Image Search with OpenAI CLIP

Implement image search with the OpenAI CLIP Model and Supabase Vector.

The OpenAI CLIP Model was trained on a variety of (image, text)-pairs. You can use the CLIP model for:

Text-to-Image / Image-To-Text / Image-to-Image / Text-to-Text Search
You can fine-tune it on your own image and text data with the regular SentenceTransformers training code.

SentenceTransformers provides models that allow you to embed images and text into the same vector space. You can use this to find similar images as well as to implement image search.

You can find the full application code as a Python Poetry project on GitHub.

Create a new Python Project with Poetry#

Poetry provides packaging and dependency management for Python. If you haven't already, install poetry via pip:

pip install poetry

Then initialize a new project:

poetry new image-search

Setup Supabase project#

If you haven't already, install the Supabase CLI, then initialize Supabase in the root of your newly created poetry project:

supabase init

Next, start your local Supabase stack:

supabase start

This will start up the Supabase stack locally and print out a bunch of environtment details, including your local DB URL. Make a note of that for later user.

Install the Dependencies#

We will need to add the following dependencies to our project:

vecs: Supabase Vector Python Client.
sentence-transformers: a framework for sentence, text and image embeddings (used with OpenAI CLIP model)
matplotlib: for displaying our image result

poetry add vecs sentence-transformers matplotlib

Import the necessary dependencies#

At the top of your main python script, import the dependencies and store your DB URL from above in a variable:

from PIL import Image
from sentence_transformers import SentenceTransformer
import vecs
from matplotlib import pyplot as plt
from matplotlib import image as mpimg

DB_CONNECTION = "postgresql://postgres:postgres@localhost:54322/postgres"

Create embeddings for your images#

In the root of your project, create a new folder called images and add some images. You can use the images from the example project on GitHub or you can find license free images on unsplash.

Next, create a seed method, which will create a new Supabase Vector Collection, generate embeddings for your images, and upsert the embeddings into your database:

def seed():
    # create vector store client
    vx = vecs.create_client(DB_CONNECTION)

    # create a collection of vectors with 3 dimensions
    images = vx.create_collection(name="image_vectors", dimension=512)

    # Load CLIP model
    model = SentenceTransformer('clip-ViT-B-32')

    # Encode an image:
    img_emb1 = model.encode(Image.open('./images/one.jpg'))
    img_emb2 = model.encode(Image.open('./images/two.jpg'))
    img_emb3 = model.encode(Image.open('./images/three.jpg'))
    img_emb4 = model.encode(Image.open('./images/four.jpg'))

    # add records to the *images* collection
    images.upsert(
        vectors=[
            (
                "one.jpg",        # the vector's identifier
                img_emb1,          # the vector. list or np.array
                {"type": "jpg"}   # associated  metadata
            ), (
                "two.jpg",
                img_emb2,
                {"type": "jpg"}
            ), (
                "three.jpg",
                img_emb3,
                {"type": "jpg"}
            ), (
                "four.jpg",
                img_emb4,
                {"type": "jpg"}
            )
        ]
    )
    print("Inserted images")

    # index the collection for fast search performance
    images.create_index()
    print("Created index")

Add this method as a script in your pyproject.toml file:

[tool.poetry.scripts]
seed = "image_search.main:seed"
search = "image_search.main:search"

After activating the virtual environtment with poetry shell you can now run your seed script via poetry run seed. You can inspect the generated embeddings in your local database by visiting the local Supabase dashboard at localhost:54323, selecting the vecs schema, and the image_vectors database.

Perform an Image Search from a Text Query#

With Supabase Vector we can easily query our embeddings. We can use either an image as search input or alternative we can generate an embedding from a string input and use that as the query input:

def search():
    # create vector store client
    vx = vecs.create_client(DB_CONNECTION)
    images = vx.get_collection(name="image_vectors")

    # Load CLIP model
    model = SentenceTransformer('clip-ViT-B-32')
    # Encode text query
    query_string = "a bike in front of a red brick wall"
    text_emb = model.encode(query_string)

    # query the collection filtering metadata for "type" = "jpg"
    results = images.query(
        query_vector=text_emb,              # required
        limit=1,                            # number of records to return
        filters={"type": {"$eq": "jpg"}},   # metadata filters
    )
    result = results[0]
    print(result)
    plt.title(result)
    image = mpimg.imread('./images/' + result)
    plt.imshow(image)
    plt.show()

By limiting the query to one result, we can show the most relevant image to the user. Finally we use matplotlib to show the image result to the user.

That's it, go ahead and test it out by running poetry run search and you will be presented with an image of a "bike in front of a red brick wall".

Conclusion#

With just a couple of lines of Python you are able to implement image search as well as reverse image search using OpenAI's CLIP model and Supabase Vector.

Edit this page on GitHub