Fix a few issues found during fine-tuning (#2582)

ming030890

2025-07-04 a750595594321833b48dc19798eed66876a100b4

Fix a few issues found during fine-tuning (#2582)

* Fix wandb log

* fix validation loss is not logged

batch_idx got reset for each epoch.
use the global step counter instead

* LR should only be updated per step, not per step+ per epoch

* add early stopping

* Fix bf16 handling

scaler is only needed for fp16

* more logs

---------

Co-authored-by: Tony Mak <tony@Tonys-MacBook-Air-1800.local>

 funasr/bin/train_ds.py

@@ -149,7 +149,7 @@
    dataloader = dataloader_class(**kwargs)
    # dataloader_tr, dataloader_val = dataloader_class(**kwargs)

    scaler = GradScaler(enabled=True) if trainer.use_fp16 or trainer.use_bf16 else None
    scaler = GradScaler(enabled=True) if trainer.use_fp16 else None
    scaler = ShardedGradScaler(enabled=trainer.use_fp16) if trainer.use_fsdp else scaler

    trainer.resume_checkpoint(
@@ -158,6 +158,10 @@
        scheduler=scheduler,
        scaler=scaler,
    )

    early_stopping_patience = kwargs.get("train_conf", {}).get("early_stopping_patience", 0)
    best_val_loss = float("inf")
    epochs_no_improve = 0

    dataloader_tr, dataloader_val = None, None
    for epoch in range(trainer.start_epoch, trainer.max_epoch):
@@ -199,7 +203,19 @@

        trainer.start_data_split_i = 0
        trainer.validate_epoch(model=model, dataloader_val=dataloader_val, epoch=epoch + 1)
        scheduler.step()
        current_val = trainer.val_loss_avg

        if current_val < best_val_loss:
            logging.info(f"current_val: {current_val}, best_val_loss: {best_val_loss}")
            best_val_loss = current_val
            epochs_no_improve = 0
        else:
            epochs_no_improve += 1
            logging.info(f"No val_loss improvement for {epochs_no_improve}/{early_stopping_patience} epochs")
        if early_stopping_patience > 0 and epochs_no_improve >= early_stopping_patience:
            logging.info(f"Early stopping triggered at epoch {epoch+1}")
            break

        trainer.step_in_epoch = 0
        trainer.save_checkpoint(
            epoch + 1, model=model, optim=optim, scheduler=scheduler, scaler=scaler

 funasr/train_utils/trainer.py

@@ -715,7 +715,7 @@
            if self.use_wandb and wandb is not None:
                wandb.log(
                    description_dict,
                    setp=self.batch_total,
                    step=self.batch_total,
                )

    def close(self, writer=None):

 funasr/train_utils/trainer_ds.py

@@ -30,9 +30,8 @@
            yield
    else:
        if dtype == torch.float16 or dtype == torch.bfloat16:
            yield
            # with autocast(enabled=True, dtype=dtype):
            #     yield
            with autocast(enabled=True, dtype=dtype):
                yield
        else:
            yield

@@ -684,7 +683,7 @@
            scaled_loss = model.backward(loss)
        else:
            loss = loss / self.accum_grad
            if self.use_fp16 or self.use_bf16:
            if scaler:
                scaler.scale(loss).backward()
            else:
                loss.backward()
@@ -712,7 +711,7 @@
                # Execute an optimization step (update model parameters)
                if self.use_ddp or self.use_fsdp:
                    dist.barrier()
                if self.use_fp16 or self.use_bf16:
                if scaler:
                    scaler.step(optim)
                    scaler.update()
                else:
@@ -736,6 +735,9 @@
        Args:
            epoch (int): The current epoch number.
        """
        self.val_loss_avg = 0.0
        self.val_acc_avg  = 0.0

        if self.use_ddp or self.use_fsdp or self.use_deepspeed:
            dist.barrier()
        logging.info(f"Validate epoch: {epoch}, rank: {self.rank}\n")
@@ -757,7 +759,7 @@
                    "data_split_i": kwargs.get("data_split_i", 0),
                    "data_split_num": kwargs.get("data_split_num", 1),
                    "log_step": batch_idx + kwargs.get("start_step", 0),
                    "batch_total": batch_idx + 1,
                    "batch_total": self.batch_total,
                    "step_in_epoch": batch_idx + 1,
                    "lr": 0.0,
                }
@@ -883,7 +885,7 @@
            if self.use_wandb and wandb is not None:
                wandb.log(
                    description_dict,
                    setp=batch_total,
                    step=batch_total,
                )

    def close(self, writer=None):