def _do_loss_eval()

in ubteacher/engine/hooks.py [0:0]

• 29 lines of code
• 12 McCabe index (conditional complexity)

    def _do_loss_eval(self):
        record_acc_dict = {}
        with inference_context(self._model), torch.no_grad():
            for _, inputs in enumerate(self._data_loader):
                record_dict = self._get_loss(inputs, self._model)
                # accumulate the losses
                for loss_type in record_dict.keys():
                    if loss_type not in record_acc_dict.keys():
                        record_acc_dict[loss_type] = record_dict[loss_type]
                    else:
                        record_acc_dict[loss_type] += record_dict[loss_type]
            # average
            for loss_type in record_acc_dict.keys():
                record_acc_dict[loss_type] = record_acc_dict[loss_type] / len(
                    self._data_loader
                )

            # divide loss and other metrics
            loss_acc_dict = {}
            for key in record_acc_dict.keys():
                if key[:4] == "loss":
                    loss_acc_dict[key] = record_acc_dict[key]

            # only output the results of major node
            if comm.is_main_process():
                total_losses_reduced = sum(loss for loss in loss_acc_dict.values())
                self.trainer.storage.put_scalar(
                    "val_total_loss_val" + self._model_name, total_losses_reduced
                )

                record_acc_dict = {
                    "val_" + k + self._model_name: record_acc_dict[k]
                    for k in record_acc_dict.keys()
                }

                if len(record_acc_dict) > 1:
                    self.trainer.storage.put_scalars(**record_acc_dict)