Method train_epoch

funasr/train_utils/trainer.py:345–540

Defines the training process for a single epoch with gradient accumulation. Args: epoch (int): The current epoch number.

(
        self,
        model=None,
        optim=None,
        scheduler=None,
        scaler=None,
        dataloader_train=None,
        dataloader_val=None,
        epoch=None,
        writer=None,
        **kwargs,
    )

            dist.barrier()

    def train_epoch(
        self,
        model=None,
        optim=None,
        scheduler=None,
        scaler=None,
        dataloader_train=None,
        dataloader_val=None,
        epoch=None,
        writer=None,
        **kwargs,
    ):
        """
        Defines the training process for a single epoch with gradient accumulation.
        Args:
            epoch (int): The current epoch number.
        """
        if self.use_ddp or self.use_fsdp:
            dist.barrier()
        logging.info(f"Train epoch: {epoch}, rank: {self.rank}\n")
        model.train()

        # Set the number of steps for gradient accumulation
        accum_grad = self.accum_grad
        # Initialize the gradient accumulation
        optim.zero_grad()
        speed_stats = {}

        iterator_stop = torch.tensor(0).to(self.device)

        dataloader_train.batch_sampler.set_epoch(epoch)
        time_beg = time.perf_counter()
        time5 = time_beg
        for batch_idx, batch in enumerate(dataloader_train):
            # if self.use_ddp or self.use_fsdp:
            #     dist.all_reduce(iterator_stop, dist.ReduceOp.SUM)
            #     if iterator_stop > 0:
            #         break
            self.batch_total += 1
            self.step_in_epoch += 1
            time1 = time.perf_counter()
            speed_stats["data_load"] = f"{time1-time_beg:0.3f}"

            batch = to_device(batch, self.device, non_blocking=True)

            my_context = nullcontext
            if self.use_ddp or self.use_fsdp:
                my_context = model.no_sync if batch_idx % accum_grad != 0 else my_context
            with my_context():
                time2 = time.perf_counter()
                with maybe_autocast(self.amp_enabled, dtype=self.amp_dtype):
                    retval = model(**batch)

                    # if (
                    #     self.reset_gpu_cache
                    #     and (torch.cuda.max_memory_reserved() / 1024 / 1024 / 1024) > 70
                    # ):
                    #     torch.cuda.empty_cache()