Dev gzf exp (#1647)

zhifu gao

2024-04-23 2ac38adbe5f4e1374a079e032ed4b504351a207c

 funasr/train_utils/trainer.py

@@ -252,6 +252,7 @@
                dataloader_val=None,
                epoch=None,
                writer=None,
                **kwargs,
                    ):
        """
        Defines the training process for a single epoch with gradient accumulation.
@@ -268,10 +269,12 @@
        # Initialize the gradient accumulation
        optim.zero_grad()
        speed_stats = {}
        time5 = time.perf_counter()
        
        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)
@@ -279,11 +282,13 @@
                    break
            self.batch_total += 1
            time1 = time.perf_counter()
            speed_stats["data_load"] = f"{time1-time5:0.3f}"
            speed_stats["data_load"] = f"{time1-time_beg:0.3f}"

            batch = to_device(batch, self.device)
            
            my_context = model.no_sync if batch_idx % accum_grad != 0 else nullcontext

            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.use_fp16):
@@ -370,6 +375,8 @@
                         stats=stats,
                         writer=writer,
                         tag="train",
                         data_split_i=kwargs.get("data_split_i", 0),
                         data_split_num=kwargs.get("data_split_num", 1),
                         )

            if (batch_idx + 1) % self.validate_interval == 0:
@@ -384,6 +391,7 @@
            if (batch_idx+1) % self.save_checkpoint_interval == 0:
                self.save_checkpoint(epoch, model=model, optim=optim, scheduler=scheduler, scaler=scaler, step=batch_idx+1)

            time_beg = time.perf_counter()
        else:
            if self.use_ddp or self.use_fsdp:
                iterator_stop.fill_(1)
@@ -502,6 +510,9 @@
            stats=None,
            writer=None,
            tag="train",
            data_split_i=0,
            data_split_num=1,
            **kwargs,
            ):
        
        if (batch_idx + 1) % self.log_interval == 0:
@@ -521,6 +532,7 @@
                f"{tag}, "
                f"rank: {self.local_rank}, "
                f"epoch: {epoch}/{self.max_epoch}, "
                f"data_slice: {data_split_i}/{data_split_num}, "
                f"step: {batch_idx + 1}/{batch_num_epoch}, total step: {self.batch_total}, "
                f"(loss_avg_rank: {loss:.3f}), "
                f"(loss_avg_epoch: {loss_avg_epoch:.3f}), "