Method log

funasr/train_utils/trainer_ds.py:848–927

Log. Args: loss_dict: TODO. tag: TODO. **kwargs: Additional keyword arguments.

(
        self,
        loss_dict: dict = None,
        tag="train",
        **kwargs,
    )

        model.train()

    def log(
        self,
        loss_dict: dict = None,
        tag="train",
        **kwargs,
    ):
        """Log.
        
            Args:
                loss_dict: TODO.
                tag: TODO.
                **kwargs: Additional keyword arguments.
            """
        loss = loss_dict["loss"].detach().cpu().item()
        epoch = loss_dict["epoch"]
        batch_idx = loss_dict["batch_idx"]
        step_in_epoch = loss_dict["step_in_epoch"]
        batch_total = loss_dict["batch_total"]
        batch_num_epoch = loss_dict["batch_num_epoch"]
        lr = loss_dict["lr"]

        speed_stats = loss_dict["speed_stats"]
        stats = loss_dict["stats"]
        data_split_i = loss_dict["data_split_i"]
        data_split_num = loss_dict["data_split_num"]
        log_step = loss_dict.get("log_step", None)

        if (batch_idx + 1) % self.log_interval == 0:
            batch_idx = log_step if log_step is not None else batch_idx
            gpu_info = (
                "GPU, memory: usage: {:.3f} GB, "
                "peak: {:.3f} GB, "
                "cache: {:.3f} GB, "
                "cache_peak: {:.3f} GB".format(
                    torch.cuda.memory_allocated() / 1024 / 1024 / 1024,
                    torch.cuda.max_memory_allocated() / 1024 / 1024 / 1024,
                    torch.cuda.memory_reserved() / 1024 / 1024 / 1024,
                    torch.cuda.max_memory_reserved() / 1024 / 1024 / 1024,
                )
            )

            loss_avg_epoch = getattr(self, f"{tag}_loss_avg")
            acc_avg_epoch = getattr(self, f"{tag}_acc_avg")
            description = (
                f"{tag}, "
                f"rank: {self.rank}, "
                f"epoch: {epoch}/{self.max_epoch}, "
                f"data_slice: {data_split_i}/{data_split_num}, "
                f"step_in_slice: {batch_idx + 1}/{batch_num_epoch}, step_in_epoch: {step_in_epoch}, total step: {batch_total}, "
                f"(loss_avg_rank: {loss:.3f}), "
                f"(loss_avg_slice: {loss_avg_epoch:.3f}), "
                f"(ppl_avg_slice: {math.exp(loss_avg_epoch):.3e}), "
                f"(acc_avg_slice: {acc_avg_epoch:.3f}), "
                f"(lr: {lr:.3e}), "
                f"{[(k, round(v.detach().cpu().item(), 3)) for k, v in stats.items()]}, "
                f"{speed_stats}, "
                f"{gpu_info}"
            )