data_loader.py

def load_embeddings(
    embedding_list_train: list,
    label_path: str,
    mapping_path: str,
    data_dir: str,
    batch_size: int = 4,
) -> Tuple[np.ndarray, np.ndarray]:
    encoding_filename_mapping = json.load(open(mapping_path, "r"))
    raw_labels = json.load(open(label_path, "r"))
    X_train, y_train = (), ()
    for key in embedding_list_train:
        real_filename = encoding_filename_mapping[key.replace(
            ".tfrecords", "")]
        loaded = np.load("{}.npy".format(os.path.join(data_dir, key.replace(
            ".tfrecords", ""))))
        # logging.info("Video - {}".format(key))
        X_train += (*_get_chunk_array(loaded, batch_size),)
        # get flicker frame indexes
        flicker_idxs = np.array(raw_labels[real_filename]) - 1
        # buffer zeros array frame video embedding
        buf_label = np.zeros(loaded.shape[0], dtype=np.uint8)
        # set indexes in zeros array based on flicker frame indexes
        buf_label[flicker_idxs] = 1
        y_train += tuple(
            1 if sum(x) else 0
            for x in _get_chunk_array(buf_label, batch_size)
        )  # consider using tf reduce sum for multiclass

    return np.array(X_train), np.array(y_train)