optimized scrfd code

machine-learning/app/models/facial_recognition/detection.py

@@ -1,40 +1,33 @@
-from pathlib import Path
 from typing import Any
 
 import numpy as np
-from insightface.model_zoo import RetinaFace
+import onnxruntime as ort
 from numpy.typing import NDArray
 
 from app.models.base import InferenceModel
-from app.models.transforms import decode_cv2
+from app.models.session import ort_has_batch_dim, ort_expand_outputs
+from app.models.transforms import decode_pil
 from app.schemas import FaceDetectionOutput, ModelSession, ModelTask, ModelType
-
+from .scrfd import SCRFD
+from PIL import Image
+from PIL.ImageOps import pad
 
 class FaceDetector(InferenceModel):
     depends = []
     identity = (ModelType.DETECTION, ModelTask.FACIAL_RECOGNITION)
 
-    def __init__(
-        self,
-        model_name: str,
-        min_score: float = 0.7,
-        cache_dir: Path | str | None = None,
-        **model_kwargs: Any,
-    ) -> None:
-        self.min_score = model_kwargs.pop("minScore", min_score)
-        super().__init__(model_name, cache_dir, **model_kwargs)
-
     def _load(self) -> ModelSession:
         session = self._make_session(self.model_path)
-        self.model = RetinaFace(session=session)
+        if isinstance(session, ort.InferenceSession) and not ort_has_batch_dim(session):
-        self.model.prepare(ctx_id=0, det_thresh=self.min_score, input_size=(640, 640))
+            ort_expand_outputs(session)
+        self.model = SCRFD(session=session)
 
         return session
 
-    def _predict(self, inputs: NDArray[np.uint8] | bytes, **kwargs: Any) -> FaceDetectionOutput:
+    def _predict(self, inputs: NDArray[np.uint8] | bytes | Image.Image, **kwargs: Any) -> FaceDetectionOutput:
-        inputs = decode_cv2(inputs)
+        inputs = self._transform(inputs)
 
-        bboxes, landmarks = self._detect(inputs)
+        [bboxes], [landmarks] = self.model.detect(inputs, threshold=kwargs.pop("minScore", 0.7))
         return {
             "boxes": bboxes[:, :4].round(),
             "scores": bboxes[:, 4],
@@ -44,5 +37,7 @@ class FaceDetector(InferenceModel):
     def _detect(self, inputs: NDArray[np.uint8] | bytes) -> tuple[NDArray[np.float32], NDArray[np.float32]]:
         return self.model.detect(inputs)  # type: ignore
 
-    def configure(self, **kwargs: Any) -> None:
+    def _transform(self, inputs: NDArray[np.uint8] | bytes | Image.Image) -> NDArray[np.uint8]:
-        self.model.det_thresh = kwargs.pop("minScore", self.model.det_thresh)
+        image = decode_pil(inputs)
+        padded = pad(image, (640, 640), method=Image.Resampling.BICUBIC)
+        return np.array(padded, dtype=np.uint8)[None, ...]

machine-learning/app/models/facial_recognition/recognition.py

@@ -2,16 +2,15 @@ from pathlib import Path
 from typing import Any
 
 import numpy as np
-import onnx
 import onnxruntime as ort
 from insightface.model_zoo import ArcFaceONNX
 from insightface.utils.face_align import norm_crop
 from numpy.typing import NDArray
-from onnx.tools.update_model_dims import update_inputs_outputs_dims
 from PIL import Image
 
 from app.config import clean_name, log
 from app.models.base import InferenceModel
+from app.models.session import ort_add_batch_dim, ort_has_batch_dim
 from app.models.transforms import decode_cv2
 from app.schemas import FaceDetectionOutput, FacialRecognitionOutput, ModelSession, ModelTask, ModelType
 
@@ -32,8 +31,9 @@ class FaceRecognizer(InferenceModel):
 
     def _load(self) -> ModelSession:
         session = self._make_session(self.model_path)
-        if not self._has_batch_dim(session):
+        if isinstance(session, ort.InferenceSession) and not ort_has_batch_dim(session):
-            self._add_batch_dim(self.model_path)
+            log.info(f"Adding batch dimension to recognition model {self.model_name}")
+            ort_add_batch_dim(self.model_path, self.model_path)
             session = self._make_session(self.model_path)
         self.model = ArcFaceONNX(
             self.model_path.with_suffix(".onnx").as_posix(),
@@ -62,16 +62,3 @@ class FaceRecognizer(InferenceModel):
 
     def _crop(self, image: NDArray[np.uint8], faces: FaceDetectionOutput) -> list[NDArray[np.uint8]]:
         return [norm_crop(image, landmark) for landmark in faces["landmarks"]]
-
-    def _has_batch_dim(self, session: ort.InferenceSession) -> bool:
-        return not isinstance(session, ort.InferenceSession) or session.get_inputs()[0].shape[0] == "batch"
-
-    def _add_batch_dim(self, model_path: Path) -> None:
-        log.debug(f"Adding batch dimension to model {model_path}")
-        proto = onnx.load(model_path)
-        static_input_dims = [shape.dim_value for shape in proto.graph.input[0].type.tensor_type.shape.dim[1:]]
-        static_output_dims = [shape.dim_value for shape in proto.graph.output[0].type.tensor_type.shape.dim[1:]]
-        input_dims = {proto.graph.input[0].name: ["batch"] + static_input_dims}
-        output_dims = {proto.graph.output[0].name: ["batch"] + static_output_dims}
-        updated_proto = update_inputs_outputs_dims(proto, input_dims, output_dims)
-        onnx.save(updated_proto, model_path)

machine-learning/app/models/facial_recognition/scrfd.py

@@ -0,0 +1,325 @@
+# Based on InsightFace-REST by SthPhoenix https://github.com/SthPhoenix/InsightFace-REST/blob/master/src/api_trt/modules/model_zoo/detectors/scrfd.py
+# Primary changes made:
+# 1. Removed CuPy-related code
+# 2. Adapted proposal generation to be thread-safe
+# 3. Added typing
+# 4. Assume RGB input
+# 5. Removed unused variables
+
+# Copyright 2021 SthPhoenix
+
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+
+# http://www.apache.org/licenses/LICENSE-2.0
+
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# -*- coding: utf-8 -*-
+# Based on Jia Guo reference implementation at
+# https://github.com/deepinsight/insightface/blob/master/detection/scrfd/tools/scrfd.py
+
+
+from __future__ import division
+
+import cv2
+import numpy as np
+from numba import njit
+from app.schemas import ModelSession
+from numpy.typing import NDArray
+
+
+@njit(cache=True, nogil=True)
+def nms(dets, threshold: float = 0.4) -> NDArray[np.float32]:
+    x1 = dets[:, 0]
+    y1 = dets[:, 1]
+    x2 = dets[:, 2]
+    y2 = dets[:, 3]
+    scores = dets[:, 4]
+
+    areas = (x2 - x1 + 1) * (y2 - y1 + 1)
+    order = scores.argsort()[::-1]
+
+    keep = []
+    while order.size > 0:
+        i = order[0]
+        keep.append(i)
+        xx1 = np.maximum(x1[i], x1[order[1:]])
+        yy1 = np.maximum(y1[i], y1[order[1:]])
+        xx2 = np.minimum(x2[i], x2[order[1:]])
+        yy2 = np.minimum(y2[i], y2[order[1:]])
+
+        w = np.maximum(0.0, xx2 - xx1 + 1)
+        h = np.maximum(0.0, yy2 - yy1 + 1)
+        inter = w * h
+        ovr = inter / (areas[i] + areas[order[1:]] - inter)
+
+        inds = np.where(ovr <= threshold)[0]
+        order = order[inds + 1]
+
+    return np.asarray(keep)
+
+
+@njit(fastmath=True, cache=True, nogil=True)
+def single_distance2bbox(point: NDArray[np.float32], distance: NDArray[np.float32], stride: int) -> NDArray[np.float32]:
+    """
+    Fast conversion of single bbox distances to coordinates
+
+    :param point: Anchor point
+    :param distance: Bbox distances from anchor point
+    :param stride: Current stride scale
+    :return: bbox
+    """
+    distance[0] = point[0] - distance[0] * stride
+    distance[1] = point[1] - distance[1] * stride
+    distance[2] = point[0] + distance[2] * stride
+    distance[3] = point[1] + distance[3] * stride
+    return distance
+
+
+@njit(fastmath=True, cache=True, nogil=True)
+def single_distance2kps(point: NDArray[np.float32], distance: NDArray[np.float32], stride: int) -> NDArray[np.float32]:
+    """
+    Fast conversion of single keypoint distances to coordinates
+
+    :param point: Anchor point
+    :param distance: Keypoint distances from anchor point
+    :param stride: Current stride scale
+    :return: keypoint
+    """
+    for ix in range(0, distance.shape[0], 2):
+        distance[ix] = distance[ix] * stride + point[0]
+        distance[ix + 1] = distance[ix + 1] * stride + point[1]
+    return distance
+
+
+@njit(fastmath=True, cache=True, nogil=True)
+def generate_proposals(
+    score_blob: NDArray[np.float32],
+    bbox_blob: NDArray[np.float32],
+    kpss_blob: NDArray[np.float32],
+    stride: int,
+    anchors: NDArray[np.float32],
+    threshold: float,
+) -> tuple[NDArray[np.float32], NDArray[np.float32], NDArray[np.float32]]:
+    """
+    Convert distances from anchors to actual coordinates on source image
+    and filter proposals by confidence threshold.
+
+    :param score_blob: Raw scores for stride
+    :param bbox_blob: Raw bbox distances for stride
+    :param kpss_blob: Raw keypoints distances for stride
+    :param stride: Stride scale
+    :param anchors: Precomputed anchors for stride
+    :param threshold: Confidence threshold
+    :return: Filtered scores, bboxes and keypoints
+    """
+
+    idxs = []
+    for ix in range(score_blob.shape[0]):
+        if score_blob[ix][0] > threshold:
+            idxs.append(ix)
+
+    score_out = np.empty((len(idxs), 1), dtype="float32")
+    bbox_out = np.empty((len(idxs), 4), dtype="float32")
+    kpss_out = np.empty((len(idxs), 10), dtype="float32")
+
+    for i in range(len(idxs)):
+        ix = idxs[i]
+        score_out[i] = score_blob[ix]
+        bbox_out[i] = single_distance2bbox(anchors[ix], bbox_blob[ix], stride)
+        kpss_out[i] = single_distance2kps(anchors[ix], kpss_blob[ix], stride)
+
+    return score_out, bbox_out, kpss_out
+
+
+@njit(fastmath=True, cache=True, nogil=True)
+def filter(
+    bboxes_list: NDArray[np.float32],
+    kpss_list: NDArray[np.float32],
+    scores_list: NDArray[np.float32],
+    nms_threshold: float = 0.4,
+) -> tuple[NDArray[np.float32], NDArray[np.float32]]:
+    """
+    Filter postprocessed network outputs with NMS
+
+    :param bboxes_list: List of bboxes (np.ndarray)
+    :param kpss_list: List of keypoints (np.ndarray)
+    :param scores_list: List of scores (np.ndarray)
+    :return: Face bboxes with scores [t,l,b,r,score], and key points
+    """
+
+    pre_det = np.hstack((bboxes_list, scores_list))
+    keep = nms(pre_det, threshold=nms_threshold)
+    det = pre_det[keep, :]
+    kpss = kpss_list[keep, :]
+    kpss = kpss.reshape((kpss.shape[0], -1, 2))
+
+    return det, kpss
+
+
+class SCRFD:
+    def __init__(self, session: ModelSession):
+        self.session = session
+        self.center_cache: dict[tuple[int, int], NDArray[np.float32]] = {}
+        self.nms_threshold = 0.4
+        self.fmc = 3
+        self._feat_stride_fpn = [8, 16, 32]
+        self._num_anchors = 2
+
+    def prepare(self, nms_threshold: float = 0.4) -> None:
+        """
+        Populate class parameters
+
+        :param nms_threshold: Threshold for NMS IoU
+        """
+
+        self.nms_threshold = nms_threshold
+
+    def detect(
+        self, imgs: NDArray[np.uint8], threshold: float = 0.5
+    ) -> tuple[list[NDArray[np.float32]], list[NDArray[np.float32]]]:
+        """
+        Run detection pipeline for provided images
+
+        :param img: Raw image as nd.ndarray with HWC shape
+        :param threshold: Confidence threshold
+        :return: Face bboxes with scores [t,l,b,r,score], and key points
+        """
+
+        height, width = imgs.shape[1:3]
+        blob = self._preprocess(imgs)
+        net_outs = self._forward(blob)
+
+        batch_bboxes, batch_kpss, batch_scores = self._postprocess(net_outs, height, width, threshold)
+
+        dets_list = []
+        kpss_list = []
+        for e in range(imgs.shape[0]):
+            if len(batch_bboxes[e]) == 0:
+                det, kpss = np.zeros((0, 5), dtype="float32"), np.zeros((0, 10), dtype="float32")
+            else:
+                det, kpss = filter(batch_bboxes[e], batch_kpss[e], batch_scores[e], self.nms_threshold)
+
+            dets_list.append(det)
+            kpss_list.append(kpss)
+
+        return dets_list, kpss_list
+
+    @staticmethod
+    def _build_anchors(
+        input_height: int, input_width: int, strides: list[int], num_anchors: int
+    ) -> NDArray[np.float32]:
+        """
+        Precompute anchor points for provided image size
+
+        :param input_height: Input image height
+        :param input_width: Input image width
+        :param strides: Model strides
+        :param num_anchors: Model num anchors
+        :return: box centers
+        """
+
+        centers = []
+        for stride in strides:
+            height = input_height // stride
+            width = input_width // stride
+
+            anchor_centers = np.stack(np.mgrid[:height, :width][::-1], axis=-1).astype(np.float32)
+            anchor_centers = (anchor_centers * stride).reshape((-1, 2))
+            if num_anchors > 1:
+                anchor_centers = np.stack([anchor_centers] * num_anchors, axis=1).reshape((-1, 2))
+            centers.append(anchor_centers)
+        return centers
+
+    def _preprocess(self, images: NDArray[np.uint8]):
+        """
+        Normalize image on CPU if backend can't provide CUDA stream,
+        otherwise preprocess image on GPU using CuPy
+
+        :param img: Raw image as np.ndarray with HWC shape
+        :return: Preprocessed image or None if image was processed on device
+        """
+
+        input_size = tuple(images[0].shape[0:2][::-1])
+        return cv2.dnn.blobFromImages(images, 1.0 / 128, input_size, (127.5, 127.5, 127.5), swapRB=False)
+
+    def _forward(self, blob: NDArray[np.float32]) -> list[NDArray[np.float32]]:
+        """
+        Send input data to inference backend.
+
+        :param blob: Preprocessed image of shape NCHW or None
+        :return: network outputs
+        """
+
+        return self.session.run(None, {"input.1": blob})
+
+    def _postprocess(
+        self, net_outs: list[NDArray[np.float32]], height: int, width: int, threshold: float
+    ) -> tuple[list[NDArray[np.float32]], list[NDArray[np.float32]], list[NDArray[np.float32]]]:
+        """
+        Precompute anchor points for provided image size and process network outputs
+
+        :param net_outs: Network outputs
+        :param input_height: Input image height
+        :param input_width: Input image width
+        :param threshold: Confidence threshold
+        :return: filtered bboxes, keypoints and scores
+        """
+
+        key = (height, width)
+
+        if not self.center_cache.get(key):
+            self.center_cache[key] = self._build_anchors(height, width, self._feat_stride_fpn, self._num_anchors)
+        anchor_centers = self.center_cache[key]
+        bboxes, kpss, scores = self._process_strides(net_outs, threshold, anchor_centers)
+        return bboxes, kpss, scores
+
+    def _process_strides(
+        self, net_outs: list[NDArray[np.float32]], threshold: float, anchors: NDArray[np.float32]
+    ) -> tuple[list[NDArray[np.float32]], list[NDArray[np.float32]], list[NDArray[np.float32]]]:
+        """
+        Process network outputs by strides and return results proposals filtered by threshold
+
+        :param net_outs: Network outputs
+        :param threshold: Confidence threshold
+        :param anchor_centers: Precomputed anchor centers for all strides
+        :return: filtered bboxes, keypoints and scores
+        """
+
+        batch_size = net_outs[0].shape[0]
+        bboxes_by_img = []
+        kpss_by_img = []
+        scores_by_img = []
+
+        for batch in range(batch_size):
+            scores_strided = []
+            bboxes_strided = []
+            kpss_strided = []
+            for idx, stride in enumerate(self._feat_stride_fpn):
+                score_blob = net_outs[idx][batch]
+                bbox_blob = net_outs[idx + self.fmc][batch]
+                kpss_blob = net_outs[idx + self.fmc * 2][batch]
+                stride_anchors = anchors[idx]
+                score_list, bbox_list, kpss_list = generate_proposals(
+                    score_blob,
+                    bbox_blob,
+                    kpss_blob,
+                    stride,
+                    stride_anchors,
+                    threshold,
+                )
+
+                scores_strided.append(score_list)
+                bboxes_strided.append(bbox_list)
+                kpss_strided.append(kpss_list)
+            bboxes_by_img.append(np.concatenate(bboxes_strided, axis=0))
+            kpss_by_img.append(np.concatenate(kpss_strided, axis=0))
+            scores_by_img.append(np.concatenate(scores_strided, axis=0))
+
+        return bboxes_by_img, kpss_by_img, scores_by_img

machine-learning/app/models/session.py

@@ -0,0 +1,34 @@
+from pathlib import Path
+
+import numpy as np
+import onnx
+import onnxruntime as ort
+from numpy.typing import NDArray
+from onnx.shape_inference import infer_shapes
+from onnx.tools.update_model_dims import update_inputs_outputs_dims
+
+
+def ort_has_batch_dim(session: ort.InferenceSession) -> bool:
+    return session.get_inputs()[0].shape[0] == "batch"
+
+
+def ort_expand_outputs(session: ort.InferenceSession) -> None:
+    original_run = session.run
+
+    def run(output_names: list[str], input_feed: dict[str, NDArray[np.float32]]) -> list[NDArray[np.float32]]:
+        out: list[NDArray[np.float32]] = original_run(output_names, input_feed)
+        out = [np.expand_dims(o, axis=0) for o in out]
+        return out
+
+    session.run = run
+
+
+def ort_add_batch_dim(input_path: Path, output_path: Path) -> None:
+    proto = onnx.load(input_path)
+    static_input_dims = [shape.dim_value for shape in proto.graph.input[0].type.tensor_type.shape.dim[1:]]
+    static_output_dims = [shape.dim_value for shape in proto.graph.output[0].type.tensor_type.shape.dim[1:]]
+    input_dims = {proto.graph.input[0].name: ["batch"] + static_input_dims}
+    output_dims = {proto.graph.output[0].name: ["batch"] + static_output_dims}
+    updated_proto = update_inputs_outputs_dims(proto, input_dims, output_dims)
+    inferred = infer_shapes(updated_proto)
+    onnx.save(inferred, output_path)

machine-learning/app/models/transforms.py

@@ -3,6 +3,7 @@ from typing import IO
 
 import cv2
 import numpy as np
+from numba import njit
 from numpy.typing import NDArray
 from PIL import Image
 
@@ -30,10 +31,11 @@ def to_numpy(img: Image.Image) -> NDArray[np.float32]:
     return np.asarray(img if img.mode == "RGB" else img.convert("RGB"), dtype=np.float32) / 255.0
 
 
+@njit(cache=True, fastmath=True, nogil=True)
 def normalize(
     img: NDArray[np.float32], mean: float | NDArray[np.float32], std: float | NDArray[np.float32]
 ) -> NDArray[np.float32]:
-    return np.divide(img - mean, std, dtype=np.float32)
+    return (img - mean) / std
 
 
 def get_pil_resampling(resample: str) -> Image.Resampling:

machine-learning/app/test_main.py

@@ -7,6 +7,7 @@ from types import SimpleNamespace
 from typing import Any, Callable
 from unittest import mock
 
+from app.models.session import ort_add_batch_dim, ort_has_batch_dim, ort_squeeze_outputs
 import cv2
 import numpy as np
 import onnxruntime as ort
@@ -734,3 +735,69 @@ class TestEndpoints:
             assert expected_face["boundingBox"] == actual_face["boundingBox"]
             assert np.allclose(expected_face["embedding"], actual_face["embedding"])
             assert np.allclose(expected_face["score"], actual_face["score"])
+
+
+class TestSessionUtils:
+    def test_ort_has_batch_dim(self, mocker: MockerFixture) -> None:
+        mock_session = mocker.Mock(spec=ort.InferenceSession)
+        mock_session.get_inputs.return_value = [SimpleNamespace(shape=["batch", 3, 224, 224], name="input.1")]
+
+        assert ort_has_batch_dim(mock_session) is True
+
+    def test_ort_has_no_batch_dim(self, mocker: MockerFixture) -> None:
+        mock_session = mocker.Mock(spec=ort.InferenceSession)
+        mock_session.get_inputs.return_value = [SimpleNamespace(shape=[1, 3, 224, 224], name="input.1")]
+
+        assert ort_has_batch_dim(mock_session) is False
+
+    def test_ort_squeeze_outputs(self, mocker: MockerFixture) -> None:
+        mock_session = mocker.Mock(spec=ort.InferenceSession)
+        mock_session.run.return_value = [np.random.rand(1, 3, 224, 224).astype(np.float32)]
+
+        ort_squeeze_outputs(mock_session)
+        out = mock_session.run(["output"], {"input": np.random.rand(1, 3, 224, 224).astype(np.float32)})
+
+        assert len(out) == 1
+        assert out[0].shape == (3, 224, 224)
+
+    def test_ort_add_batch_dim(self, mocker: MockerFixture) -> None:
+        mock_proto = mocker.Mock()
+        mock_proto.graph.input = [
+            SimpleNamespace(name="input", type=mock.Mock(tensor_type=SimpleNamespace(shape=SimpleNamespace())))
+        ]
+        mock_proto.graph.output = [
+            SimpleNamespace(name="output", type=SimpleNamespace(tensor_type=SimpleNamespace(shape=SimpleNamespace())))
+        ]
+        mock_proto.graph.input[0].type.tensor_type.shape.dim = [
+            SimpleNamespace(dim_value=3),
+            SimpleNamespace(dim_value=224),
+            SimpleNamespace(dim_value=224),
+        ]
+        mock_proto.graph.output[0].type.tensor_type.shape.dim = [
+            SimpleNamespace(dim_value=3),
+            SimpleNamespace(dim_value=224),
+            SimpleNamespace(dim_value=224),
+        ]
+
+        mock_load = mocker.patch("app.models.session.onnx.load")
+        mock_load.return_value = mock_proto
+
+        mock_update_dims = mocker.patch("app.models.session.update_inputs_outputs_dims")
+        mock_updated = mocker.Mock()
+        mock_update_dims.return_value = mock_updated
+
+        mock_infer_shapes = mocker.patch("app.models.session.infer_shapes")
+        mock_inferred = mocker.Mock()
+        mock_infer_shapes.return_value = mock_inferred
+
+        mock_save = mocker.patch("app.models.session.onnx.save")
+
+        input_path, output_path = Path("input.onnx"), Path("output.onnx")
+        ort_add_batch_dim(input_path, output_path)
+
+        mock_load.assert_called_once_with(input_path)
+        mock_save.assert_called_once_with(mock_inferred, output_path)
+        mock_update_dims.assert_called_once_with(mock_proto, mock.ANY, mock.ANY)
+        assert mock_update_dims.call_args_list == [
+            mock.call(mock_proto, {"input": ["batch", 224, 224]}, {"output": ["batch", 224, 224]})
+        ]

machine-learning/poetry.lock

@@ -1528,6 +1528,36 @@ files = [
 lint = ["pre-commit (>=3.3)"]
 test = ["pytest (>=7.4)", "pytest-cov (>=4.1)"]
 
+[[package]]
+name = "llvmlite"
+version = "0.42.0"
+description = "lightweight wrapper around basic LLVM functionality"
+optional = false
+python-versions = ">=3.9"
+files = [
+    {file = "llvmlite-0.42.0-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:3366938e1bf63d26c34fbfb4c8e8d2ded57d11e0567d5bb243d89aab1eb56098"},
+    {file = "llvmlite-0.42.0-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:c35da49666a21185d21b551fc3caf46a935d54d66969d32d72af109b5e7d2b6f"},
+    {file = "llvmlite-0.42.0-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:70f44ccc3c6220bd23e0ba698a63ec2a7d3205da0d848804807f37fc243e3f77"},
+    {file = "llvmlite-0.42.0-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:763f8d8717a9073b9e0246998de89929071d15b47f254c10eef2310b9aac033d"},
+    {file = "llvmlite-0.42.0-cp310-cp310-win_amd64.whl", hash = "sha256:8d90edf400b4ceb3a0e776b6c6e4656d05c7187c439587e06f86afceb66d2be5"},
+    {file = "llvmlite-0.42.0-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:ae511caed28beaf1252dbaf5f40e663f533b79ceb408c874c01754cafabb9cbf"},
+    {file = "llvmlite-0.42.0-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:81e674c2fe85576e6c4474e8c7e7aba7901ac0196e864fe7985492b737dbab65"},
+    {file = "llvmlite-0.42.0-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:bb3975787f13eb97629052edb5017f6c170eebc1c14a0433e8089e5db43bcce6"},
+    {file = "llvmlite-0.42.0-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:c5bece0cdf77f22379f19b1959ccd7aee518afa4afbd3656c6365865f84903f9"},
+    {file = "llvmlite-0.42.0-cp311-cp311-win_amd64.whl", hash = "sha256:7e0c4c11c8c2aa9b0701f91b799cb9134a6a6de51444eff5a9087fc7c1384275"},
+    {file = "llvmlite-0.42.0-cp312-cp312-macosx_10_9_x86_64.whl", hash = "sha256:08fa9ab02b0d0179c688a4216b8939138266519aaa0aa94f1195a8542faedb56"},
+    {file = "llvmlite-0.42.0-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:b2fce7d355068494d1e42202c7aff25d50c462584233013eb4470c33b995e3ee"},
+    {file = "llvmlite-0.42.0-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:ebe66a86dc44634b59a3bc860c7b20d26d9aaffcd30364ebe8ba79161a9121f4"},
+    {file = "llvmlite-0.42.0-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:d47494552559e00d81bfb836cf1c4d5a5062e54102cc5767d5aa1e77ccd2505c"},
+    {file = "llvmlite-0.42.0-cp312-cp312-win_amd64.whl", hash = "sha256:05cb7e9b6ce69165ce4d1b994fbdedca0c62492e537b0cc86141b6e2c78d5888"},
+    {file = "llvmlite-0.42.0-cp39-cp39-macosx_10_9_x86_64.whl", hash = "sha256:bdd3888544538a94d7ec99e7c62a0cdd8833609c85f0c23fcb6c5c591aec60ad"},
+    {file = "llvmlite-0.42.0-cp39-cp39-macosx_11_0_arm64.whl", hash = "sha256:d0936c2067a67fb8816c908d5457d63eba3e2b17e515c5fe00e5ee2bace06040"},
+    {file = "llvmlite-0.42.0-cp39-cp39-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:a78ab89f1924fc11482209f6799a7a3fc74ddc80425a7a3e0e8174af0e9e2301"},
+    {file = "llvmlite-0.42.0-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:d7599b65c7af7abbc978dbf345712c60fd596aa5670496561cc10e8a71cebfb2"},
+    {file = "llvmlite-0.42.0-cp39-cp39-win_amd64.whl", hash = "sha256:43d65cc4e206c2e902c1004dd5418417c4efa6c1d04df05c6c5675a27e8ca90e"},
+    {file = "llvmlite-0.42.0.tar.gz", hash = "sha256:f92b09243c0cc3f457da8b983f67bd8e1295d0f5b3746c7a1861d7a99403854a"},
+]
+
 [[package]]
 name = "locust"
 version = "2.28.0"
@@ -1864,6 +1894,40 @@ doc = ["nb2plots (>=0.7)", "nbconvert (<7.9)", "numpydoc (>=1.6)", "pillow (>=9.
 extra = ["lxml (>=4.6)", "pydot (>=1.4.2)", "pygraphviz (>=1.11)", "sympy (>=1.10)"]
 test = ["pytest (>=7.2)", "pytest-cov (>=4.0)"]
 
+[[package]]
+name = "numba"
+version = "0.59.1"
+description = "compiling Python code using LLVM"
+optional = false
+python-versions = ">=3.9"
+files = [
+    {file = "numba-0.59.1-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:97385a7f12212c4f4bc28f648720a92514bee79d7063e40ef66c2d30600fd18e"},
+    {file = "numba-0.59.1-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:0b77aecf52040de2a1eb1d7e314497b9e56fba17466c80b457b971a25bb1576d"},
+    {file = "numba-0.59.1-cp310-cp310-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:3476a4f641bfd58f35ead42f4dcaf5f132569c4647c6f1360ccf18ee4cda3990"},
+    {file = "numba-0.59.1-cp310-cp310-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:525ef3f820931bdae95ee5379c670d5c97289c6520726bc6937a4a7d4230ba24"},
+    {file = "numba-0.59.1-cp310-cp310-win_amd64.whl", hash = "sha256:990e395e44d192a12105eca3083b61307db7da10e093972ca285c85bef0963d6"},
+    {file = "numba-0.59.1-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:43727e7ad20b3ec23ee4fc642f5b61845c71f75dd2825b3c234390c6d8d64051"},
+    {file = "numba-0.59.1-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:411df625372c77959570050e861981e9d196cc1da9aa62c3d6a836b5cc338966"},
+    {file = "numba-0.59.1-cp311-cp311-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:2801003caa263d1e8497fb84829a7ecfb61738a95f62bc05693fcf1733e978e4"},
+    {file = "numba-0.59.1-cp311-cp311-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:dd2842fac03be4e5324ebbbd4d2d0c8c0fc6e0df75c09477dd45b288a0777389"},
+    {file = "numba-0.59.1-cp311-cp311-win_amd64.whl", hash = "sha256:0594b3dfb369fada1f8bb2e3045cd6c61a564c62e50cf1f86b4666bc721b3450"},
+    {file = "numba-0.59.1-cp312-cp312-macosx_10_9_x86_64.whl", hash = "sha256:1cce206a3b92836cdf26ef39d3a3242fec25e07f020cc4feec4c4a865e340569"},
+    {file = "numba-0.59.1-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:8c8b4477763cb1fbd86a3be7050500229417bf60867c93e131fd2626edb02238"},
+    {file = "numba-0.59.1-cp312-cp312-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:7d80bce4ef7e65bf895c29e3889ca75a29ee01da80266a01d34815918e365835"},
+    {file = "numba-0.59.1-cp312-cp312-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:f7ad1d217773e89a9845886401eaaab0a156a90aa2f179fdc125261fd1105096"},
+    {file = "numba-0.59.1-cp312-cp312-win_amd64.whl", hash = "sha256:5bf68f4d69dd3a9f26a9b23548fa23e3bcb9042e2935257b471d2a8d3c424b7f"},
+    {file = "numba-0.59.1-cp39-cp39-macosx_10_9_x86_64.whl", hash = "sha256:4e0318ae729de6e5dbe64c75ead1a95eb01fabfe0e2ebed81ebf0344d32db0ae"},
+    {file = "numba-0.59.1-cp39-cp39-macosx_11_0_arm64.whl", hash = "sha256:0f68589740a8c38bb7dc1b938b55d1145244c8353078eea23895d4f82c8b9ec1"},
+    {file = "numba-0.59.1-cp39-cp39-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:649913a3758891c77c32e2d2a3bcbedf4a69f5fea276d11f9119677c45a422e8"},
+    {file = "numba-0.59.1-cp39-cp39-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:9712808e4545270291d76b9a264839ac878c5eb7d8b6e02c970dc0ac29bc8187"},
+    {file = "numba-0.59.1-cp39-cp39-win_amd64.whl", hash = "sha256:8d51ccd7008a83105ad6a0082b6a2b70f1142dc7cfd76deb8c5a862367eb8c86"},
+    {file = "numba-0.59.1.tar.gz", hash = "sha256:76f69132b96028d2774ed20415e8c528a34e3299a40581bae178f0994a2f370b"},
+]
+
+[package.dependencies]
+llvmlite = "==0.42.*"
+numpy = ">=1.22,<1.27"
+
 [[package]]
 name = "numpy"
 version = "1.26.3"
@@ -2037,11 +2101,8 @@ description = "ONNX Runtime is a runtime accelerator for Machine Learning models
 optional = false
 python-versions = "*"
 files = [
-    {file = "onnxruntime_openvino-1.17.1-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:6ed693011b472f9a617b2d5c4785d5fa1e1b77f7cb2b02e47b899534ec6c6396"},
     {file = "onnxruntime_openvino-1.17.1-cp310-cp310-win_amd64.whl", hash = "sha256:5152b5e56e83e022ced2986700d68dd8ba7b1466761725ce774f679c5710ab87"},
-    {file = "onnxruntime_openvino-1.17.1-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:2ce3b1aa06d6b8b732d314d217028ec4735de5806215c44d3bdbcad03b9260d5"},
     {file = "onnxruntime_openvino-1.17.1-cp311-cp311-win_amd64.whl", hash = "sha256:21133a701bb07ea19e01f48b8c23beee575f2e879f49173843f275d7c91a625a"},
-    {file = "onnxruntime_openvino-1.17.1-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:76824dac3c392ad4b812f29c18be2055ab3bba2e3c111e44baae847b33d5b081"},
 ]
 
 [package.dependencies]
@@ -2601,7 +2662,6 @@ files = [
     {file = "PyYAML-6.0.1-cp311-cp311-win_amd64.whl", hash = "sha256:bf07ee2fef7014951eeb99f56f39c9bb4af143d8aa3c21b1677805985307da34"},
     {file = "PyYAML-6.0.1-cp312-cp312-macosx_10_9_x86_64.whl", hash = "sha256:855fb52b0dc35af121542a76b9a84f8d1cd886ea97c84703eaa6d88e37a2ad28"},
     {file = "PyYAML-6.0.1-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:40df9b996c2b73138957fe23a16a4f0ba614f4c0efce1e9406a184b6d07fa3a9"},
-    {file = "PyYAML-6.0.1-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:a08c6f0fe150303c1c6b71ebcd7213c2858041a7e01975da3a99aed1e7a378ef"},
     {file = "PyYAML-6.0.1-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:6c22bec3fbe2524cde73d7ada88f6566758a8f7227bfbf93a408a9d86bcc12a0"},
     {file = "PyYAML-6.0.1-cp312-cp312-musllinux_1_1_x86_64.whl", hash = "sha256:8d4e9c88387b0f5c7d5f281e55304de64cf7f9c0021a3525bd3b1c542da3b0e4"},
     {file = "PyYAML-6.0.1-cp312-cp312-win32.whl", hash = "sha256:d483d2cdf104e7c9fa60c544d92981f12ad66a457afae824d146093b8c294c54"},
@@ -3572,4 +3632,4 @@ testing = ["coverage (>=5.0.3)", "zope.event", "zope.testing"]
 [metadata]
 lock-version = "2.0"
 python-versions = ">=3.10,<3.12"
-content-hash = "db51ad1e631b569e106927683a13124252bd80974def1f2edbe23ac87d89c461"
+content-hash = "a44e079d565fc1166458690ca2dc5826e198cc07ccab0ebaf71b5ab5e0eed150"

machine-learning/pyproject.toml

@@ -23,6 +23,7 @@ orjson = ">=3.9.5"
 gunicorn = ">=21.1.0"
 huggingface-hub = ">=0.20.1,<1.0"
 tokenizers = ">=0.15.0,<1.0"
+numba = "^0.59.1"
 
 [tool.poetry.group.dev.dependencies]
 mypy = ">=1.3.0"