def on_phase_end()

in classy_vision/hooks/visdom_hook.py [0:0]

• 37 lines of code
• 14 McCabe index (conditional complexity)

    def on_phase_end(self, task) -> None:
        """
        Plot the metrics on visdom.
        """
        phase_type = task.phase_type
        metrics = self.metrics
        batches = len(task.losses)

        if batches == 0:
            return

        # Loss for the phase
        loss = sum(task.losses) / batches
        loss_key = phase_type + "_loss"
        if loss_key not in metrics:
            metrics[loss_key] = []
        metrics[loss_key].append(loss)

        # Optimizer LR for the phase
        optimizer_lr = task.optimizer.options_view.lr
        lr_key = phase_type + "_learning_rate"
        if lr_key not in metrics:
            metrics[lr_key] = []
        metrics[lr_key].append(optimizer_lr)

        # Calculate meters
        for meter in task.meters:
            if isinstance(meter.value, collections.MutableMapping):
                flattened_meters_dict = flatten_dict(meter.value, prefix=meter.name)
                for k, v in flattened_meters_dict.items():
                    metric_key = phase_type + "_" + k
                    if metric_key not in metrics:
                        metrics[metric_key] = []
                    metrics[metric_key].append(v)
            else:
                metric_key = phase_type + "_" + meter.name
                if metric_key not in metrics:
                    metrics[metric_key] = []
                metrics[metric_key].append(meter.value)

        # update learning curve visualizations:
        phase_type = "train" if task.train else "test"
        title = "%s-%s" % (phase_type, task.base_model.__class__.__name__)
        title += self.title_suffix

        if not task.train and is_primary():
            logging.info("Plotting learning curves to visdom")
            plot_learning_curves(
                metrics, visdom_server=self.visdom, env=self.env, win=title, title=title
            )