ruclip-vit-base-patch16-224

RuCLIP (Russian Contrastive Language–Image Pretraining) is a multimodal model for obtaining images and text similarities and rearranging captions and pictures. RuCLIP builds on a large body of work on zero-shot transfer, computer vision, natural language processing and multimodal learning.

Model was trained by Sber AI and SberDevices teams.

• Task: text ranking; image ranking; zero-shot image classification;
• Type: encoder
• Num Parameters: 150M
• Training Data Volume: 240 million text-image pairs
• Language: Russian
• Context Length: 77
• Transformer Layers: 12
• Transformer Width: 512
• Transformer Heads: 8
• Image Size: 224
• Vision Layers: 12
• Vision Width: 768
• Vision Patch Size: 16

Usage Github

pip install ruclip

clip, processor = ruclip.load("ruclip-vit-base-patch16-224", device="cuda")

Performance

We have evaluated the performance on the following datasets:

Dataset	Metric Name	Metric Result
Food101	acc	0.552
CIFAR10	acc	0.810
CIFAR100	acc	0.496
Birdsnap	acc	0.117
SUN397	acc	0.462
Stanford Cars	acc	0.487
DTD	acc	0.401
MNIST	acc	0.464
STL10	acc	0.932
PCam	acc	0.505
CLEVR	acc	0.128
Rendered SST2	acc	0.527
ImageNet	acc	0.401
FGVC Aircraft	mean-per-class	0.043
Oxford Pets	mean-per-class	0.595
Caltech101	mean-per-class	0.775
Flowers102	mean-per-class	0.388
HatefulMemes	roc-auc	0.516

Authors

• Alex Shonenkov: Github, Kaggle GM
• Daniil Chesakov: Github
• Denis Dimitrov: Github
• Igor Pavlov: Github