clip-cpp 0.5.0

CLIP inference with no big dependencies such as PyTorch, TensorFlow, Numpy or ONNX

Python bindings for clip.cpp

This package provides basic Python bindings for clip.cpp.

It requires no third-party libraries and no big dependencies such as PyTorch, TensorFlow, Numpy, ONNX etc.

Install

If you are on a X64 Linux distribution, you can simply Pip-install it:

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. Se clip.cpp for more info.

All you need to do is to compile with the -DBUILD_SHARED_LIBS=ON option and copy libclip.so to examples/python_bindings/clip_cpp.

Usage

from clip_cpp import Clip

## you can either pass repo_id or .gguf file
## you can type `clip-cpp-models` in your terminal to see what models are available for download
## in case you pass repo_id and it has more than .bin file
## it's recommended to specify which file to download with `model_file`
repo_id = 'mys/ggml_CLIP-ViT-B-32-laion2B-s34B-b79K'
model_file = 'CLIP-ViT-B-32-laion2B-s34B-b79K_ggml-model-f16.gguf'

model = Clip(
    model_path_or_repo_id=repo_id,
    model_file=model_file,
    verbosity=2
)

text_2encode = 'cat on a Turtle'

tokens = model.tokenize(text_2encode)
text_embed = model.encode_text(tokens)

## load and extract embeddings of an image from the disk
image_2encode = '/path/to/cat.jpg'
image_embed = model.load_preprocess_encode_image(image_2encode)

## calculate the similarity between the image and the text
score = model.calculate_similarity(text_embed, image_embed)

# Alternatively, you can just do:
# score = model.compare_text_and_image(text, image_path)

print(f"Similarity score: {score}")

Clip Class

The Clip class provides a Python interface to clip.cpp, allowing you to perform various tasks such as text and image encoding, similarity scoring, and text-image comparison. Below are the constructor and public methods of the Clip class:

Constructor

def __init__(
    self, model_path_or_repo_id: str,
    model_file: Optional[str] = None,
    revision: Optional[str] = None,
    verbosity: int = 0):

• Description: Initializes a Clip instance with the specified CLIP model file and optional verbosity level.
• model_path_or_repo_id (str): The path to the CLIP model file file | HF repo_id.
• model_file (str, optional): if model_path_or_repo_id is repo_id that has multiple .gguf files you can specify which .gguf file to download
• verbosity (int, optional): An integer specifying the verbosity level (default is 0).

Public Methods

1. vision_config

@property
def vision_config(self) -> Dict[str, Any]:

• Description: Retrieves the configuration parameters related to the vision component of the CLIP model.

2. text_config

@property
def text_config(self) -> Dict[str, Any]:

• Description: Retrieves the configuration parameters related to the text component of the CLIP model.

3. tokenize

def tokenize(self, text: str) -> List[int]:

• Description: Tokenizes a text input into a list of token IDs.
• text (str): The input text to be tokenized.

4. encode_text

def encode_text(
    self, tokens: List[int], n_threads: int = os.cpu_count(), normalize: bool = True
) -> List[float]:

• Description: Encodes a list of token IDs into a text embedding.
• tokens (List[int]): A list of token IDs obtained through tokenization.
• n_threads (int, optional): The number of CPU threads to use for encoding (default is the number of CPU cores).
• normalize (bool, optional): Whether or not to normalize the output vector (default is True).

5. load_preprocess_encode_image

def load_preprocess_encode_image(
    self, image_path: str, n_threads: int = os.cpu_count(), normalize: bool = True
) -> List[float]:

• Description: Loads an image, preprocesses it, and encodes it into an image embedding.
• image_path (str): The path to the image file to be encoded.
• n_threads (int, optional): The number of CPU threads to use for encoding (default is the number of CPU cores).
• normalize (bool, optional): Whether or not to normalize the output vector (default is True).

6. calculate_similarity

def calculate_similarity(
    self, text_embedding: List[float], image_embedding: List[float]
) -> float:

• Description: Calculates the similarity score between a text embedding and an image embedding.
• text_embedding (List[float]): The text embedding obtained from encode_text.
• image_embedding (List[float]): The image embedding obtained from load_preprocess_encode_image.

7. compare_text_and_image

def compare_text_and_image(
    self, text: str, image_path: str, n_threads: int = os.cpu_count()
) -> float:

• Description: Compares a text input and an image file, returning a similarity score.
• text (str): The input text.
• image_path (str): The path to the image file for comparison.
• n_threads (int, optional): The number of CPU threads to use for encoding (default is the number of CPU cores).

8. zero_shot_label_image

def zero_shot_label_image(
        self, image_path: str, labels: List[str], n_threads: int = os.cpu_count()
    ) -> Tuple[List[float], List[int]]:

• Description: Zero-shot labels an image with given candidate labels, returning a tuple of sorted scores and indices.
• image_path (str): The path to the image file to be labelled.
• labels (List[str]): A list of candidate labels to be scored.
• n_threads (int, optional): The number of CPU threads to use for encoding (default is the number of CPU cores).

9. __del__

def __del__(self):

• Description: Destructor that frees resources associated with the Clip instance.

With the Clip class, you can easily work with the CLIP model for various natural language understanding and computer vision tasks.

Example

A basic example can be found in the clip.cpp examples.

python example_main.py --help
usage: clip [-h] -m MODEL [-fn FILENAME] [-v VERBOSITY] -t TEXT [TEXT ...] -i IMAGE                                     
                                                                                                                        
optional arguments:                                                                                                     
  -h, --help            show this help message and exit                                                                 
  -m MODEL, --model MODEL                                                                                               
                        path to GGML file or repo_id                                                                    
  -fn FILENAME, --filename FILENAME                                                                                     
                        path to GGML file in the Hugging face repo                                                      
  -v VERBOSITY, --verbosity VERBOSITY                                                                                   
                        Level of verbosity. 0 = minimum, 2 = maximum                                                    
  -t TEXT [TEXT ...], --text TEXT [TEXT ...]                                                                            
                        text to encode. Multiple values allowed. In this case, apply zero-shot labeling                 
  -i IMAGE, --image IMAGE                                                                                               
                        path to an image file