Shortcuts

Source code for catalyst.tools.meters.averagevaluemeter

"""
Average value meter
"""
import numpy as np

from . import meter


[docs]class AverageValueMeter(meter.Meter): """ Average value meter stores mean and standard deviation for population of input values. Meter updates are applied online, one value for each update. Values are not cached, only the last added. """
[docs] def __init__(self): """Constructor method for the ``AverageValueMeter`` class.""" super(AverageValueMeter, self).__init__() self.n = 0 self.val = 0.0 self.mean = np.nan self.mean_old = 0.0 self.m_s = 0.0 self.std = np.nan self.n_samples = 0
[docs] def add(self, value, batch_size) -> None: """Add a new observation. Updates of mean and std are going online, with `Welford's online algorithm <https://en.wikipedia.org/wiki/Algorithms_for_calculating_variance>`_. Args: value (float): value for update, can be scalar number or PyTorch tensor .. note:: Because of algorithm design, you can update meter values with only one value a time. """ self.val = value self.n += 1 self.n_samples += batch_size if self.n == 1: self.mean = 0.0 + value # Force a copy in torch/numpy self.std = 0.0 self.mean_old = self.mean self.m_s = 0.0 else: self.mean = self.mean_old + ( value - self.mean_old ) * batch_size / float(self.n_samples) self.m_s += ( (value - self.mean_old) * (value - self.mean) * batch_size ) self.mean_old = self.mean self.std = np.sqrt(self.m_s / (self.n_samples - 1.0))
[docs] def value(self): """Returns meter values. Returns: mean (float): Mean that has been updated online. std (float): Standard deviation that has been updated online. """ return self.mean, self.std
[docs] def reset(self): """Resets the meter to default settings.""" self.n = 0 self.val = 0.0 self.mean = np.nan self.mean_old = 0.0 self.m_s = 0.0 self.std = np.nan self.n_samples = 0