RuCLIP

Zero-shot image classification model for Russian language

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. This repo has the prototypes model of OpenAI CLIP's Russian version following this paper.

Models

• ruclip-vit-base-patch32-224 🤗
• ruclip-vit-base-patch16-224 🤗
• ruclip-vit-large-patch14-224 🤗
• ruclip-vit-base-patch32-384 🤗
• ruclip-vit-large-patch14-336 🤗
• ruclip-vit-base-patch16-384 🤗

Installing

pip install ruclip==0.0.2

Usage

Standart RuCLIP API

RuCLIP + SberVqgan

ONNX example

Init models

import ruclip

device = 'cuda'
clip, processor = ruclip.load('ruclip-vit-base-patch32-384', device=device)

Zero-Shot Classification [Minimal Example]

import torch
import base64
import requests
import matplotlib.pyplot as plt
from PIL import Image
from io import BytesIO

# prepare images
bs4_urls = requests.get('https://raw.githubusercontent.com/sberbank-ai/ru-dolph/master/pics/pipelines/cats_vs_dogs_bs4.json').json()
images = [Image.open(BytesIO(base64.b64decode(bs4_url))) for bs4_url in bs4_urls]

# prepare classes
classes = ['кошка', 'собака']
templates = ['{}', 'это {}', 'на картинке {}', 'это {}, домашнее животное']

# predict
predictor = ruclip.Predictor(clip, processor, device, bs=8, templates=templates)
with torch.no_grad():
    text_latents = predictor.get_text_latents(classes)
    pred_labels = predictor.run(images, text_latents)

# show results
f, ax = plt.subplots(2,4, figsize=(12,6))
for i, (pil_img, pred_label) in enumerate(zip(images, pred_labels)):
    ax[i//4, i%4].imshow(pil_img)
    ax[i//4, i%4].set_title(classes[pred_label])

Cosine similarity Visualization Example

Softmax Scores Visualization Example

Linear Probe and ZeroShot Correlation Results

Linear Probe Example

train = CIFAR100(root, download=True, train=True)
test = CIFAR100(root, download=True, train=False)

with torch.no_grad():
    X_train = predictor.get_image_latents((pil_img for pil_img, _ in train)).cpu().numpy()
    X_test = predictor.get_image_latents((pil_img for pil_img, _ in test)).cpu().numpy()
    y_train, y_test = np.array(train.targets), np.array(test.targets)

clf = LogisticRegression(solver='lbfgs', penalty='l2', max_iter=1000, verbose=1)
clf.fit(X_train, y_train)
y_pred = clf.predict(X_test)
accuracy = np.mean((y_test == y_pred).astype(np.float)) * 100.
print(f"Accuracy = {accuracy:.3f}")

>>> Accuracy = 75.680

Performance

We have evaluated the performance zero-shot image classification on the following datasets:

And for linear-prob evaluation:

Also, we have created speed comparison based on CIFAR100 dataset using Nvidia-V100 for evaluation:

Authors

• Alex Shonenkov: Github, Kaggle GM
• Daniil Chesakov: Github
• Denis Dimitrov: Github
• Igor Pavlov: Github
• Andrey Kuznetsov: Github
• Anastasia Maltseva: Github

Supported by

Social Media