preprocessor/feature_extractor.py

import json
from pathlib import Path

import numpy as np


class HapticFeatureExtractor:
    def __init__(self, config):
        self.config = config
        self.window_size = int(config["window_size"])
        self.padding_value = float(config.get("padding_value", 0.0))
        self.return_attention_mask = bool(config.get("return_attention_mask", True))
        normalization = config.get("normalization", {})
        self.mean = np.asarray(normalization.get("mean", []), dtype=np.float32)
        self.std = np.asarray(normalization.get("std", []), dtype=np.float32)

    @classmethod
    def from_pretrained(cls, root):
        root_path = Path(root)
        config_path = root_path / "preprocessor" / "preprocessor_config.json"
        with config_path.open("r", encoding="utf-8") as handle:
            config = json.load(handle)
        return cls(config)

    def _normalize(self, values):
        if not self.config.get("normalize", True):
            return values
        if self.mean.size == 0 or self.std.size == 0:
            return values
        denom = np.where(self.std == 0, 1.0, self.std)
        return (values - self.mean) / denom

    def __call__(self, values):
        values = np.asarray(values, dtype=np.float32)
        if values.ndim != 2:
            raise ValueError("Expected input shape [sequence_length, num_channels].")

        values = self._normalize(values)
        length, channels = values.shape

        if length >= self.window_size:
            trimmed = values[: self.window_size]
            attention_mask = np.ones(self.window_size, dtype=np.int64)
        else:
            pad_amount = self.window_size - length
            padding = np.full((pad_amount, channels), self.padding_value, dtype=np.float32)
            trimmed = np.concatenate([values, padding], axis=0)
            attention_mask = np.concatenate(
                [
                    np.ones(length, dtype=np.int64),
                    np.zeros(pad_amount, dtype=np.int64),
                ]
            )

        result = {"input_values": trimmed}
        if self.return_attention_mask:
            result["attention_mask"] = attention_mask
        return result