Class StreamingAverage

src/transformers/optimization.py:581–618

Rolling window average for smoothing metric values. Maintains a sliding window of values and computes their average, useful for smoothing noisy metric values before making learning rate decisions. Args: window_size (`int`): The maximum number of values to keep in th

class StreamingAverage:
    """Rolling window average for smoothing metric values.

    Maintains a sliding window of values and computes their average,
    useful for smoothing noisy metric values before making learning rate decisions.

    Args:
        window_size (`int`):
            The maximum number of values to keep in the rolling window.
    """

    def __init__(self, window_size: int) -> None:
        self.window_size: int = window_size
        self.values: list[float] = []
        self.sum: float = 0.0

    def streamavg(self, value: float) -> float:
        """Add a value and return the current rolling average."""
        self.values.append(value)
        self.sum += value

        if len(self.values) > self.window_size:
            removed = self.values.pop(0)
            self.sum -= removed

        return self.sum / len(self.values)

    def state_dict(self) -> dict[str, Any]:
        return {
            "window_size": self.window_size,
            "values": self.values.copy(),
            "sum": self.sum,
        }

    def load_state_dict(self, state_dict: dict[str, Any]) -> None:
        self.window_size = state_dict.get("window_size", self.window_size)
        self.values = state_dict.get("values", []).copy()
        self.sum = state_dict.get("sum", 0.0)


class GreedyLR: