COYO-Labeled-300M

🐺 COYO-Labeled-300M: Image-Multi-Label Pair Dataset

COYO-Labeled-300M

COYO-Labeled-300M is a dataset of machine-labeled 300M image-multi-label pairs. We labeled subset of COYO-700M with a large model (efficientnetv2-xl) trained on imagenet-21k. We followed the same evaluation pipeline as in efficientnet-v2. The labels are top 50 most likely labels out of 21,841 classes from imagenet-21k. The label probabilies are also provided with labels so that the user can select threshold of their choice for multi-label classification use or can take top-1 class for single class classification use.

In other words, COYO-Labeled-300M is a ImageNet-like dataset. Instead of human labeled 1.25 million samples, it's machine-labeled 300 million samples. This dataset is similar to JFT-300M which is not released to the public.

We found that our ViT implementation trained on COYO-Labeled-300M performs similar to the performance numbers in the ViT paper trained on JFT-300M.

We also provide weights for the pretrained ViT model on COYO-Labeled-300M as well as its training & fine-tuning code.

The basic instruction, licenses and contributors are the same as for the coyo-700m.

Dataset Preview

id	url	imagehash	labels	label_probs	width	height
315		daf5a50aae4aa54a	[8087, 11054, 808...	[0.4453125, 0.304...	240	240
463		8f3fd06161c41e3e	[12213, 12215, 12...	[0.71728516, 0.38...	219	230
601		8e70855e6ba3f825	[4945, 4953, 4942...	[0.8496094, 0.069...	320	240
1105		ea90d5978594cbc6	[12352, 12658, 12...	[0.2763672, 0.267...	310	310
2308		c0cf983c7c7074f4	[6728, 14101, 158...	[0.32128906, 0.02...	702	336
3710		8f9694594b26371b	[15721, 15722, 49...	[0.4152832, 0.175...	600	467

Meta-Attributes

Attributes

name	type	description
id	long	Unique 64-bit integer ID generated by monotonically_increasing_id() which is the same value that is mapped with the existing COYO-700M.
url	string	The image URL extracted from the src attribute of the <img>
imagehash	string	The perceptual hash(pHash) of the image
labels	sequence[integer]	Inference results of EfficientNetV2-XL model trained on ImageNet-21K dataset (Top 50 indices among 21,841 classes)
label_probs	sequence[float]	Inference results of EfficientNetV2-XL model trained on ImageNet-21K dataset (Top 50 indices among 21,841 probabilites)
width	integer	The width of the image
height	integer	The height of the image

Statistics

• Statistics for threshold-based label distribution

Threshold	Labels per Image	Unique Labels	Sampling Ratio
0.00	50.00	21,841	100.00%
0.05	3.16	18,922	96.69%
0.10	1.98	18,471	85.83%
0.15	1.58	18,122	74.49%
0.20	1.37	17,781	64.30%
0.25	1.26	17,479	55.60%

• Image Size

  

• Top1 Label Distribution

Getting Started

Download

• You can download the dataset from Huggingface Dataset
• For more information on downloading the image dataset, please refer to download/README.md.

Experiments

• We validated the quality of COYO-Labeled-300M dataset by re-implementing popular models, ViT.
• We also provide pretraining & fine-tuning code and weights file for ViT
• We report here the performance of vit trained on the coyo dataset.