fastembed-bio 0.1.0__py3-none-any.whl

fastembed/late_interaction_multimodal/colpali.py

@@ -0,0 +1,327 @@
+from typing import Any, Iterable, Sequence, Type
+
+import numpy as np
+from tokenizers import Encoding
+
+from fastembed.common import OnnxProvider, ImageInput
+from fastembed.common.onnx_model import OnnxOutputContext
+from fastembed.common.types import NumpyArray, Device
+from fastembed.common.utils import define_cache_dir, iter_batch
+from fastembed.late_interaction_multimodal.late_interaction_multimodal_embedding_base import (
+    LateInteractionMultimodalEmbeddingBase,
+)
+from fastembed.late_interaction_multimodal.onnx_multimodal_model import (
+    OnnxMultimodalModel,
+    TextEmbeddingWorker,
+    ImageEmbeddingWorker,
+)
+from fastembed.common.model_description import DenseModelDescription, ModelSource
+
+supported_colpali_models: list[DenseModelDescription] = [
+    DenseModelDescription(
+        model="Qdrant/colpali-v1.3-fp16",
+        dim=128,
+        description="Text embeddings, Multimodal (text&image), English, 50 tokens query length truncation, 2024.",
+        license="mit",
+        size_in_GB=6.5,
+        sources=ModelSource(hf="Qdrant/colpali-v1.3-fp16"),
+        additional_files=["model.onnx_data"],
+        model_file="model.onnx",
+    ),
+]
+
+
+class ColPali(LateInteractionMultimodalEmbeddingBase, OnnxMultimodalModel[NumpyArray]):
+    QUERY_PREFIX = "Query: "
+    BOS_TOKEN = "<s>"
+    PAD_TOKEN = "<pad>"
+    QUERY_MARKER_TOKEN_ID = [2, 5098]
+    IMAGE_PLACEHOLDER_SIZE = (3, 448, 448)
+    EMPTY_TEXT_PLACEHOLDER = np.array(
+        [257152] * 1024 + [2, 50721, 573, 2416, 235265, 108]
+    )  # This is a tokenization of '<image>' * 1024 + '<bos>Describe the image.\n' line which is used as placeholder
+    # while processing an image
+    EVEN_ATTENTION_MASK = np.array([1] * 1030)
+
+    def __init__(
+        self,
+        model_name: str,
+        cache_dir: str | None = None,
+        threads: int | None = None,
+        providers: Sequence[OnnxProvider] | None = None,
+        cuda: bool | Device = Device.AUTO,
+        device_ids: list[int] | None = None,
+        lazy_load: bool = False,
+        device_id: int | None = None,
+        specific_model_path: str | None = None,
+        **kwargs: Any,
+    ):
+        """
+        Args:
+            model_name (str): The name of the model to use.
+            cache_dir (str, optional): The path to the cache directory.
+                                       Can be set using the `FASTEMBED_CACHE_PATH` env variable.
+                                       Defaults to `fastembed_cache` in the system's temp directory.
+            threads (int, optional): The number of threads single onnxruntime session can use. Defaults to None.
+            providers (Optional[Sequence[OnnxProvider]], optional): The list of onnxruntime providers to use.
+                Mutually exclusive with the `cuda` and `device_ids` arguments. Defaults to None.
+            cuda (Union[bool, Device], optional): Whether to use cuda for inference. Mutually exclusive with `providers`
+                Defaults to Device.AUTO.
+            device_ids (Optional[list[int]], optional): The list of device ids to use for data parallel processing in
+                workers. Should be used with `cuda` equals to `True`, `Device.AUTO` or `Device.CUDA`, mutually exclusive
+                with `providers`. Defaults to None.
+            lazy_load (bool, optional): Whether to load the model during class initialization or on demand.
+                Should be set to True when using multiple-gpu and parallel encoding. Defaults to False.
+            device_id (Optional[int], optional): The device id to use for loading the model in the worker process.
+
+        Raises:
+            ValueError: If the model_name is not in the format <org>/<model> e.g. BAAI/bge-base-en.
+        """
+
+        super().__init__(model_name, cache_dir, threads, **kwargs)
+        self.providers = providers
+        self.lazy_load = lazy_load
+        self._extra_session_options = self._select_exposed_session_options(kwargs)
+
+        # List of device ids, that can be used for data parallel processing in workers
+        self.device_ids = device_ids
+        self.cuda = cuda
+
+        # This device_id will be used if we need to load model in current process
+        self.device_id: int | None = None
+        if device_id is not None:
+            self.device_id = device_id
+        elif self.device_ids is not None:
+            self.device_id = self.device_ids[0]
+
+        self.model_description = self._get_model_description(model_name)
+        self.cache_dir = str(define_cache_dir(cache_dir))
+
+        self._specific_model_path = specific_model_path
+        self._model_dir = self.download_model(
+            self.model_description,
+            self.cache_dir,
+            local_files_only=self._local_files_only,
+            specific_model_path=self._specific_model_path,
+        )
+        self.mask_token_id = None
+        self.pad_token_id = None
+
+        if not self.lazy_load:
+            self.load_onnx_model()
+
+    @classmethod
+    def _list_supported_models(cls) -> list[DenseModelDescription]:
+        """Lists the supported models.
+
+        Returns:
+            list[DenseModelDescription]: A list of DenseModelDescription objects containing the model information.
+        """
+        return supported_colpali_models
+
+    def load_onnx_model(self) -> None:
+        self._load_onnx_model(
+            model_dir=self._model_dir,
+            model_file=self.model_description.model_file,
+            threads=self.threads,
+            providers=self.providers,
+            cuda=self.cuda,
+            device_id=self.device_id,
+            extra_session_options=self._extra_session_options,
+        )
+
+    def _post_process_onnx_image_output(
+        self,
+        output: OnnxOutputContext,
+    ) -> Iterable[NumpyArray]:
+        """
+        Post-process the ONNX model output to convert it into a usable format.
+
+        Args:
+            output (OnnxOutputContext): The raw output from the ONNX model.
+
+        Returns:
+            Iterable[NumpyArray]: Post-processed output as NumPy arrays.
+        """
+        assert self.model_description.dim is not None, "Model dim is not defined"
+        return output.model_output.reshape(
+            output.model_output.shape[0], -1, self.model_description.dim
+        )
+
+    def _post_process_onnx_text_output(
+        self,
+        output: OnnxOutputContext,
+    ) -> Iterable[NumpyArray]:
+        """
+        Post-process the ONNX model output to convert it into a usable format.
+
+        Args:
+            output (OnnxOutputContext): The raw output from the ONNX model.
+
+        Returns:
+            Iterable[NumpyArray]: Post-processed output as NumPy arrays.
+        """
+        return output.model_output
+
+    def tokenize(self, documents: list[str], **kwargs: Any) -> list[Encoding]:
+        texts_query: list[str] = []
+        for query in documents:
+            query = self.BOS_TOKEN + self.QUERY_PREFIX + query + self.PAD_TOKEN * 10
+            query += "\n"
+
+            texts_query.append(query)
+        encoded = self.tokenizer.encode_batch(texts_query)  # type: ignore[union-attr]
+        return encoded
+
+    def token_count(
+        self,
+        texts: str | Iterable[str],
+        batch_size: int = 1024,
+        include_extension: bool = False,
+        **kwargs: Any,
+    ) -> int:
+        if not hasattr(self, "model") or self.model is None:
+            self.load_onnx_model()  # loads the tokenizer as well
+        token_num = 0
+        texts = [texts] if isinstance(texts, str) else texts
+        assert self.tokenizer is not None
+        tokenize_func = self.tokenize if include_extension else self.tokenizer.encode_batch
+        for batch in iter_batch(texts, batch_size):
+            token_num += sum([sum(encoding.attention_mask) for encoding in tokenize_func(batch)])
+        return token_num
+
+    def _preprocess_onnx_text_input(
+        self, onnx_input: dict[str, NumpyArray], **kwargs: Any
+    ) -> dict[str, NumpyArray]:
+        onnx_input["input_ids"] = np.array(
+            [
+                self.QUERY_MARKER_TOKEN_ID + input_ids[2:].tolist()  # type: ignore[index]
+                for input_ids in onnx_input["input_ids"]
+            ]
+        )
+        empty_image_placeholder: NumpyArray = np.zeros(
+            self.IMAGE_PLACEHOLDER_SIZE, dtype=np.float32
+        )
+        onnx_input["pixel_values"] = np.array(
+            [empty_image_placeholder for _ in onnx_input["input_ids"]],
+        )
+        return onnx_input
+
+    def _preprocess_onnx_image_input(
+        self, onnx_input: dict[str, np.ndarray], **kwargs: Any
+    ) -> dict[str, NumpyArray]:
+        """
+        Add placeholders for text input when processing image data for ONNX.
+        Args:
+            onnx_input (Dict[str, NumpyArray]): Preprocessed image inputs.
+            **kwargs: Additional arguments.
+        Returns:
+            Dict[str, NumpyArray]: ONNX input with text placeholders.
+        """
+        onnx_input["input_ids"] = np.array(
+            [self.EMPTY_TEXT_PLACEHOLDER for _ in onnx_input["pixel_values"]]
+        )
+        onnx_input["attention_mask"] = np.array(
+            [self.EVEN_ATTENTION_MASK for _ in onnx_input["pixel_values"]]
+        )
+        return onnx_input
+
+    def embed_text(
+        self,
+        documents: str | Iterable[str],
+        batch_size: int = 256,
+        parallel: int | None = None,
+        **kwargs: Any,
+    ) -> Iterable[NumpyArray]:
+        """
+        Encode a list of documents into list of embeddings.
+
+        Args:
+            documents: Iterator of documents or single document to embed
+            batch_size: Batch size for encoding -- higher values will use more memory, but be faster
+            parallel:
+                If > 1, data-parallel encoding will be used, recommended for offline encoding of large datasets.
+                If 0, use all available cores.
+                If None, don't use data-parallel processing, use default onnxruntime threading instead.
+
+        Returns:
+            List of embeddings, one per document
+        """
+        yield from self._embed_documents(
+            model_name=self.model_name,
+            cache_dir=str(self.cache_dir),
+            documents=documents,
+            batch_size=batch_size,
+            parallel=parallel,
+            providers=self.providers,
+            cuda=self.cuda,
+            device_ids=self.device_ids,
+            local_files_only=self._local_files_only,
+            specific_model_path=self._specific_model_path,
+            extra_session_options=self._extra_session_options,
+            **kwargs,
+        )
+
+    def embed_image(
+        self,
+        images: ImageInput | Iterable[ImageInput],
+        batch_size: int = 16,
+        parallel: int | None = None,
+        **kwargs: Any,
+    ) -> Iterable[NumpyArray]:
+        """
+        Encode a list of images into list of embeddings.
+
+        Args:
+            images: Iterator of image paths or single image path to embed
+            batch_size: Batch size for encoding -- higher values will use more memory, but be faster
+            parallel:
+                If > 1, data-parallel encoding will be used, recommended for offline encoding of large datasets.
+                If 0, use all available cores.
+                If None, don't use data-parallel processing, use default onnxruntime threading instead.
+
+        Returns:
+            List of embeddings, one per document
+        """
+        yield from self._embed_images(
+            model_name=self.model_name,
+            cache_dir=str(self.cache_dir),
+            images=images,
+            batch_size=batch_size,
+            parallel=parallel,
+            providers=self.providers,
+            cuda=self.cuda,
+            device_ids=self.device_ids,
+            local_files_only=self._local_files_only,
+            specific_model_path=self._specific_model_path,
+            extra_session_options=self._extra_session_options,
+            **kwargs,
+        )
+
+    @classmethod
+    def _get_text_worker_class(cls) -> Type[TextEmbeddingWorker[NumpyArray]]:
+        return ColPaliTextEmbeddingWorker
+
+    @classmethod
+    def _get_image_worker_class(cls) -> Type[ImageEmbeddingWorker[NumpyArray]]:
+        return ColPaliImageEmbeddingWorker
+
+
+class ColPaliTextEmbeddingWorker(TextEmbeddingWorker[NumpyArray]):
+    def init_embedding(self, model_name: str, cache_dir: str, **kwargs: Any) -> ColPali:
+        return ColPali(
+            model_name=model_name,
+            cache_dir=cache_dir,
+            threads=1,
+            **kwargs,
+        )
+
+
+class ColPaliImageEmbeddingWorker(ImageEmbeddingWorker[NumpyArray]):
+    def init_embedding(self, model_name: str, cache_dir: str, **kwargs: Any) -> ColPali:
+        return ColPali(
+            model_name=model_name,
+            cache_dir=cache_dir,
+            threads=1,
+            **kwargs,
+        )

fastembed/late_interaction_multimodal/late_interaction_multimodal_embedding.py

@@ -0,0 +1,189 @@
+from typing import Any, Iterable, Sequence, Type
+from dataclasses import asdict
+
+from fastembed.common import OnnxProvider, ImageInput
+from fastembed.common.types import NumpyArray, Device
+from fastembed.late_interaction_multimodal.colpali import ColPali
+from fastembed.late_interaction_multimodal.colmodernvbert import ColModernVBERT
+
+from fastembed.late_interaction_multimodal.late_interaction_multimodal_embedding_base import (
+    LateInteractionMultimodalEmbeddingBase,
+)
+from fastembed.common.model_description import DenseModelDescription
+
+
+class LateInteractionMultimodalEmbedding(LateInteractionMultimodalEmbeddingBase):
+    EMBEDDINGS_REGISTRY: list[Type[LateInteractionMultimodalEmbeddingBase]] = [
+        ColPali,
+        ColModernVBERT,
+    ]
+
+    @classmethod
+    def list_supported_models(cls) -> list[dict[str, Any]]:
+        """
+        Lists the supported models.
+
+        Returns:
+            list[dict[str, Any]]: A list of dictionaries containing the model information.
+
+            Example:
+                ```
+                [
+                    {
+                         "model": "Qdrant/colpali-v1.3-fp16",
+                         "dim": 128,
+                         "description": "Text embeddings, Unimodal (text), Aligned to image latent space, ColBERT-compatible, 512 tokens max, 2024.",
+                         "license": "mit",
+                         "size_in_GB": 6.06,
+                         "sources": {
+                            "hf": "Qdrant/colpali-v1.3-fp16",
+                         },
+                         "additional_files": [
+                            "model.onnx_data",
+                         ],
+                        "model_file": "model.onnx",
+                    },
+                ]
+                ```
+        """
+        return [asdict(model) for model in cls._list_supported_models()]
+
+    @classmethod
+    def _list_supported_models(cls) -> list[DenseModelDescription]:
+        result: list[DenseModelDescription] = []
+        for embedding in cls.EMBEDDINGS_REGISTRY:
+            result.extend(embedding._list_supported_models())
+        return result
+
+    def __init__(
+        self,
+        model_name: str,
+        cache_dir: str | None = None,
+        threads: int | None = None,
+        providers: Sequence[OnnxProvider] | None = None,
+        cuda: bool | Device = Device.AUTO,
+        device_ids: list[int] | None = None,
+        lazy_load: bool = False,
+        **kwargs: Any,
+    ):
+        super().__init__(model_name, cache_dir, threads, **kwargs)
+        for EMBEDDING_MODEL_TYPE in self.EMBEDDINGS_REGISTRY:
+            supported_models = EMBEDDING_MODEL_TYPE._list_supported_models()
+            if any(model_name.lower() == model.model.lower() for model in supported_models):
+                self.model = EMBEDDING_MODEL_TYPE(
+                    model_name,
+                    cache_dir,
+                    threads=threads,
+                    providers=providers,
+                    cuda=cuda,
+                    device_ids=device_ids,
+                    lazy_load=lazy_load,
+                    **kwargs,
+                )
+                return
+
+        raise ValueError(
+            f"Model {model_name} is not supported in LateInteractionMultimodalEmbedding."
+            "Please check the supported models using `LateInteractionMultimodalEmbedding.list_supported_models()`"
+        )
+
+    @property
+    def embedding_size(self) -> int:
+        """Get the embedding size of the current model"""
+        if self._embedding_size is None:
+            self._embedding_size = self.get_embedding_size(self.model_name)
+        return self._embedding_size
+
+    @classmethod
+    def get_embedding_size(cls, model_name: str) -> int:
+        """Get the embedding size of the passed model
+
+        Args:
+            model_name (str): The name of the model to get embedding size for.
+
+        Returns:
+            int: The size of the embedding.
+
+        Raises:
+            ValueError: If the model name is not found in the supported models.
+        """
+        descriptions = cls._list_supported_models()
+        embedding_size: int | None = None
+        for description in descriptions:
+            if description.model.lower() == model_name.lower():
+                embedding_size = description.dim
+                break
+        if embedding_size is None:
+            model_names = [description.model for description in descriptions]
+            raise ValueError(
+                f"Embedding size for model {model_name} was None. "
+                f"Available model names: {model_names}"
+            )
+        return embedding_size
+
+    def embed_text(
+        self,
+        documents: str | Iterable[str],
+        batch_size: int = 256,
+        parallel: int | None = None,
+        **kwargs: Any,
+    ) -> Iterable[NumpyArray]:
+        """
+        Encode a list of documents into list of embeddings.
+
+        Args:
+            documents: Iterator of documents or single document to embed
+            batch_size: Batch size for encoding -- higher values will use more memory, but be faster
+            parallel:
+                If > 1, data-parallel encoding will be used, recommended for offline encoding of large datasets.
+                If 0, use all available cores.
+                If None, don't use data-parallel processing, use default onnxruntime threading instead.
+
+        Returns:
+            List of embeddings, one per document
+        """
+        yield from self.model.embed_text(documents, batch_size, parallel, **kwargs)
+
+    def embed_image(
+        self,
+        images: ImageInput | Iterable[ImageInput],
+        batch_size: int = 16,
+        parallel: int | None = None,
+        **kwargs: Any,
+    ) -> Iterable[NumpyArray]:
+        """
+        Encode a list of images into list of embeddings.
+
+        Args:
+            images: Iterator of image paths or single image path to embed
+            batch_size: Batch size for encoding -- higher values will use more memory, but be faster
+            parallel:
+                If > 1, data-parallel encoding will be used, recommended for offline encoding of large datasets.
+                If 0, use all available cores.
+                If None, don't use data-parallel processing, use default onnxruntime threading instead.
+
+        Returns:
+            List of embeddings, one per image
+        """
+        yield from self.model.embed_image(images, batch_size, parallel, **kwargs)
+
+    def token_count(
+        self,
+        texts: str | Iterable[str],
+        batch_size: int = 1024,
+        include_extension: bool = False,
+        **kwargs: Any,
+    ) -> int:
+        """Returns the number of tokens in the texts.
+
+        Args:
+            texts (str | Iterable[str]): The list of texts to embed.
+            batch_size (int): Batch size for encoding
+            include_extension (bool): Whether to include tokens added by preprocessing
+
+        Returns:
+            int: Sum of number of tokens in the texts.
+        """
+        return self.model.token_count(
+            texts, batch_size=batch_size, include_extension=include_extension, **kwargs
+        )

fastembed/late_interaction_multimodal/late_interaction_multimodal_embedding_base.py

@@ -0,0 +1,86 @@
+from typing import Iterable, Any
+
+
+from fastembed.common import ImageInput
+from fastembed.common.model_description import DenseModelDescription
+from fastembed.common.model_management import ModelManagement
+from fastembed.common.types import NumpyArray
+
+
+class LateInteractionMultimodalEmbeddingBase(ModelManagement[DenseModelDescription]):
+    def __init__(
+        self,
+        model_name: str,
+        cache_dir: str | None = None,
+        threads: int | None = None,
+        **kwargs: Any,
+    ):
+        self.model_name = model_name
+        self.cache_dir = cache_dir
+        self.threads = threads
+        self._local_files_only = kwargs.pop("local_files_only", False)
+        self._embedding_size: int | None = None
+
+    def embed_text(
+        self,
+        documents: str | Iterable[str],
+        batch_size: int = 256,
+        parallel: int | None = None,
+        **kwargs: Any,
+    ) -> Iterable[NumpyArray]:
+        """
+        Embeds a list of documents into a list of embeddings.
+
+        Args:
+            documents (Iterable[str]): The list of texts to embed.
+            batch_size: Batch size for encoding -- higher values will use more memory, but be faster
+            parallel:
+                If > 1, data-parallel encoding will be used, recommended for offline encoding of large datasets.
+                If 0, use all available cores.
+                If None, don't use data-parallel processing, use default onnxruntime threading instead.
+            **kwargs: Additional keyword argument to pass to the embed method.
+
+        Yields:
+            Iterable[NumpyArray]: The embeddings.
+        """
+        raise NotImplementedError()
+
+    def embed_image(
+        self,
+        images: ImageInput | Iterable[ImageInput],
+        batch_size: int = 16,
+        parallel: int | None = None,
+        **kwargs: Any,
+    ) -> Iterable[NumpyArray]:
+        """
+        Encode a list of images into list of embeddings.
+        Args:
+            images: Iterator of image paths or single image path to embed
+            batch_size: Batch size for encoding -- higher values will use more memory, but be faster
+            parallel:
+                If > 1, data-parallel encoding will be used, recommended for offline encoding of large datasets.
+                If 0, use all available cores.
+                If None, don't use data-parallel processing, use default onnxruntime threading instead.
+
+        Returns:
+            List of embeddings, one per image
+        """
+        raise NotImplementedError()
+
+    @classmethod
+    def get_embedding_size(cls, model_name: str) -> int:
+        """Returns embedding size of the chosen model."""
+        raise NotImplementedError("Subclasses must implement this method")
+
+    @property
+    def embedding_size(self) -> int:
+        """Returns embedding size for the current model"""
+        raise NotImplementedError("Subclasses must implement this method")
+
+    def token_count(
+        self,
+        texts: str | Iterable[str],
+        **kwargs: Any,
+    ) -> int:
+        """Returns the number of tokens in the texts."""
+        raise NotImplementedError("Subclasses must implement this method")