pretrained

feat.pretrained

Helper functions specifically for working with included pre-trained models

detector_capabilities()

Per-detector supported models, the single source of truth for what each detector class accepts. JSON-serializable (None stays None -> null).

Source code in feat/pretrained.py

def detector_capabilities():
    """Per-detector supported models, the single source of truth for what each
    detector class accepts. JSON-serializable (None stays None -> null)."""
    from feat.detector import Detectorv1
    from feat.detector_v2 import Detectorv2
    from feat.MPDetector import MPDetector
    return {
        "Detectorv1": Detectorv1.SUPPORTED_MODELS,
        "Detectorv2": Detectorv2.SUPPORTED_MODELS,
        "MPDetector": MPDetector.SUPPORTED_MODELS,
    }

fetch_model(model_type, model_name)

Fetch a pre-trained model class constructor. Used by detector init

Source code in feat/pretrained.py

def fetch_model(model_type, model_name):
    """Fetch a pre-trained model class constructor. Used by detector init"""
    if model_name is None:
        raise ValueError(f"{model_type} must be a valid string model name, not None")
    model_type = PRETRAINED_MODELS[model_type]
    matches = list(filter(lambda e: model_name in e.keys(), model_type))[0]
    return list(matches.values())[0]

get_pretrained_models(face_model, landmark_model, au_model, emotion_model, facepose_model, identity_model, verbose)

Helper function that validates the request model names and downloads them if necessary using the URLs in the included JSON file. User by detector init

Source code in feat/pretrained.py

def get_pretrained_models(
    face_model,
    landmark_model,
    au_model,
    emotion_model,
    facepose_model,
    identity_model,
    verbose,
):
    """Helper function that validates the request model names and downloads them if
    necessary using the URLs in the included JSON file. User by detector init"""

    # Get supported model URLs
    with open(os.path.join(get_resource_path(), "model_list.json"), "r") as f:
        model_urls = json.load(f)

    get_names = lambda s: list(
        map(
            lambda e: list(e.keys())[0],
            PRETRAINED_MODELS[s],
        )
    )

    # Face model
    if face_model is None:
        raise ValueError(
            f"face_model must be one of {[list(e.keys())[0] for e in PRETRAINED_MODELS['face_model']]}"
        )
    else:
        face_model = face_model.lower()
        if face_model not in get_names("face_model"):
            raise ValueError(
                f"Requested face_model was {face_model}. Must be one of {[list(e.keys())[0] for e in PRETRAINED_MODELS['face_model']]}"
            )
        for url in model_urls["face_detectors"][face_model]["urls"]:
            download_url(url, get_resource_path(), verbose=verbose)

    # Landmark model
    if landmark_model is None:
        raise ValueError(
            f"landmark_model must be one of {[list(e.keys())[0] for e in PRETRAINED_MODELS['landmark_model']]}"
        )
    else:
        landmark_model = landmark_model.lower()
        if landmark_model not in get_names("landmark_model"):
            raise ValueError(
                f"Requested landmark_model was {landmark_model}. Must be one of {[list(e.keys())[0] for e in PRETRAINED_MODELS['landmark_model']]}"
            )
        for url in model_urls["landmark_detectors"][landmark_model]["urls"]:
            download_url(url, get_resource_path(), verbose=verbose)

    # AU model
    if au_model is None:
        raise ValueError(
            f"au_model must be one of {[list(e.keys())[0] for e in PRETRAINED_MODELS['au_model']]}"
        )
    else:
        au_model = au_model.lower()
        if au_model not in get_names("au_model"):
            raise ValueError(
                f"Requested au_model was {au_model}. Must be one of {[list(e.keys())[0]for e in PRETRAINED_MODELS['au_model']]}"
            )

        for url in model_urls["au_detectors"][au_model]["urls"]:
            download_url(url, get_resource_path(), verbose=verbose)
            if au_model in ["xgb", "svm"]:
                download_url(
                    model_urls["au_detectors"]["hog-pca"]["urls"][0],
                    get_resource_path(),
                    verbose=verbose,
                )
                download_url(
                    model_urls["au_detectors"]["hog-pca"]["urls"][1],
                    get_resource_path(),
                    verbose=verbose,
                )
                download_url(
                    model_urls["au_detectors"]["hog-pca"]["urls"][2],
                    get_resource_path(),
                    verbose=verbose,
                )
                download_url(
                    model_urls["au_detectors"]["hog-pca"]["urls"][3],
                    get_resource_path(),
                    verbose=verbose,
                )
                download_url(
                    model_urls["au_detectors"]["hog-pca"]["urls"][4],
                    get_resource_path(),
                    verbose=verbose,
                )
                download_url(
                    model_urls["au_detectors"]["hog-pca"]["urls"][5],
                    get_resource_path(),
                    verbose=verbose,
                )
    # Emotion model
    if emotion_model is None:
        raise ValueError(
            f"emotion_model must be one of {[list(e.keys())[0] for e in PRETRAINED_MODELS['emotion_model']]}"
        )
    else:
        emotion_model = emotion_model.lower()
        if emotion_model not in get_names("emotion_model"):
            raise ValueError(
                f"Requested emotion_model was {emotion_model}. Must be one of {[list(e.keys())[0] for e in PRETRAINED_MODELS['emotion_model']]}"
            )
        for url in model_urls["emotion_detectors"][emotion_model]["urls"]:
            download_url(url, get_resource_path(), verbose=verbose)
            if emotion_model in ["svm"]:
                download_url(
                    model_urls["emotion_detectors"]["emo_pca"]["urls"][0],
                    get_resource_path(),
                    verbose=verbose,
                )
                download_url(
                    model_urls["emotion_detectors"]["emo_scalar"]["urls"][0],
                    get_resource_path(),
                    verbose=verbose,
                )

    # Facepose model
    if facepose_model is None:
        raise ValueError(
            f"facepose_model must be one of {[list(e.keys())[0] for e in PRETRAINED_MODELS['facepose_model']]}"
        )
    else:
        facepose_model = facepose_model.lower()
        if facepose_model not in get_names("facepose_model"):
            raise ValueError(
                f"Requested facepose_model was {facepose_model}. Must be one of {[list(e.keys())[0] for e in PRETRAINED_MODELS['facepose_model']]}"
            )
        for url in model_urls["facepose_detectors"][facepose_model]["urls"]:
            download_url(url, get_resource_path(), verbose=verbose)

    # Face Identity model
    if identity_model is None:
        raise ValueError(
            f"identity_model must be one of {[list(e.keys())[0] for e in PRETRAINED_MODELS['identity_model']]}"
        )
    else:
        identity_model = identity_model.lower()
        if identity_model not in get_names("identity_model"):
            raise ValueError(
                f"Requested identity_model was {identity_model}. Must be one of {[list(e.keys())[0] for e in PRETRAINED_MODELS['identity_model']]}"
            )
        for url in model_urls["identity_detectors"][identity_model]["urls"]:
            download_url(url, get_resource_path(), verbose=verbose)

    return (
        face_model,
        landmark_model,
        au_model,
        emotion_model,
        facepose_model,
        identity_model,
    )

load_model_weights(model_type='au', model='xgb', location='huggingface')

Load weights for the AU models

Source code in feat/pretrained.py

def load_model_weights(model_type="au", model="xgb", location="huggingface"):
    """Load weights for the AU models"""
    if model_type == "au":
        if model == "xgb":
            if location == "huggingface":
                # Load the entire model from skops serialized file
                model_path = hf_hub_download_with_fallback(
                    repo_id="py-feat/xgb_au",
                    filename="xgb_au_classifier_v2.skops",
                    fallback_filename="xgb_au_classifier.skops",
                    cache_dir=get_resource_path(),
                )
                unknown_types = get_untrusted_types(file=model_path)
                loaded_model = load(model_path, trusted=unknown_types)
                return {
                    "scaler_upper": loaded_model.scaler_upper,
                    "pca_model_upper": loaded_model.pca_model_upper,
                    "scaler_lower": loaded_model.scaler_lower,
                    "pca_model_lower": loaded_model.pca_model_lower,
                    "scaler_full": loaded_model.scaler_full,
                    "pca_model_full": loaded_model.pca_model_full,
                    "au_classifiers": loaded_model.classifiers,
                }
            elif location == "local":
                # Load weights from local Resources folder
                scaler_upper = load_classifier_pkl(
                    os.path.join(get_resource_path(), "all_data_Upperscalar_June30.pkl")
                )
                pca_model_upper = load_classifier_pkl(
                    os.path.join(get_resource_path(), "all_data_Upperpca_June30.pkl")
                )
                scaler_lower = load_classifier_pkl(
                    os.path.join(get_resource_path(), "all_data_Lowerscalar_June30.pkl")
                )
                pca_model_lower = load_classifier_pkl(
                    os.path.join(get_resource_path(), "all_data_Lowerpca_June30.pkl")
                )
                scaler_full = load_classifier_pkl(
                    os.path.join(get_resource_path(), "all_data_Fullscalar_June30.pkl")
                )
                pca_model_full = load_classifier_pkl(
                    os.path.join(get_resource_path(), "all_data_Fullpca_June30.pkl")
                )

                au_keys = [
                    "AU1",
                    "AU2",
                    "AU4",
                    "AU5",
                    "AU6",
                    "AU7",
                    "AU9",
                    "AU10",
                    "AU11",
                    "AU12",
                    "AU14",
                    "AU15",
                    "AU17",
                    "AU20",
                    "AU23",
                    "AU24",
                    "AU25",
                    "AU26",
                    "AU28",
                    "AU43",
                ]
                classifiers = {}
                for key in au_keys:
                    classifier = xgb.XGBClassifier()
                    classifier.load_model(
                        os.path.join(get_resource_path(), f"July4_{key}_XGB.ubj")
                    )
                    classifiers[key] = classifier
                return {
                    "scaler_upper": scaler_upper,
                    "pca_model_upper": pca_model_upper,
                    "scaler_lower": scaler_lower,
                    "pca_model_lower": pca_model_lower,
                    "scaler_full": scaler_full,
                    "pca_model_full": pca_model_full,
                    "au_classifiers": classifiers,
                }

        elif model == "svm":
            if location == "huggingface":
                # Load the entire model from skops serialized file
                model_path = hf_hub_download_with_fallback(
                    repo_id="py-feat/svm_au",
                    filename="svm_au_classifier_v2.skops",
                    fallback_filename="svm_au_classifier.skops",
                    cache_dir=get_resource_path(),
                )
                unknown_types = get_untrusted_types(file=model_path)
                loaded_model = load(model_path, trusted=unknown_types)
                return {
                    "scaler_upper": loaded_model.scaler_upper,
                    "pca_model_upper": loaded_model.pca_model_upper,
                    "scaler_lower": loaded_model.scaler_lower,
                    "pca_model_lower": loaded_model.pca_model_lower,
                    "scaler_full": loaded_model.scaler_full,
                    "pca_model_full": loaded_model.pca_model_full,
                    "au_classifiers": loaded_model.classifiers,
                }
            elif location == "local":
                # Load weights from local Resources folder
                scaler_upper = load_classifier_pkl(
                    os.path.join(get_resource_path(), "all_data_Upperscalar_June30.pkl")
                )
                pca_model_upper = load_classifier_pkl(
                    os.path.join(get_resource_path(), "all_data_Upperpca_June30.pkl")
                )
                scaler_lower = load_classifier_pkl(
                    os.path.join(get_resource_path(), "all_data_Lowerscalar_June30.pkl")
                )
                pca_model_lower = load_classifier_pkl(
                    os.path.join(get_resource_path(), "all_data_Lowerpca_June30.pkl")
                )
                scaler_full = load_classifier_pkl(
                    os.path.join(get_resource_path(), "all_data_Fullscalar_June30.pkl")
                )
                pca_model_full = load_classifier_pkl(
                    os.path.join(get_resource_path(), "all_data_Fullpca_June30.pkl")
                )
                classifiers = load_classifier_pkl(
                    os.path.join(get_resource_path(), "svm_60_July2023.pkl")
                )
                return {
                    "scaler_upper": scaler_upper,
                    "pca_model_upper": pca_model_upper,
                    "scaler_lower": scaler_lower,
                    "pca_model_lower": pca_model_lower,
                    "scaler_full": scaler_full,
                    "pca_model_full": pca_model_full,
                    "au_classifiers": classifiers,
                }
            else:
                raise ValueError(f"This function does not support {model_type} {model}")
    elif model_type == "emotion":
        if model == "svm":
            if location == "huggingface":
                # Load the entire model from skops serialized file
                model_path = hf_hub_download(
                    repo_id="py-feat/svm_emo",
                    filename="svm_emo_classifier.skops",
                    cache_dir=get_resource_path(),
                )
                unknown_types = get_untrusted_types(file=model_path)
                loaded_model = load(model_path, trusted=unknown_types)
                return {
                    "scaler_full": loaded_model.scaler_full,
                    "pca_model_full": loaded_model.pca_model_full,
                    "emo_classifiers": loaded_model.classifiers,
                }
            elif location == "local":
                # Load weights from local Resources folder
                scaler_full = load_classifier_pkl(
                    os.path.join(get_resource_path(), "emo_data_Fullscalar_Jun30.pkl")
                )
                pca_model_full = load_classifier_pkl(
                    os.path.join(get_resource_path(), "emo_data_Fullpca_Jun30.pkl")
                )
                classifiers = load_classifier_pkl(
                    os.path.join(get_resource_path(), "July4_emo_SVM.pkl")
                )
                return {
                    "scaler_full": scaler_full,
                    "pca_model_full": pca_model_full,
                    "emo_classifiers": classifiers,
                }
        else:
            raise ValueError(f"This function does not support {model_type} {model}")
    else:
        raise ValueError(f"This function does not support {model_type}")