StatisticsCollector

class tfsnippet.utils.StatisticsCollector(shape=())

Bases: object

Computing \(\mathrm{E}[X]\) and \(\operatorname{Var}[X]\) online.

Attributes Summary

counter	Get the counter of collected values.
has_value	Whether or not any value has been collected?
mean	Get the mean of the values, i.e., \(\mathrm{E}[X]\).
shape	Get the shape of the values.
square	Get \(\mathrm{E}[X^2]\) of the values.
stddev	Get the std of the values, i.e., \(\sqrt{\operatorname{Var}[X]}\).
var	Get the variance of the values, i.e., \(\operatorname{Var}[X]\).
weight_sum	Get the weight summation.

Methods Summary

collect(values[, weight])	Update the statistics from values.
reset()	Reset the collector to initial state.

Attributes Documentation

counter

Get the counter of collected values.

has_value

Whether or not any value has been collected?

mean

Get the mean of the values, i.e., \(\mathrm{E}[X]\).

shape

Get the shape of the values.

square

Get \(\mathrm{E}[X^2]\) of the values.

stddev

Get the std of the values, i.e., \(\sqrt{\operatorname{Var}[X]}\).

var

Get the variance of the values, i.e., \(\operatorname{Var}[X]\).

weight_sum

Get the weight summation.

Methods Documentation

collect(values, weight=1.0)

Update the statistics from values.

This method uses the following equation to update mean and square:

\[\frac{\sum_{i=1}^n w_i f(x_i)}{\sum_{j=1}^n w_j} = \frac{\sum_{i=1}^m w_i f(x_i)}{\sum_{j=1}^m w_j} + \frac{\sum_{i=m+1}^n w_i}{\sum_{j=1}^n w_j} \Bigg( \frac{\sum_{i=m+1}^n w_i f(x_i)}{\sum_{j=m+1}^n w_j} - \frac{\sum_{i=1}^m w_i f(x_i)}{\sum_{j=1}^m w_j} \Bigg)\]

Parameters:

• values – Values to be collected in batch, numpy array or scalar whose shape ends with self.shape. The leading shape in front of self.shape is regarded as the batch shape.
• weight – Weights of the values, should be broadcastable against the batch shape. (default is 1)

Raises:

ValueError – If the shape of values does not end with self.shape.

reset()

Reset the collector to initial state.