Working with the Metadata

Downloading all the metadata files at once

Install the huggingface-cli utility (via pip). You may then use the following command:

huggingface-cli download Spawning/PD12M --repo-type dataset --local-dir metadata --include "metadata/*"

metadata format

The metadata files are in parquet format, and contain the following attributes:

id: A unique identifier for the image.
url: The URL of the image.
s3_key: The S3 file key of the image.
caption: A caption for the image.
hash: The MD5 hash of the image file.
width: The width of the image in pixels.
height: The height of the image in pixels.
mime_type: The MIME type of the image file.
license: The URL of the license.
source: The source organization of the image.

Open a metadata file

The files are in parquet format, and can be opened with a tool like pandas in Python.

import pandas as pd
df = pd.read_parquet('pd12m.000.parquet')

Get URLs from metadata

Once you have opened a maetadata file with pandas, you can get the URLs of the images with the following command:

urls = df['url']

Download all files mentioned in metadata

If you want to just grab all files referenced by a metadata collection, you may try this (adjust to taste):

img2dataset --url_list $file --input_format "parquet" \
--url_col "url" --caption_col "caption" --output_format files \
--output_folder $dir --processes_count 16 --thread_count 64 \
--skip_reencode true --min_image_sizel 654 --max_aspect_ratio=1.77