/coyo-dataset

COYO-700M: Large-scale Image-Text Pair Dataset

Primary LanguagePython

🐺 COYO-700M: Image-Text Pair Dataset

COYO-700M is a large-scale dataset that contains 747M image-text pairs as well as many other meta-attributes to increase the usability to train various models. Our dataset follows a similar strategy to previous vision-and-language datasets, collecting many informative pairs of alt-text and its associated image in HTML documents. We expect COYO to be used to train popular large-scale foundation models complementary to other similar datasets.

More details on the data acquisition process can be found in [our paper] (which will be updated soon).

Updates

Data Collection Process

  • We collected about 10 billion pairs of alt-text and image sources in HTML documents in CommonCrawl from Oct. 2020 to Aug. 2021. and eliminated uninformative pairs through the image and text level filtering process with minimal cost. The following figure outlines our data collection procedure.
https://en.wikipedia.org/wiki/Napoleon

Data Filtering

Image Level

  • Included all image formats that Pillow library can decode. (JPEG, WEBP, PNG, BMP, ...)
  • Removed images less than 5KB image size.
  • Removed images with an aspect ratio greater than 3.0.
  • Removed images with min(width, height) < 200.
  • Removed images with a score of OpenNSFW2 or GantMan/NSFW higher than 0.5.
  • Removed all duplicate images based on the image pHash value from external public datasets.
    • ImageNet-1K/21K, Flickr-30K, MS-COCO, CC-3M, CC-12M

Text Level

  • Collected only English text using cld3.
  • Replaced consecutive whitespace characters with a single whitespace and removed the whitespace before and after the sentence.
    • e.g. "\n \n Load image into Gallery viewer, valentine&amp;#39;s day roses\n \n" → "Load image into Gallery viewer, valentine&amp;#39;s day roses"
  • Removed texts with a length of 5 or less.
  • Removed texts that do not have a noun form.
  • Removed texts with less than 3 words or more than 256 words and texts over 1000 in length.
  • Removed texts appearing more than 10 times.
    • e.g. “thumbnail for”, “image for”, “picture of”
  • Removed texts containing NSFW words collected from profanity_filter, better_profanity, and google_twunter_lol.

Image-Text Level

  • Removed duplicated samples based on (image_phash, text).
    • Different text may exist for the same image URL.

Dataset Preview

id                                                                              url                                                                              text width height image_phash text_length word_count num_tokens_bert num_tokens_gpt num_faces clip_similarity_vitb32 clip_similarity_vitl14 nsfw_score_opennsfw2 nsfw_score_gantman watermark_score aesthetic_score_laion_v2
4896263451343 Fishing Fleet (Monterey), California art by Art Riley. HD giclee art prints for sale at CaliforniaWatercolor.com - original California paintings, & premium giclee prints for sale 600 447 bac58374982e0fc7 178 25 39 40 0 0.319336 0.248169 2.54512e-05 0.0293861 0.0406009 7.04812
1425929344479 The Gate by Pete2453 600 347 8374726575bc0f8a 20 4 6 6 0 0.24939 0.203735 6.97374e-06 0.00823276 0.0721415 6.98521
7456063527931 Beautiful Pictures From the Shores of the Mythical Land (42 600 320 949d1fe559e2cc90 59 10 11 14 0 0.290771 0.179321 0.0130615 0.0178628 0.489642 6.94643
3221225511175 contemporary expensive lighting fixtures with minimum lighting 800 499 e5ea35075ab912c6 62 7 7 8 0 0.263916 0.217896 0.000990868 0.0137114 0.0960748 4.57594
5626407855002 Nintendo Co.'s Super Mario is displayed on coffee mugs for sale at the Nintendo World store in New York, U.S., on Friday, May 17, 2013. 2000 1309 9311891e9437f4f3 135 27 37 35 0 0.400878 0.316650 0.00362968 0.0317519 0.0022693 6.324910
1125282207474 FILE PHOTO: A rainbow appears on the Auckland skyline featuring Sky Tower in New Zealand 800 525 85b89c0166ee63be 88 15 16 16 0 0.4453125 0.3505859 2.640485e-05 0.012074 0.0219129 5.294523
1434519186493 A man covers himself with algae as he poses for photographs on a beach in Qingdao, Shandong province on Tuesday, July 23, 2013. -- FILE PHOTO: REUTERS 860 573 f2c48dabbf93810a 150 26 35 36 7 0.4165039 0.3427734 0.025009 0.01608 0.072775 6.833739

Dataset Numbers

count ratio
# of image-text pairs 746,972,269 100.00%
# of unique urls 656,114,783 87.84%
# of unique image_phash 579,679,137 77.60%
# of unique text 566,253,888 75.81%

Meta-Attributes

Attributes

name type description
id long Unique 64-bit integer ID generated by monotonically_increasing_id()
url string The image URL extracted from the src attribute of the <img> tag
text string The text extracted from the alt attribute of the <img> tag
width integer The width of the image
height integer The height of the image
image_phash string The perceptual hash(pHash) of the image
text_length integer The length of the text
word_count integer The number of words separated by spaces.
num_tokens_bert integer The number of tokens using BertTokenizer
num_tokens_gpt integer The number of tokens using GPT2TokenizerFast
num_faces integer The number of faces in the image detected by SCRFD
clip_similarity_vitb32 float The cosine similarity between text and image(ViT-B/32) embeddings by OpenAI CLIP
clip_similarity_vitl14 float The cosine similarity between text and image(ViT-L/14) embeddings by OpenAI CLIP
nsfw_score_opennsfw2 float The NSFW score of the image by OpenNSFW2
nsfw_score_gantman float The NSFW score of the image by GantMan/NSFW
watermark_score float The watermark probability of the image by our internal model
aesthetic_score_laion_v2 float The aesthetic score of the image by LAION-Aesthetics-Predictor-V2

Statistics

  • Statistics for numeric columns
width height text_length word_count num_tokens_bert num_tokens_gpt num_faces
mean 621.78 540.99 68.53 11.13 15.75 17.24 0.60
min 200 200 6 3 1 3 0
max 21449 22507 1000 323 811 1523 736
watermark_score clip_similarity_vitb32 clip_similarity_vitl14 aesthetic_score_laion_v2 nsfw_score_opennsfw2
mean 0.178544 0.291266 0.254632 4.769132 0.012903
min 0.0 -0.080871 -0.176269 1.171712 0.0
max 1.0 0.591796 0.581542 8.082607 0.499755
  • Image Size statistics_image_size.png
  • CLIP Similarity statistics_clip_similarity.png
  • Text Length & Word Size img.png
  • Watermark score img.png
  • Aesthetic score img.png
  • Number of faces img.png
  • For more detailed statistics on COYO-700M, please see the Data Studio report on COYO-700M.

Getting Started

Download

Usage

Experiments

We empirically validated the quality of COYO dataset by re-implementing popular models such as ALIGN, unCLIP, and ViT. We trained these models on COYO-700M or its subsets from scratch, achieving competitive performance to the reported numbers or generated samples in the original papers. Since this observation supports the high quality of our dataset, we hope it to be continuously updated with open collaboration. Our pre-trained models and training codes will be released soon along with the technical report.

ALIGN

Model Data ImageNet KNN COCO I2T COCO T2I
EfficientNet-B7 + BERT-base ALIGN-1.8B 69.300 55.400 41.700
EfficientNet-B7 + BERT-base COYO-700M 68.618 59.000 42.419
  • Our experiment setup followed ALIGN.
    • We increased the batch size from 16K to 64K and reduced training steps by 1/4 for faster training.
A high quality picture of a medieval knight with golden armor A person with the head of a cat in the style of Andy Warhol
A pencil drawing of an astronaut riding a horse Goryeo celadon in the shape of darth vader
  • We implemented the smaller version of unCLIP to validate the effectiveness of COYO for the text-conditional generation tasks.
  • Specifically, we tried to reproduce three components of the original unCLIP: diffusion-based prior, decoder with some modifications, and super-resolution model for upscaling 64x64 into 256x256px.
  • Detailed information on our modified version of unCLIP and quantitative analysis would be included in the upcoming technical report.

ViT

Model Data ImageNet
Validation
Top-1 Acc
ViT-L/16 JFT-300M 87.76%
ViT-L/16 COYO-Labeled-300M 87.24%
  • We also provide COYO-Labeled-300M by adding machine-generated vision labels to a subset of COYO-700M for comparison with the JFT-300M.
    • We first removed the duplicated images by image_phash.
    • Then, we labeled 300M unique images into 21,841 classes by EfficientNetV2-XL trained with ImageNet-21K dataset.
  • Our experiment setup followed ViT.
  • We also provide vit pre-training, fine-tuning code for reproducibility with weight files.

Citation

If you apply this dataset to any project and research, please cite our code:

@misc{kakaobrain2022coyo-700m,
  title         = {COYO-700M: Image-Text Pair Dataset},
  author        = {Byeon, Minwoo and Park, Beomhee and Kim, Haecheon and Lee, Sungjun and Baek, Woonhyuk and Kim, Saehoon},
  year          = {2022},
  howpublished  = {\url{https://github.com/kakaobrain/coyo-dataset}},
}

People

Disclaimer & Content Warning

The COYO dataset is recommended to be used for research purposes. Kakao Brain tried to construct a "Safe" dataset when building the COYO dataset. (See Data Filtering Section) Kakao Brain is constantly making efforts to create more "Safe" datasets. However, despite these efforts, this large-scale dataset was not hand-picked by humans to avoid the risk due to its very large size (over 700M). Keep in mind that the unscreened nature of the dataset means that the collected images can lead to strongly discomforting and disturbing content for humans. The COYO dataset may contain some inappropriate data, and any problems resulting from such data are the full responsibility of the user who used it. Therefore, it is strongly recommended that this dataset be used only for research, keeping this in mind when using the dataset, and Kakao Brain does not recommend using this dataset as it is without special processing to clear inappropriate data to create commercial products.

License

The COYO dataset of Kakao Brain is licensed under CC-BY-4.0 License. The full license can be found in the LICENSE.cc-by-4.0 file. The dataset includes “Image URL” and “Text” collected from various sites by analyzing Common Crawl data, an open data web crawling project. The collected data (images and text) is subject to the license to which each content belongs.

Obligation to use

While Open Source may be free to use, that does not mean it is free of obligation. To determine whether your intended use of the COYO dataset is suitable for the CC-BY-4.0 license, please consider the license guide. If you violate the license, you may be subject to legal action such as the prohibition of use or claim for damages depending on the use.

Contact

COYO dataset was released as an open source in the hope that it will be helpful to many research institutes and startups for research purposes. We look forward to contacting us from various places who wish to cooperate with us.

coyo@kakaobrain.com