Dev gzf exp (#1647)

zhifu gao

2024-04-23 2ac38adbe5f4e1374a079e032ed4b504351a207c

 funasr/train_utils/trainer.py

@@ -71,21 +71,19 @@
        self.use_ddp = use_ddp
        self.use_fsdp = use_fsdp
        self.device = kwargs.get('device', "cuda")
        self.avg_nbest_model = kwargs.get("avg_nbest_model", 5)
        # self.kwargs = kwargs
        self.log_interval = kwargs.get("log_interval", 50)
        self.batch_total = 0
        self.use_fp16 = use_fp16
        self.disable_gpu_cache = kwargs.get("disable_gpu_cache", True)
        # scaler = GradScaler(enabled=use_fp16) if use_fp16 else None
        # scaler = ShardedGradScaler(enabled=use_fp16) if use_fsdp else scaler
        # self.scaler = scaler
        self.save_checkpoint_interval = kwargs.get("save_checkpoint_interval", 5000)
        self.keep_nbest_models = kwargs.get("keep_nbest_models", -1)
        self.validate_interval = kwargs.get("validate_interval", 5000)
        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)
        self.accum_grad = kwargs.get("accum_grad", 1)
        self.grad_clip = kwargs.get("grad_clip", 10.0)
        self.grad_clip_type = kwargs.get("grad_clip_type", 2.0)
        self.validate_interval = kwargs.get("validate_interval", 5000)
        
        
    
        try:
@@ -103,8 +101,10 @@
        self.val_loss_avg = 0.0
        self.best_acc_idx = 0
        self.saved_ckpts = {}
        self.val_acc_list = []
        self.step_or_epoch = -1
        self.best_step_or_epoch = ""
        self.val_acc_step_or_eoch = {}
        self.val_loss_step_or_eoch = {}
       
    def save_checkpoint(self, epoch,
                        step=None,
@@ -124,14 +124,17 @@
        
        if self.rank == 0:
            logging.info(f"Save checkpoint: {epoch}, rank: {self.local_rank}\n")
            self.step_or_epoch += 1
            # self.step_or_epoch += 1
            state = {
                'epoch': epoch,
                'state_dict': model.state_dict(),
                'optimizer': optim.state_dict(),
                'scheduler': scheduler.state_dict(),
                "acc": self.val_acc_list,
                "step_or_epoch": self.step_or_epoch,
                "saved_ckpts": self.saved_ckpts,
                "val_acc_step_or_eoch": self.val_acc_step_or_eoch,
                "val_loss_step_or_eoch": self.val_loss_step_or_eoch,
                "best_step_or_epoch": self.best_step_or_epoch,
                "avg_keep_nbest_models_type": self.avg_keep_nbest_models_type,
            }
            if hasattr(model, "module"):
                state["state_dict"] = model.module.state_dict()
@@ -150,23 +153,37 @@
            logging.info(f'\nCheckpoint saved to {filename}\n')
            latest = Path(os.path.join(self.output_dir, f'model.pt'))
            torch.save(state, latest)
            
            if self.val_acc_list[self.step_or_epoch] >= self.val_acc_list[self.best_acc_idx]:
                self.best_acc_idx = self.step_or_epoch
                best_ckpt = Path(os.path.join(self.output_dir, f'model.pt.best'))
                torch.save(state, best_ckpt)
                logging.info(f"Update best acc: {self.val_acc_list[self.best_acc_idx]}, {best_ckpt}")
            if self.best_step_or_epoch == "":
                self.best_step_or_epoch = ckpt_name
             
            if self.avg_keep_nbest_models_type == "acc":
                if self.val_acc_step_or_eoch[ckpt_name] >= self.val_acc_step_or_eoch[self.best_step_or_epoch]:
                    self.best_step_or_epoch = ckpt_name
                    best_ckpt = Path(os.path.join(self.output_dir, f'model.pt.best'))
                    torch.save(state, best_ckpt)
                    logging.info(f"Update best acc: {self.val_acc_step_or_eoch[self.best_step_or_epoch]:.4f}, {best_ckpt}")
                else:
                    logging.info(f"No improvement in acc: {self.val_acc_step_or_eoch[ckpt_name]:.4f} < {self.val_acc_step_or_eoch[self.best_step_or_epoch]:.4f}")
            elif self.avg_keep_nbest_models_type == "loss":
                if self.val_loss_step_or_eoch[ckpt_name] <= self.val_loss_step_or_eoch[self.best_step_or_epoch]:
                    self.best_step_or_epoch = ckpt_name
                    best_ckpt = Path(os.path.join(self.output_dir, f'model.pt.best'))
                    torch.save(state, best_ckpt)
                    logging.info(f"Update best loss: {self.val_loss_step_or_eoch[self.best_step_or_epoch]:.4f}, {best_ckpt}")
                else:
                    logging.info(f"No improvement in loss: {self.val_loss_step_or_eoch[ckpt_name]:.4f} > {self.val_loss_step_or_eoch[self.best_step_or_epoch]:.4f}")
            else:
                logging.info(f"No improvement in acc: {self.val_acc_list[self.best_acc_idx]}")
            
                print("Undo")
            self.saved_ckpts[ckpt_name] = getattr(self, f"val_{self.avg_keep_nbest_models_type}_step_or_eoch")[ckpt_name]
            if self.keep_nbest_models > 0:
                self.saved_ckpts[ckpt_name] = self.val_acc_list[-1]
                if len(self.saved_ckpts) > self.keep_nbest_models:

                    min_key = min(self.saved_ckpts, key=self.saved_ckpts.get)
                    if min_key in self.saved_ckpts:
                        del self.saved_ckpts[min_key]
                    filename = os.path.join(self.output_dir, min_key)
                    if self.avg_keep_nbest_models_type == "acc":
                        key = min(self.saved_ckpts, key=self.saved_ckpts.get)
                    else:
                        key = max(self.saved_ckpts, key=self.saved_ckpts.get)
                    if key in self.saved_ckpts:
                        del self.saved_ckpts[key]
                    filename = os.path.join(self.output_dir, key)
                    logging.info(f"Delete: {filename}")
                    if os.path.exists(filename):
                        os.remove(filename)
@@ -190,7 +207,7 @@
        if self.resume:
            ckpt = os.path.join(self.output_dir, "model.pt")
            if os.path.isfile(ckpt):
                checkpoint = torch.load(ckpt)
                checkpoint = torch.load(ckpt, map_location="cpu")
                self.start_epoch = checkpoint['epoch'] + 1
                # self.model.load_state_dict(checkpoint['state_dict'])
                src_state = checkpoint['state_dict']
@@ -213,9 +230,11 @@
                if scaler is not None and 'scaler_state' in checkpoint:
                    scaler.load_state_dict(checkpoint['scaler_state'])
                
                self.val_acc_list = checkpoint["acc"]
                self.step_or_epoch = checkpoint["step_or_epoch"]
                
                self.saved_ckpts = checkpoint["saved_ckpts"]
                self.val_acc_step_or_eoch = checkpoint["val_acc_step_or_eoch"] if "val_acc_step_or_eoch" in checkpoint else {}
                self.val_loss_step_or_eoch = checkpoint["val_loss_step_or_eoch"] if "val_loss_step_or_eoch" in checkpoint else {}
                self.best_step_or_epoch = checkpoint["best_step_or_epoch"] if "best_step_or_epoch" in checkpoint else ""
                model.to(self.device)
                print(f"Checkpoint loaded successfully from '{ckpt}'")
            else:
                print(f"No checkpoint found at '{ckpt}', does not resume status!")
@@ -233,12 +252,15 @@
                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.local_rank}\n")
        model.train()

@@ -247,11 +269,12 @@
        # Initialize the gradient accumulation
        optim.zero_grad()
        speed_stats = {}
        time5 = time.perf_counter()
        
        iterator_stop = torch.tensor(0).to(self.device)
        dist.barrier()
        print(f"before iter, iterator_stop: {iterator_stop}\n")

        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)
@@ -259,11 +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):
@@ -297,13 +322,13 @@
                self.train_loss_avg = (self.train_loss_avg*batch_idx + loss.detach().cpu().item())/(batch_idx+1)
                if "acc" in stats:
                    self.train_acc_avg = (self.train_acc_avg * batch_idx + stats["acc"].detach().cpu().item()) / (batch_idx + 1)
                # 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
                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
                
            
            # Perform an optimizer step only after accumulating enough gradients
@@ -339,7 +364,7 @@
    
                speed_stats["total_time"] = total_time
                lr = scheduler.get_last_lr()[0]
                batch_num_epoch = -1
                batch_num_epoch = 1
                if hasattr(dataloader_train, "__len__"):
                    batch_num_epoch = len(dataloader_train)
                self.log(epoch, batch_idx,
@@ -350,6 +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:
@@ -357,12 +384,14 @@
                    model=model,
                    dataloader_val=dataloader_val,
                    epoch=epoch,
                    writer=writer
                    writer=writer,
                    step=batch_idx+1,
                )

            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)
@@ -370,6 +399,7 @@
                
        if self.use_ddp or self.use_fsdp:
            dist.barrier()
            iterator_stop = torch.tensor(0).to(self.device)
        
        

@@ -387,6 +417,8 @@
        Args:
            epoch (int): The current epoch number.
        """
        if self.use_ddp or self.use_fsdp:
            dist.barrier()
        logging.info(f"Validate epoch: {epoch}, rank: {self.local_rank}\n")
        model.eval()
        
@@ -395,13 +427,10 @@
            speed_stats = {}
            time5 = time.perf_counter()
            iterator_stop = torch.tensor(0).to(self.device)
            dist.barrier()
            print(f"before iter, iterator_stop: {iterator_stop}\n")
            dataloader_val.batch_sampler.set_epoch(epoch)
            for batch_idx, batch in enumerate(dataloader_val):
                if self.use_ddp or self.use_fsdp:
                    dist.all_reduce(iterator_stop, dist.ReduceOp.SUM)
                    if epoch >= 1:
                        print(f"iterator_stop: {iterator_stop}\n")
                    if iterator_stop > 0:
                        break
                time1 = time.perf_counter()
@@ -417,7 +446,7 @@
                    # Apply weighted averaging for loss and stats
                    loss = (loss * weight.type(loss.dtype)).sum()
                    # if distributed, this method can also apply all_reduce()
                    stats, weight = recursive_average(stats, weight, distributed=True)
                    # stats, weight = recursive_average(stats, weight, distributed=True)
                    if self.use_ddp or self.use_fsdp:
                        dist.all_reduce(weight, op=dist.ReduceOp.SUM)
                    # Now weight is summation over all workers
@@ -432,15 +461,15 @@
                self.val_loss_avg = (self.val_loss_avg*batch_idx + loss.detach().cpu().item())/(batch_idx+1)
                if "acc" in stats:
                    self.val_acc_avg = (self.val_acc_avg * batch_idx + stats["acc"].detach().cpu().item()) / (batch_idx + 1)
                # if self.use_ddp or self.use_fsdp:
                #     val_loss_avg = torch.tensor(self.val_loss_avg, dtype=torch.float32).to(self.device)
                #     val_acc_avg = torch.tensor(self.val_acc_avg, dtype=torch.float32).to(self.device)
                #     dist.all_reduce(val_loss_avg, op=dist.ReduceOp.SUM)
                #     dist.all_reduce(val_acc_avg, op=dist.ReduceOp.SUM)
                #     self.val_loss_avg = val_loss_avg.detach().cpu().item() / self.world_size
                #     self.val_acc_avg = val_acc_avg.detach().cpu().item() / self.world_size
                
                batch_num_epoch = -1
                if self.use_ddp or self.use_fsdp:
                    val_loss_avg = torch.tensor(self.val_loss_avg, dtype=torch.float32).to(self.device)
                    val_acc_avg = torch.tensor(self.val_acc_avg, dtype=torch.float32).to(self.device)
                    dist.all_reduce(val_loss_avg, op=dist.ReduceOp.SUM)
                    dist.all_reduce(val_acc_avg, op=dist.ReduceOp.SUM)
                    self.val_loss_avg = val_loss_avg.detach().cpu().item() / self.world_size
                    self.val_acc_avg = val_acc_avg.detach().cpu().item() / self.world_size
                time5 = time.perf_counter()
                batch_num_epoch = 1
                if hasattr(dataloader_val, "__len__"):
                    batch_num_epoch = len(dataloader_val)
                self.log(epoch, batch_idx,
@@ -457,14 +486,18 @@
                if self.use_ddp or self.use_fsdp:
                    iterator_stop.fill_(1)
                    dist.all_reduce(iterator_stop, dist.ReduceOp.SUM)
                    
        self.val_acc_list.append(self.val_acc_avg)

        if kwargs.get("step", None) is None:
            ckpt_name = f'model.pt.ep{epoch}'
        else:
            ckpt_name = f'model.pt.ep{epoch}.{kwargs.get("step")}'
        self.val_acc_step_or_eoch[ckpt_name] = self.val_acc_avg
        self.val_loss_step_or_eoch[ckpt_name] = self.val_loss_avg
        model.train()



        if self.use_ddp or self.use_fsdp:
            dist.barrier()
            iterator_stop = torch.tensor(0).to(self.device)
        
        
    def log(self,
@@ -477,6 +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:
@@ -496,10 +532,11 @@
                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}), "
                f"(ppl_avg_epoch: {math.exp(loss_avg_epoch):.3f}), "
                f"(ppl_avg_epoch: {math.exp(loss_avg_epoch):.3e}), "
                f"(acc_avg_epoch: {acc_avg_epoch:.3f}), "
                f"(lr: {lr:.3e}), "
                f"{[(k, round(v.detach().cpu().item(), 3)) for k, v in stats.items()]}, "