Method train

src/base/base_inference.py:93–135

Full training logic

(self)

        return device, list_ids

    def train(self):
        """
        Full training logic
        """
        epoch = 0
        result = self._train_epoch(epoch)

        # save logged informations into log dict
        log = {'epoch': epoch}
        for key, value in result.items():
            if key == 'metrics':
                log.update({mtr.__name__: value[i] for i, mtr in enumerate(self.metrics)})
            elif key == 'val_metrics':
                log.update({'val_' + mtr.__name__: value[i] for i, mtr in enumerate(self.metrics)})
            else:
                log[key] = value

        # print logged informations to the screen
        if self.train_logger is not None:
            self.train_logger.add_entry(log)
            if self.verbosity >= 1:
                for key, value in log.items():
                    self.logger.info('    {:15s}: {}'.format(str(key), value))

        # evaluate model performance according to configured metric, save best checkpoint as model_best
        best = False
        monitor_value = None
        if self.monitor_mode != 'off':
            try:
                if (self.monitor_mode == 'min' and log[self.monitor] < self.monitor_best) or\
                        (self.monitor_mode == 'max' and log[self.monitor] > self.monitor_best):
                    self.monitor_best = log[self.monitor]
                    best = True
                monitor_value = log[self.monitor]

            except KeyError:
                if epoch == 1:
                    msg = "Warning: Can\'t recognize metric named '{}' ".format(self.monitor)\
                        + "for performance monitoring. model_best checkpoint won\'t be updated."
                    self.logger.warning(msg)

        if epoch % self.save_freq == 0 or best:
            self._save_checkpoint(epoch, save_best=best, monitor_value=monitor_value)
            

    def _train_epoch(self, epoch):