extensible.extensions.utils

Classes

Avg(target[, accum, visualize, fig, counter])

Computes the average of a scalar fixture across an entire evaluation epoch

class extensible.extensions.utils.Avg(target, accum=None, visualize: bool = True, fig: str = 'avgs/{datasource_name}/{target}', counter='num_batches')

Bases: Extension

Computes the average of a scalar fixture across an entire evaluation epoch

Optionally, plots a visualization of the average at the end of the evaluation epoch.

This extension can be used by adding an instance of it to a train manager after an extension that generates the fixture to average. Another possibility is to derive it, e.g., :

class ThresholdedLoss(Avg):
    def __init__(self, *args, **kwargs):
        super().__init__("thresh_loss", *args, **kwargs)

    def post_eval_step_batch(self, train_manager, fixtures, batch, prediction):

        # Add the fixture to average.
        hard_prediction = prediction>0.5
        fixtures["thresh_loss"] = train_manager.loss(targets, hard_prediction)

        # Call the super's hook to carry out the averaging
        fixtures(super().post_eval_step_batch)

Exposes the following fixtures:

• self.accum, which by default is 'accum_' + self.target

Depends on EvalState

Parameters:

• target – The name of the fixture to accumulate

• accum – The name of the fixture to create as an accumulator

• visualize – Whether to create a scalar graph

• fig – The name of the figure visualization will be placed – can include formatting placeholders for any fixture and target and accum.

• counter' – The counter that will be used to divide the running sum