chromadb/test/ef/test_multimodal_ef.py

from typing import Generator, cast
import numpy as np
import pytest
import chromadb
from chromadb.api.types import (
    Embeddable,
    EmbeddingFunction,
    Embeddings,
    Image,
    Document,
)
from chromadb.test.property.strategies import hashing_embedding_function
from chromadb.test.property.invariants import _exact_distances


# A 'standard' multimodal embedding function, which converts inputs to strings
# then hashes them to a fixed dimension.
class hashing_multimodal_ef(EmbeddingFunction[Embeddable]):
    def __init__(self) -> None:
        self._hef = hashing_embedding_function(dim=10, dtype=np.float_)

    def __call__(self, input: Embeddable) -> Embeddings:
        to_texts = [str(i) for i in input]
        embeddings = np.array(self._hef(to_texts))
        # Normalize the embeddings
        # This is so we can generate random unit vectors and have them be close to the embeddings
        embeddings /= np.linalg.norm(embeddings, axis=1, keepdims=True)
        return cast(Embeddings, embeddings.tolist())


def random_image() -> Image:
    return np.random.randint(0, 255, size=(10, 10, 3), dtype=np.int32)


def random_document() -> Document:
    return str(random_image())


@pytest.fixture
def multimodal_collection(
    default_ef: EmbeddingFunction[Embeddable] = hashing_multimodal_ef(),
) -> Generator[chromadb.Collection, None, None]:
    client = chromadb.Client()
    collection = client.create_collection(
        name="multimodal_collection", embedding_function=default_ef
    )
    yield collection
    client.clear_system_cache()


# Test adding and querying of a multimodal collection consisting of images and documents
def test_multimodal(
    multimodal_collection: chromadb.Collection,
    default_ef: EmbeddingFunction[Embeddable] = hashing_multimodal_ef(),
    n_examples: int = 10,
    n_query_results: int = 3,
) -> None:
    # Fix numpy's random seed for reproducibility
    random_state = np.random.get_state()
    np.random.seed(0)

    image_ids = [str(i) for i in range(n_examples)]
    images = [random_image() for _ in range(n_examples)]
    image_embeddings = default_ef(images)

    document_ids = [str(i) for i in range(n_examples, 2 * n_examples)]
    documents = [random_document() for _ in range(n_examples)]
    document_embeddings = default_ef(documents)

    # Trying to add a document and an image at the same time should fail
    with pytest.raises(
        ValueError, match="You can only provide documents or images, not both."
    ):
        multimodal_collection.add(
            ids=image_ids[0], documents=documents[0], images=images[0]
        )

    # Add some documents
    multimodal_collection.add(ids=document_ids, documents=documents)
    # Add some images
    multimodal_collection.add(ids=image_ids, images=images)

    # get() should return all the documents and images
    # ids corresponding to images should not have documents
    get_result = multimodal_collection.get(include=["documents"])
    assert len(get_result["ids"]) == len(document_ids) + len(image_ids)
    for i, id in enumerate(get_result["ids"]):
        assert id in document_ids or id in image_ids
        assert get_result["documents"] is not None
        if id in document_ids:
            assert get_result["documents"][i] == documents[document_ids.index(id)]
        if id in image_ids:
            assert get_result["documents"][i] is None

    # Generate a random query image
    query_image = random_image()
    query_image_embedding = default_ef([query_image])

    image_neighbor_indices, _ = _exact_distances(
        query_image_embedding, image_embeddings + document_embeddings
    )
    # Get the ids of the nearest neighbors
    nearest_image_neighbor_ids = [
        image_ids[i] if i < n_examples else document_ids[i % n_examples]
        for i in image_neighbor_indices[0][:n_query_results]
    ]

    # Generate a random query document
    query_document = random_document()
    query_document_embedding = default_ef([query_document])
    document_neighbor_indices, _ = _exact_distances(
        query_document_embedding, image_embeddings + document_embeddings
    )
    nearest_document_neighbor_ids = [
        image_ids[i] if i < n_examples else document_ids[i % n_examples]
        for i in document_neighbor_indices[0][:n_query_results]
    ]

    # Querying with both images and documents should fail
    with pytest.raises(ValueError):
        multimodal_collection.query(
            query_images=[query_image], query_texts=[query_document]
        )

    # Query with images
    query_result = multimodal_collection.query(
        query_images=[query_image], n_results=n_query_results, include=["documents"]
    )

    assert query_result["ids"][0] == nearest_image_neighbor_ids

    # Query with documents
    query_result = multimodal_collection.query(
        query_texts=[query_document], n_results=n_query_results, include=["documents"]
    )

    assert query_result["ids"][0] == nearest_document_neighbor_ids
    np.random.set_state(random_state)


@pytest.mark.xfail
def test_multimodal_update_with_image(
    multimodal_collection: chromadb.Collection,
) -> None:
    # Updating an entry with an existing document should remove the documentß

    document = random_document()
    image = random_image()
    id = "0"

    multimodal_collection.add(ids=id, documents=document)

    multimodal_collection.update(ids=id, images=image)

    get_result = multimodal_collection.get(ids=id, include=["documents"])
    assert get_result["documents"] is not None
    assert get_result["documents"][0] is None