Implementation:Facebookresearch Habitat lab WindowedRunningMean

Knowledge Sources	Facebookresearch_Habitat_lab
Domains	Embodied_AI, Statistics
Last Updated	2026-02-15 00:00 GMT

Overview

WindowedRunningMean is an efficient implementation of a windowed running mean that supports both finite window sizes and an infinite (cumulative) window using a circular buffer.

Description

WindowedRunningMean uses the attrs library for a compact, slot-based dataclass definition. It maintains a running sum and count. When the window size is finite, a NumPy circular buffer of the specified size stores recent values; once the buffer is full, the oldest value is subtracted from the running sum when a new value is added. When the window is infinite (signaled by math.isinf(window_size) or a non-positive value), it simply accumulates all values. The class provides properties for mean, sum, count, and infinite_window, and supports += via __iadd__ and float() conversion via __float__.

Usage

Use WindowedRunningMean when you need to track a running average over a sliding window of recent values, such as for smoothing training metrics or timing measurements. Pass float('inf') as the window size for cumulative averaging.

Code Reference

Source Location

• Repository: Facebookresearch_Habitat_lab
• File: habitat-baselines/habitat_baselines/common/windowed_running_mean.py
• Lines: 16-74

Signature

@attr.s(auto_attribs=True, slots=True, repr=False)
class WindowedRunningMean:
    window_size: Union[int, float]

Import

from habitat_baselines.common.windowed_running_mean import WindowedRunningMean

I/O Contract

Inputs

Name	Type	Required	Description
window_size	Union[int, float]	Yes	Size of the sliding window. Use float('inf') or a non-positive value for cumulative (infinite) averaging.

Outputs

Name	Type	Description
mean	float	The current running mean of values in the window
sum	float	The current running sum of values in the window
count	int	Number of values currently tracked (up to window_size)

Key Methods

add

def add(self, v_i: Union[numbers.Real, float, int]) -> None

Adds a single value to the running mean.

add_many

def add_many(self, vs: Sequence[Union[numbers.Real, float, int]])

Adds multiple values sequentially.

Usage Examples

Basic Usage

from habitat_baselines.common.windowed_running_mean import WindowedRunningMean

# Finite window of 100 values
tracker = WindowedRunningMean(window_size=100)
for value in range(200):
    tracker.add(value)

print(tracker.mean)   # Mean of the last 100 values (150.0 - 199.0)
print(tracker.count)  # 100

# Infinite window (cumulative mean)
cumulative = WindowedRunningMean(window_size=float("inf"))
cumulative += 10.0
cumulative += 20.0
print(float(cumulative))  # 15.0