CLIP embeddings#

CLIP (Contrastive Language-Image Pre-training) is a machine learning model developed by OpenAI. It is versatile and excels in tasks like zero-shot learning, image classification, and image-text matching without needing specific training for each task. This makes it ideal for a wide range of applications, including content recommendation, image captioning, visual search, and automated content moderation.

This document demonstrates how to build a CLIP application using BentoML, powered by the clip-vit-base-patch32 model.


Install dependencies#

Clone the project repository and install all the dependencies.

git clone
cd BentoClip
pip install -r requirements.txt

Create a BentoML Service#

Define a BentoML Service in a file to wrap the capabilities of the CLIP model, making them accessible and easy to use in a wide range of applications. Here is an example file in the project:
import bentoml
from PIL.Image import Image
import numpy as np
from typing import Dict
from typing import List
from pydantic import Field

MODEL_ID = "openai/clip-vit-base-patch32"

        "memory" : "4Gi"
class CLIP:

    def __init__(self) -> None:
        import torch
        from transformers import CLIPModel, CLIPProcessor
        self.device = "cuda" if torch.cuda.is_available() else "cpu"
        self.model = CLIPModel.from_pretrained(MODEL_ID).to(self.device)
        self.processor = CLIPProcessor.from_pretrained(MODEL_ID)
        self.logit_scale = self.model.logit_scale.item() if self.model.logit_scale.item() else 4.60517
        print("Model clip loaded", "device:", self.device)

    async def encode_image(self, items: List[Image]) -> np.ndarray:
        generate the 512-d embeddings of the images
        inputs = self.processor(images=items, return_tensors="pt", padding=True).to(self.device)
        image_embeddings = self.model.get_image_features(**inputs)
        return image_embeddings.cpu().detach().numpy()

    async def encode_text(self, items: List[str]) -> np.ndarray:
        generate the 512-d embeddings of the texts
        inputs = self.processor(text=items, return_tensors="pt", padding=True).to(self.device)
        text_embeddings = self.model.get_text_features(**inputs)
        return text_embeddings.cpu().detach().numpy()

    async def rank(self, queries: List[Image], candidates : List[str] = Field(["picture of a dog", "picture of a cat"], description="list of description candidates")) -> Dict[str, List[List[float]]]:
        return the similarity between the query images and the candidate texts
        # Encode embeddings
        query_embeds = await self.encode_image(queries)
        candidate_embeds = await self.encode_text(candidates)

        # Compute cosine similarities
        cosine_similarities = self.cosine_similarity(query_embeds, candidate_embeds)
        logit_scale = np.exp(self.logit_scale)
        # Compute softmax scores
        prob_scores = self.softmax(logit_scale * cosine_similarities)
        return {
            "probabilities": prob_scores.tolist(),
            "cosine_similarities" : cosine_similarities.tolist(),

    def cosine_similarity(query_embeds, candidates_embeds):
        # Normalize each embedding to a unit vector
        query_embeds /= np.linalg.norm(query_embeds, axis=1, keepdims=True)
        candidates_embeds /= np.linalg.norm(candidates_embeds, axis=1, keepdims=True)

        # Compute cosine similarity
        cosine_similarities = np.matmul(query_embeds, candidates_embeds.T)

        return cosine_similarities

    def softmax(scores):
        # Compute softmax scores (probabilities)
        exp_scores = np.exp(
            scores - np.max(scores, axis=-1, keepdims=True)
        )  # Subtract max for numerical stability
        return exp_scores / np.sum(exp_scores, axis=-1, keepdims=True)

Here is a breakdown of the Service code:

  1. The script uses the @bentoml.service decorator to annotate the CLIP class as a BentoML Service. You can set more configurations for the Service as needed with the decorator.

  2. In the __init__ method, the CLIP model and processor are loaded based on the specified MODEL_ID. The model is transferred to a GPU if available, otherwise, it uses the CPU. The logit_scale is set to the model’s logit scale or a default value if not available.

  3. The Service defines the following three API endpoints:

    • encode_image: Takes a list of images and generates 512-dimensional embeddings for them.

    • encode_text: Takes a list of text strings and generates 512-dimensional embeddings for them.

    • rank: Computes the similarity between a list of query images and candidate text descriptions. It uses the embeddings generated by the previous two endpoints to calculate cosine similarities and softmax scores, indicating how closely each text candidate matches each image.

  4. The Service defines the following two static methods:

    • cosine_similarity: Computes the cosine similarity between query embeddings and candidate embeddings. It normalizes each embedding to a unit vector before computing the similarity.

    • softmax: Computes softmax scores from the similarity scores, turning them into probabilities. This method includes a numerical stability trick by subtracting the maximum score before exponentiation.

This Service can be used for the following use cases:

  • Image and text embedding: Convert images and text into embeddings, which can then be utilized for various machine learning tasks like clustering and similarity search.

  • Image-text matching: Find the most relevant text descriptions for a set of images, which is useful in applications like image captioning and content recommendation.

Run bentoml serve in your project directory to start the Service.

$ bentoml serve service:CLIP

2024-01-08T09:07:28+0000 [INFO] [cli] Starting production HTTP BentoServer from "service:CLIP" listening on http://localhost:3000 (Press CTRL+C to quit)
Model clip loaded device: cuda

The server is active at http://localhost:3000. You can interact with it in different ways.

curl -s \
    -X POST \
    -F 'items=@image.jpg' \
import bentoml
from pathlib import Path

with bentoml.SyncHTTPClient("http://localhost:3000") as client:
    result = client.encode_image(

Visit http://localhost:3000, scroll down to Service APIs, and select the desired API endpoint for interaction.


This is the image sent in the request. Expected output:


Deploy to BentoCloud#

After the Service is ready, you can deploy the project to BentoCloud for better management and scalability. Sign up for a BentoCloud account and get $30 in free credits.

First, specify a configuration YAML file (bentofile.yaml) to define the build options for your application. It is used for packaging your application into a Bento. Here is an example file in the project:

service: "service:CLIP"
  owner: bentoml-team
  project: gallery
- "*.py"
  requirements_txt: "./requirements.txt"

Create an API token with Developer Operations Access to log in to BentoCloud, then run the following command to deploy the project.

bentoml deploy .

Once the Deployment is up and running on BentoCloud, you can access it via the exposed URL.



For custom deployment in your own infrastructure, use BentoML to generate an OCI-compliant image.