test

语帆

2024-02-28 eb92e79fb94e7b3df8f27c8ce3e607a70dff2a2e

 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
@@ -107,14 +108,10 @@
        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'))
        try:
            latest.unlink()
        except:
            pass
        torch.save(state, latest)

        latest.symlink_to(filename)
    
    def _resume_checkpoint(self, resume_path):
        """
@@ -160,7 +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)


@@ -182,6 +179,9 @@
            
            self.scheduler.step()

            time2 = time.perf_counter()
            time_escaped = (time2 - time1)/3600.0
            print(f"\ntime_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)
@@ -287,7 +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}, "