transform_cc12m.py

from io import BytesIO

import ujson
import webdataset as wds
from PIL import Image
from tqdm import tqdm


def load_text(txt: bytes):
    return txt.decode()


def load_image(jpg):
    return Image.open(BytesIO(jpg)).convert('RGB')


load_mapping = {
    'input.jpg': load_image,
    'output.txt': load_text
}


def valid_image(img):
    return min(img.size) >= 64


def resize_img(img, max_size=512):
    width, height = img.size

    if width < max_size and height < max_size:
        return img

    if width > height:
        new_width = max_size
        new_height = int(new_width * height / width)
    elif height > width:
        new_height = max_size
        new_width = int(new_height * width / height)
    else:
        new_height = new_width = max_size

    img = img.resize((new_width, new_height), Image.ANTIALIAS)
    return img


def img_to_meta(img):
    width, height = img.size
    return {
        'width': width,
        'height': height
    }


def get_image(img):
    if not valid_image(img):
        return None, None

    resize_img(img)
    img_stream = BytesIO()
    img.save(img_stream, format='jpeg')

    img_stream.seek(0)
    return img_stream.read(), ujson.dumps(img_to_meta(img))

change_mapping = {
    'input.jpg': get_image
}

def func(wds_dataset_str, **kwargs):
    ds = wds.WebDataset(wds_dataset_str, shardshuffle=False).map_dict(**load_mapping).map_dict(**change_mapping).to_tuple(
        'input.jpg', 'output.txt')
    dl = wds.WebLoader(ds, batch_size=None, num_workers=48, prefetch_factor=16, **kwargs)

    writer = wds.ShardWriter('%05d.tar', 10000)
    for img, txt in tqdm(dl):
        img_str, meta = img
        if img_str is None:
            continue
        sample = {
            '__key__': f'{writer.count:08}',
            'jpg': img_str,
            'txt': txt,
            'json': meta
        }
        writer.write(sample)

if __name__ == '__main__':
    func('../conceptual-captions-12m-webdataset/{0..127}/{00000..4}.tar')