wechat

游雁

2024-04-23 22d71ff774d409f8b413f9955a7c0efef5b3e288

 funasr/train_utils/trainer.py

@@ -161,17 +161,17 @@
                    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]}, {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]} < {self.val_acc_step_or_eoch[self.best_step_or_epoch]}")
                    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]}, {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]} > {self.val_loss_step_or_eoch[self.best_step_or_epoch]}")
                    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:
                print("Undo")
            self.saved_ckpts[ckpt_name] = getattr(self, f"val_{self.avg_keep_nbest_models_type}_step_or_eoch")[ckpt_name]
@@ -268,10 +268,12 @@
        # 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 +281,13 @@
                    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):
@@ -377,12 +381,14 @@
                    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)
@@ -523,7 +529,7 @@
                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()]}, "