model.py

import threading

import cv2
import mediapipe as mp
import numpy as np
import pandas as pd
from huggingface_hub import hf_hub_download
from tensorflow import keras


class Model:
    def __init__(self, model_repo_id: str, model_filename: str, hf_token: str):
        self.landmark_extractor = LandmarkExtractor()
        self.trained_model = self.create_prod_model()
        self.trained_model.compile(
            keras.optimizers.Adam(0.0005),
            loss=self.triplet_loss_init(0.2),
            metrics=[self.triplet_accuracy],
        )
        custom_objects = {
            "triplet_loss": self.triplet_loss_init(0.2),
            "triplet_accuracy": self.triplet_accuracy,
            "K": keras.backend,
            "keras": keras,
        }
        with keras.utils.custom_object_scope(custom_objects):
            weights = keras.models.load_model(
                hf_hub_download(model_repo_id, model_filename, token=hf_token, cache_dir=".cache")
            ).get_weights()
            self.trained_model.set_weights(weights)

    @staticmethod
    def fec_net(inputs):
        x = keras.layers.Dense(1024, activation="relu")(inputs)
        x = keras.layers.Dense(512, activation="relu")(x)
        x = keras.layers.Dense(512, activation="relu")(x)
        x = keras.layers.Dropout(0.5)(x)
        x = keras.layers.Dense(16)(x)
        outputs = keras.layers.Lambda(keras.backend.l2_normalize)(x)
        return outputs

    @classmethod
    def create_prod_model(cls):
        inputs = keras.layers.Input(shape=(478 * 3,))
        outputs = cls.fec_net(inputs)
        return keras.Model(inputs=inputs, outputs=outputs)

    @staticmethod
    def triplet_loss_init(alpha):
        def triplet_loss(y_true, y_pred):
            dimensions = 16
            batch_size = y_pred.shape[0]
            s1 = y_pred[:, :dimensions]
            s2 = y_pred[:, dimensions : 2 * dimensions]
            d = y_pred[:, 2 * dimensions :]

            s1_s2 = keras.backend.sum(keras.backend.square(s1 - s2), axis=1)
            s1_d = keras.backend.sum(keras.backend.square(s1 - d), axis=1)
            s2_d = keras.backend.sum(keras.backend.square(s2 - d), axis=1)

            loss = keras.backend.maximum(
                0.0, s1_s2 - s1_d + alpha
            ) + keras.backend.maximum(0.0, s1_s2 - s2_d + alpha)
            loss = keras.backend.mean(loss)

            return loss

        return triplet_loss

    @staticmethod
    def triplet_accuracy(y_true, y_pred):
        dimensions = 16
        s1 = y_pred[:, :dimensions]
        s2 = y_pred[:, dimensions : 2 * dimensions]
        d = y_pred[:, 2 * dimensions :]

        s1_s2 = keras.backend.sqrt(
            keras.backend.sum(keras.backend.square(s1 - s2), axis=1)
        )
        s1_d = keras.backend.sqrt(
            keras.backend.sum(keras.backend.square(s1 - d), axis=1)
        )
        s2_d = keras.backend.sqrt(
            keras.backend.sum(keras.backend.square(s2 - d), axis=1)
        )

        s1_match = keras.backend.less(s1_s2, s1_d)
        s2_match = keras.backend.less(s1_s2, s2_d)
        match = keras.backend.cast(
            keras.backend.all(keras.backend.stack([s1_match, s2_match]), axis=0),
            "float32",
        )

        acc = keras.backend.mean(match, axis=0)

        return acc

    @classmethod
    def load_model(cls, model_path: str):
        custom_objects = {
            "triplet_loss": cls.triplet_loss_init(0.2),
            "triplet_accuracy": cls.triplet_accuracy,
        }
        with keras.utils.custom_object_scope(custom_objects):
            return keras.models.load_model(model_path)

    def predict(self, x: np.ndarray):
        return self.trained_model.predict(x)

    def preprocess(self, image: bytes) -> np.ndarray:
        array_repr = np.asarray(bytearray(image), dtype=np.uint8)
        decoded_img = cv2.imdecode(array_repr, flags=cv2.IMREAD_COLOR)
        landmarks = self.landmark_extractor.extract_landmarks_flat(decoded_img)
        return pd.DataFrame([landmarks]).to_numpy()

    @staticmethod
    def distance(x1: np.ndarray, x2: np.ndarray):
        return np.linalg.norm(x1 - x2, ord=2)


class LandmarkExtractor:
    def __init__(self):
        self.lock = threading.Lock()
        self.face_mesh = mp.solutions.face_mesh.FaceMesh(
            static_image_mode=True,
            max_num_faces=1,
            refine_landmarks=True,
            min_detection_confidence=0.5,
        )

    @staticmethod
    def __normalise_landmarks(landmarks):
        left_eye_idx = 33
        right_eye_idx = 263
        left_eye_v = [landmarks[left_eye_idx].x, landmarks[left_eye_idx].y]
        right_eye_v = [landmarks[right_eye_idx].x, landmarks[right_eye_idx].y]
        xy_norm = np.linalg.solve(
            np.array([left_eye_v, right_eye_v]).T,
            np.array([[lmk.x, lmk.y] for lmk in landmarks]).T,
        ).T
        for i, lmk in enumerate(landmarks):
            lmk.x = xy_norm[i][0]
            lmk.y = xy_norm[i][1]
        return landmarks

    def extract_landmarks(self, img: np.ndarray):
        results = self.face_mesh.process(cv2.cvtColor(img, cv2.COLOR_BGR2RGB))
        if results.multi_face_landmarks is None:
            return None
        return self.__normalise_landmarks(results.multi_face_landmarks[0].landmark)

    def extract_landmarks_flat(self, img: np.ndarray):
        landmarks = self.extract_landmarks(img)
        if landmarks is None:
            return None
        flat_landmarks = {}
        for i, landmark in enumerate(landmarks):
            flat_landmarks.update(
                {
                    f"{i}_x": landmark.x,
                    f"{i}_y": landmark.y,
                    f"{i}_z": landmark.z,
                }
            )
        return flat_landmarks