, dl_manager):
        new_url = dl_manager.download_and_extract(_URL)
        tgz_urls = [f"{_REPO}/resolve/main/images/{file}" for file in _IMAGE_TGZ]
        archive_paths = dl_manager.download_and_extract(tgz_urls)
        tar_paths = {}
        for path in archive_paths:
            image_tar = os.listdir(f"{path}/subdirs")[0]
            tar_paths[image_tar] = f"{path}/subdirs/{image_tar}"
        # remove extra files
        for file in os.listdir(new_url):
            if os.path.isfile(new_url+"/"+file):
                if file not in ['photos.tsv000', 'url.json']:
                    os.remove(new_url+'/'+file)
        # load URLs
        res = requests.get(f"{_REPO}/raw/main/url.json")
        urls = json.loads(res.text)
        # load mapping dict
        res = requests.get(f"{_REPO}/raw/main/maps.json")
        maps = json.loads(res.text)
        # format into dataframe
        keys = list(urls.keys())
        values = list(urls.values())
        urls = pd.DataFrame({
            'photo_id': keys,
            'photo_url': values
        })
        # load the unsplash 25k data
        unsplash = pd.read_csv(new_url+"/photos.tsv000", sep='\t')
        # drop rows where there is no text description
        unsplash = unsplash[
            (unsplash['photo_description'] != "") | (unsplash['ai_description'] != "")
        ]
        # keep only required cols
        cols = [col for col in _COLS if col not in ['photo_url', 'image']]
        unsplash = unsplash[cols]
        # merge with urls
        unsplash = unsplash.merge(urls, on='photo_id', how='inner')
        # save to tsv
        unsplash.to_csv(new_url+'/unsplash.tsv', sep='\t', index=False)
        return [
            datasets.SplitGenerator(
                name=datasets.Split.TRAIN,
                gen_kwargs={
                    "filepath": new_url+"/unsplash.tsv",
                    "tar_paths": tar_paths,
                    "maps": maps,
                }
            ),
        ]

    def _generate_examples(self, filepath, tar_paths, maps):
        """This function returns the examples in the raw (text) form."""
        with open(filepath, "r") as f:
            id_ = 0
            for line in f:
                if id_ == 0:
                    cols = line.strip().split("\t")
                    id_ += 1
                else:
                    values = line.strip().split("\t")
                    if len(values) != len(cols):
                        values.extend(['']*(len(cols)-len(values)))
                    data = {
                        cols[i]: values[i] for i in range(len(cols))
                    }
                    try:
                        photo_id = data['photo_id']
                        try:
                            tar = maps[photo_id]
                            tar_path = tar_paths[tar]
                            image = Image.open(f"{tar_path}/{data['photo_id']}.jpg")
                            data['image'] = image
                            yield id_, data
                        except KeyError:
                            pass
                    except OSError:
                        # in this case don't return the record
                        pass
                    id_ += 1