update

游雁

2024-07-01 92b14aaa2a98e2872f1c8b62551ecaf82bc5688a

 funasr/train_utils/trainer.py

@@ -85,7 +85,12 @@
        self.batch_total = 0
        self.use_fp16 = use_fp16
        self.save_checkpoint_interval = kwargs.get("save_checkpoint_interval", 5000)
        self.validate_interval = kwargs.get("validate_interval", 5000)
        self.validate_interval = kwargs.get("validate_interval", -1)
        if self.validate_interval < 0:
            self.validate_interval = self.save_checkpoint_interval
        assert (
            self.save_checkpoint_interval == self.validate_interval
        ), f"save_checkpoint_interval must equal to validate_interval"
        self.keep_nbest_models = kwargs.get("keep_nbest_models", 500)
        self.avg_keep_nbest_models_type = kwargs.get("avg_keep_nbest_models_type", "acc")
        self.avg_nbest_model = kwargs.get("avg_nbest_model", 10)
@@ -357,10 +362,10 @@
        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
            # 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()
@@ -376,11 +381,11 @@
                with maybe_autocast(self.use_fp16):
                    retval = model(**batch)

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

                loss, stats, weight = retval
                stats = {k: v for k, v in stats.items() if v is not None}
@@ -410,24 +415,14 @@
                speed_stats["backward_and_AllReaduce_time"] = f"{time4 - time3:0.3f}"

                self.train_loss_avg = (
                    self.train_loss_avg * (self.step_in_epoch - 1) + loss.detach().cpu().item()
                ) / self.step_in_epoch
                    self.train_loss_avg * (batch_idx + kwargs.get("start_step", 0))
                    + loss.detach().cpu().item()
                ) / (batch_idx + kwargs.get("start_step", 0) + 1)
                if "acc" in stats:
                    self.train_acc_avg = (
                        self.train_acc_avg * (self.step_in_epoch - 1)
                        self.train_acc_avg * (batch_idx + kwargs.get("start_step", 0))
                        + stats["acc"].detach().cpu().item()
                    ) / self.step_in_epoch
                if self.use_ddp or self.use_fsdp:
                    train_loss_avg = torch.tensor(self.train_loss_avg, dtype=torch.float32).to(
                        self.device
                    )
                    train_acc_avg = torch.tensor(self.train_acc_avg, dtype=torch.float32).to(
                        self.device
                    )
                    dist.all_reduce(train_loss_avg, op=dist.ReduceOp.SUM)
                    dist.all_reduce(train_acc_avg, op=dist.ReduceOp.SUM)
                    self.train_loss_avg = train_loss_avg.detach().cpu().item() / self.world_size
                    self.train_acc_avg = train_acc_avg.detach().cpu().item() / self.world_size
                    ) / (batch_idx + kwargs.get("start_step", 0) + 1)

            # Perform an optimizer step only after accumulating enough gradients
            if (batch_idx + 1) % accum_grad == 0:
@@ -456,6 +451,19 @@
                scheduler.step()
                # Clear gradients for the next accumulation stage
                optim.zero_grad(set_to_none=True)

                if self.use_ddp or self.use_fsdp:
                    train_loss_avg = torch.tensor(self.train_loss_avg, dtype=torch.float32).to(
                        self.device
                    )
                    train_acc_avg = torch.tensor(self.train_acc_avg, dtype=torch.float32).to(
                        self.device
                    )
                    dist.all_reduce(train_loss_avg, op=dist.ReduceOp.SUM)
                    dist.all_reduce(train_acc_avg, op=dist.ReduceOp.SUM)
                    self.train_loss_avg = train_loss_avg.detach().cpu().item() / self.world_size
                    self.train_acc_avg = train_acc_avg.detach().cpu().item() / self.world_size

                total_time = f"{(time.perf_counter() - time5)/accum_grad:0.3f}"
                time5 = time.perf_counter()

@@ -473,7 +481,7 @@
                    step_in_epoch=self.step_in_epoch,
                    batch_num_epoch=batch_num_epoch,
                    lr=lr,
                    loss=loss.detach().cpu().item(),
                    loss=accum_grad * loss.detach().cpu().item(),
                    speed_stats=speed_stats,
                    stats=stats,
                    writer=writer,
@@ -508,14 +516,14 @@
                )

            time_beg = time.perf_counter()
        else:
            if self.use_ddp or self.use_fsdp:
                iterator_stop.fill_(1)
                dist.all_reduce(iterator_stop, dist.ReduceOp.SUM)
        # else:
        #     if self.use_ddp or self.use_fsdp:
        #         iterator_stop.fill_(1)
        #         dist.all_reduce(iterator_stop, dist.ReduceOp.SUM)

        if self.use_ddp or self.use_fsdp:
            dist.barrier()
            iterator_stop = torch.tensor(0).to(self.device)
            # iterator_stop = torch.tensor(0).to(self.device)

    def validate_epoch(
        self,