Dev gzf (#1377)

zhifu gao

2024-02-21 cdca62d933c4e0766a05044c6cba7cfa0596e615

 funasr/train_utils/trainer.py

@@ -3,6 +3,7 @@
import torch
import logging
from tqdm import tqdm
from datetime import datetime
import torch.distributed as dist
from contextlib import nullcontext
# from torch.utils.tensorboard import SummaryWriter
@@ -107,7 +108,7 @@
        filename = os.path.join(self.output_dir, f'model.pt.ep{epoch}')
        torch.save(state, filename)
        
        print(f'Checkpoint saved to {filename}')
        print(f'\nCheckpoint saved to {filename}\n')
        latest = Path(os.path.join(self.output_dir, f'model.pt'))
        torch.save(state, latest)

@@ -156,7 +157,7 @@
            self._resume_checkpoint(self.output_dir)
        
        for epoch in range(self.start_epoch, self.max_epoch + 1):
            
            time1 = time.perf_counter()
            self._train_epoch(epoch)


@@ -178,6 +179,9 @@
            
            self.scheduler.step()

            time2 = time.perf_counter()
            time_escaped = (time2 - time1)/3600.0
            print(f"\nrank: {self.local_rank}, time_escaped_epoch: {time_escaped:.3f} hours, estimated to finish {self.max_epoch} epoch: {(self.max_epoch-epoch)*time_escaped:.3f}\n")

        if self.rank == 0:
            average_checkpoints(self.output_dir, self.avg_nbest_model)
@@ -283,7 +287,10 @@
                                             torch.cuda.max_memory_reserved()/1024/1024/1024,
                                             )
                lr = self.scheduler.get_last_lr()[0]
                time_now = datetime.now()
                time_now = time_now.strftime("%Y-%m-%d %H:%M:%S")
                description = (
                    f"{time_now}, "
                    f"rank: {self.local_rank}, "
                    f"epoch: {epoch}/{self.max_epoch}, "
                    f"step: {batch_idx+1}/{len(self.dataloader_train)}, total: {self.batch_total}, "
@@ -295,17 +302,14 @@
                )
                pbar.set_description(description)
                if self.writer:
                    self.writer.add_scalar(f'rank{self.local_rank}_Loss/train', loss.item(),
                                           epoch*len(self.dataloader_train) + batch_idx)
                    self.writer.add_scalar(f'rank{self.local_rank}_Loss/train', loss.item(), self.batch_total)
                    self.writer.add_scalar(f'rank{self.local_rank}_lr/train', lr, self.batch_total)
                    for key, var in stats.items():
                        self.writer.add_scalar(f'rank{self.local_rank}_{key}/train', var.item(),
                                               epoch * len(self.dataloader_train) + batch_idx)
                        self.writer.add_scalar(f'rank{self.local_rank}_{key}/train', var.item(), self.batch_total)
                    for key, var in speed_stats.items():
                        self.writer.add_scalar(f'rank{self.local_rank}_{key}/train', eval(var),
                                               epoch * len(self.dataloader_train) + batch_idx)
                    
            # if batch_idx == 2:
            #     break
                        self.writer.add_scalar(f'rank{self.local_rank}_{key}/train', eval(var), self.batch_total)


        pbar.close()

    def _validate_epoch(self, epoch):
@@ -349,7 +353,10 @@
                
                if (batch_idx+1) % self.log_interval == 0 or (batch_idx+1) == len(self.dataloader_val):
                    pbar.update(self.log_interval)
                    time_now = datetime.now()
                    time_now = time_now.strftime("%Y-%m-%d %H:%M:%S")
                    description = (
                        f"{time_now}, "
                        f"rank: {self.local_rank}, "
                        f"validation epoch: {epoch}/{self.max_epoch}, "
                        f"step: {batch_idx+1}/{len(self.dataloader_val)}, "