CLIP inference in plain C/C++ with no extra dependencies

This is a dependency free implementation of well known CLIP by OpenAI, thanks to the great work in GGML. You can use it to work with CLIP models from both OpenAI and LAION in Transformers format.

CLIP is deployed for several task from semantic image search to zero-shot image labeling. It's also a part of Stable Diffusion and and the recently emerging field of large multimodal models (LMM). This repo is aimed at powering useful applications based on such models on computation- or memory-constraint devices. 4-bit quantized CLIP is only 85.6 MB!

• Dependency-free and lightweight inference thanks to ggml.
• 4-bit, 5-bit and 8-bit quantization support.
• Support inference with text-only, vision-only and two-tower model variants. It might be relevant to use a single modality in certain cases, as in encoders for large multimodal models, or building and/or searching for semantic image search.
• Dependency free Python binding without relying any large third-party packages. No need for Numpy, TensorFlow, PyTorch, ONNX etc. In fact, nothing more than the standard Python library.

clip.cpp also has a short startup time compared to large ML frameworks, which makes it suitable for serverless deployments where the cold start is an issue.

• 01/27/2024: Clojure bindings available, clip.clj.
• 09/27/2023: clip.cpp now uses a new model file structure in GGUF format. This is a breaking change. See bwlow for more information.
• 09/14/2023: All functions are C-compatible now. zsl example is updated to match Huggingface's zero-shot behavior in the zero-shot pipeline.
• 09/11/2023: Introduce Python bindings.
• 07/12/2023: Batch inference support for image encoding.
• 07/11/2023: Semantic image search example directly in C++.

PIL uses a two-pass convolutions-based bicubic interpolation in resizing with antialiasing applied. In Pytorch, antialiasing is optional. It needs some extra attention to implement this preprocessing logic that matches their results numerically. However, I found that linear interpolation is also good enough for both comparison of different embeddings from this implementation and also comparison of an embedding from this implementation and another one from Transformers. So let's use it until we craft a proper bicubic interpolation.

Preconverted Models can be found in HuggingFace Repositories tagged with clip-cpp-gguf. If you want to do conversion yourself for some reason, see below for how. Otherwise, download a model of your choice from the link above and then feel free to jump to the building section.

You can convert CLIP models from OpenAI and LAION in Transformers format. Apparently, LAION's models outperform OpenAI models in several benchmarks, so they are recommended.

1. Clone the model repo from HF Hub:

git lfs init

git clone https://huggingface.co/laion/CLIP-ViT-B-32-laion2B-s34B-b79K

2. Clone this repository:

git clone --recurse-submodules https://github.com/monatis/clip.cpp.git

cd clip.cpp/models

3. You need to install pytorch and transformers packages if you haven't already. Otherwise, you can skip this step:

pip install -r requirements.txt

4. Use models/convert_hf_to_gguf.py to convert it to GGUF format:

python convert_hf_to_gguf.py -m ../../CLIP-ViT-B-32-laion2B-s34B-b79K

The output CLIP-ViT-B-32-laion2B-s34B-b79K_ggml-model-f16.bin file is in the model directory specified in the command above.

To see other options that you can use with the conversion script, run:

python convert_hf_to_gguf.py --help

git clone --recurse-submodules https://github.com/monatis/clip.cpp.git

cd clip.cpp

mkdir build

cd build

cmake -DCLIP_NATIVE=ON ..

make

And the binaries are in the ./bin directory.

Note: Some Mac devices report x86_64 instead of arm64 architecture. If this is the case see here for a discussion. I couldn't reproduce it on my Macbook M2 pro so cannot help further. If you know a solution that I can include in CMakeLists.txt please ping me here.

clip.cpp supports q4_0, q4_1, q5_0, q5_1 and q8_0 quantization types. You can quantize a model in f32 (recommended) or f16 to one of these types by using the ./bin/quantize binary.

usage: ./bin/quantize /path/to/ggml-model-f32.gguf /path/to/ggml-model-quantized.gguf type                              
  type = 2 - q4_0                                                                                                       
  type = 3 - q4_1                                                                                                       
  type = 6 - q5_0                                                                                                       
  type = 7 - q5_1                                                                                                       
  type = 8 - q8_0                                                                                                       

For example, you can run the following to convert the model to q5_1:

./bin/quantize ./CLIP-ViT-B-32-laion2B-s34B-b79K/ggml-model-f32.gguf ./CLIP-ViT-B-32-laion2B-s34B-b79K/ggml-model-q5_1.gguf 7

Now you can use ggml-model-q5_1.gguf just like the model in F16.

Currently we have 4 examples: main, zsl and image-search.

1. main is just for demonstrating the usage of API and optionally print out some verbose timings. It simply calculates the similarity between one image and one text passed as CLI args.

Usage: ./bin/main [options]

Options:  -h, --help: Show this message and exit
  -m <path>, --model <path>: path to model. Default: models/ggml-model-f16.gguf
  -t N, --threads N: Number of threads to use for inference. Default: 4
  --text <text>: Text to encode. At least one text should be specified
  --image <path>: Path to an image file. At least one image path should be specified
  -v <level>, --verbose <level>: Control the level of verbosity. 0 = minimum, 2 = maximum. Default: 1

2. zsl is a zero-shot image labeling example. It labels an image with one of the labels. The CLI args are the same as in main, but you must specify multiple --text arguments to specify the labels.

3. image-search is an example for semantic image search with USearch. You must enable CLIP_BUILD_IMAGE_SEARCH option to compile it, and the dependency will be automatically fetched by cmake:

mkdir build

cd build

cmake -DCLIP_BUILD_IMAGE_SEARCH=ON ..

make

See examples/image-search/README.md for more info and usage.

4. extract is a tool to extract vectors to *.npy files in Numpy format. It is again dependency-free, i.e., no need for Numpy.

CLI arguments are the same as in main and zsl, but you can pass one or multiple --text and/or --image arguments. Either of them or a combination of both allowed.

Note: Support for processing *.txt files and image directories is on the way.

You can use clip.cpp in Python with no third-party libraries (no dependencies other than standard Python libraries). It uses ctypes to load a dynamically linked library (DLL) to interface the implementation in C/C++.

If you are on an X64 Linux distribution, you can simply Pip-install it with AVX2 support:

pip install clip_cpp

Colab Notebook available for quick experiment :

If you are on another operating system or architecture, or if you want to make use of support for instruction sets other than AVX2 (e.g., AVX512), you can build it from source.

All you need is to compile with -DBUILD_SHARED_LIBS=ON option to get the required DLL.

mkdir build

cd build

cmake -DBUILD_SHARED_LIBS=ON ..

make

And find the libclip.so binary in the build directory. See examples/python_bindings/README.md for more info and usage.

You can use the benchmarking utility to compare the performances of different checkpoints and quantization types.

usage: ./bin/benchmark <model_path> <images_dir> <num_images_per_dir> [output_file]

model_path: path to CLIP model in GGML format
images_dir: path to a directory of images where images are organized into subdirectories named classes
num_images_per_dir: maximum number of images to read from each one of subdirectories. if 0, read all files
output_file: optional. if specified, dump the output to this file instead of stdout

See tests/README.md for more info about benchmarking.

• Implement proper bicubic interpolation (PIL uses a convolutions-based algorithm, and it's more stable than affine transformations).