DeliriumV01D / RuCLIP

Unofficial c++ LibTorch implementation of RuCLIP (Sber AI)

RuCLIP

Unofficial c++ LibTorch implementation of RuCLIP (Sber AI)

RuCLIP (Russian Contrastive Language–Image Pretraining) is a multimodal model for obtaining images and text similarities and rearranging captions and pictures. Original PyTorch code: https://github.com/ai-forever/ru-clip Original CLIP OpenAI paper: https://arxiv.org/pdf/2103.00020.pdf

Dependencies:

libTorch(https://pytorch.org), YouTokenToMe tokenizer https://github.com/VKCOM/YouTokenToMe, OpenCV(https://opencv.org/releases/), nlohmann json(https://github.com/nlohmann/json)

Test

Test images:

Test labels: {"кот", "медведь", "лиса"}

RuCLIP probabilities: 0.8879 0.0063 0.1058 0.0014 0.0026 0.9960 0.0002 0.9994 0.0003

For minimal example see main.cpp:

CLIP clip = FromPretrained("..//data//ruclip-vit-large-patch14-336");
clip->to(device);

RuCLIPProcessor processor(
	"..//data//ruclip-vit-large-patch14-336//bpe.model",
	INPUT_IMG_SIZE,
	77,
	{ 0.48145466, 0.4578275, 0.40821073 },
	{ 0.26862954, 0.26130258, 0.27577711 }
);

//Загрузить картинки
std::vector <cv::Mat> images;
images.push_back(cv::imread("..//data//test_images//1.png", cv::ImreadModes::IMREAD_COLOR));
images.push_back(cv::imread("..//data//test_images//2.jpg", cv::ImreadModes::IMREAD_COLOR));
images.push_back(cv::imread("..//data//test_images//3.jpg", cv::ImreadModes::IMREAD_COLOR));
//resize->[336, 336]
for (auto &it : images)
	 cv::resize(it, it, cv::Size(INPUT_IMG_SIZE, INPUT_IMG_SIZE));

//Завести метки
std::vector<std::string> labels;
labels = {"кот", "медведь", "лиса"};

auto dummy_input = processor(labels, images);
try {
	torch::Tensor logits_per_image = clip->forward(dummy_input.first.to(device), dummy_input.second.to(device));
	torch::Tensor logits_per_text = logits_per_image.t();
	auto probs = logits_per_image.softmax(/*dim = */-1).detach().cpu();
	std::cout << "probs per image: " << probs << std::endl;

To load weights you need to export checkpoint to jit format:

var=(torch.ones((1,77)).long(), torch.ones((1,3,336,336)))

traced_script_module = torch.jit.trace(model, var)

traced_script_module.save("gdrive/My Drive/ruclip-vit-large-patch14-336.zip")

Run a defaul example

On Windows with CMake build can be some problems:

1. nvToolsExt not found with CUDA 12.1. You can set include directory "C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v12.1/include/nvtx3" and any library
2. After CMake generate step remove from linker dependency nvToolsExt library - they are not used in project and nvToolsExt64_1.dll has alredy in libTorch

Run example with bat file:

        set IMGS=C:\work\clip\ruclip_\CLIP\data\test_images\1.png,C:\work\clip\ruclip_\CLIP\data\test_images\2.jpg,C:\work\clip\ruclip_\CLIP\data\test_images\3.jpg
        set LABELS=cat,bear,fox

        set CLIP=C:\work\clip\ruclip_\CLIP\data\ruclip-vit-large-patch14-336
        set BPE=C:\work\clip\ruclip_\CLIP\data\ruclip-vit-large-patch14-336\bpe.model
        set SIZE=336

        RuCLIP.exe --imgs=%IMGS% --text=%LABELS% --clip=%CLIP% --bpe=%BPE% --img_size=%SIZE%