Skip to content

cene555/ru-clip-tiny

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

19 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

RuCLIPtiny

Zero-shot image classification model for Russian language


RuCLIP tiny (Russian Contrastive Language–Image Pretraining) is a neural network trained to work with different pairs (images, texts). Our model is based on ConvNeXt-tiny and DistilRuBert-tiny, and is supported by extensive research zero-shot transfer, computer vision, natural language processing, and multimodal learning.

Result evaluation

Our model achieved 46.62% top1 and 73.18% top5 zero-shot accuracy on CIFAR100

Examples

Open In Colab Evaluate & Simple usage

Open In Colab Finetuning

Open In Colab ONNX conversion and speed testing

Usage

Install rucliptiny module and requirements first. Use this trick

!gdown -O ru-clip-tiny.pkl https://drive.google.com/uc?id=1-3g3J90pZmHo9jbBzsEmr7ei5zm3VXOL
!pip install git+https://github.com/cene555/ru-clip-tiny.git

Example in 3 steps

Download CLIP image from repo

!wget -c -O CLIP.png https://github.com/openai/CLIP/blob/main/CLIP.png?raw=true
  1. Import libraries
from rucliptiny.predictor import Predictor
from rucliptiny import RuCLIPtiny
import torch

torch.manual_seed(1)
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
  1. Load model
model = RuCLIPtiny()
model.load_state_dict(torch.load('ru-clip-tiny.pkl'))
model = model.to(device).eval()
  1. Use predictor to get probabilities
predictor = Predictor()

classes = ['диаграмма', 'собака', 'кошка']
text_probs = predictor(model=model, images_path=["CLIP.png"],
                       classes=classes, get_probs=True,
                       max_len=77, device=device)

Cosine similarity Visualization Example

image

Speed Testing

NVIDIA Tesla K80 (Google Colab session)

TORCH batch encode_image encode_text total
RuCLIPtiny 2 0.011 0.004 0.015
RuCLIPtiny 8 0.011 0.004 0.015
RuCLIPtiny 16 0.012 0.005 0.017
RuCLIPtiny 32 0.014 0.005 0.019
RuCLIPtiny 64 0.013 0.006 0.019

We would like to express my gratitude to Sber AI for the grants provided, for which research was carried out, as part of the Artificial Intelligence International Junior Contest (AIIJC)

About

RuCLIP tiny (Russian Contrastive Language–Image Pretraining) is a neural network trained to work with different pairs (images, texts).

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published