From 16a976a01d110d3969759be7720cae2b6b0664f7 Mon Sep 17 00:00:00 2001
From: 游雁 <zhifu.gzf@alibaba-inc.com>
Date: 星期日, 24 三月 2024 01:27:08 +0800
Subject: [PATCH] finetune

---
 funasr/train_utils/trainer.py |   35 +++++++++++++++++++++++++++++++----
 1 files changed, 31 insertions(+), 4 deletions(-)

diff --git a/funasr/train_utils/trainer.py b/funasr/train_utils/trainer.py
index 49b9fbc..cf23483 100644
--- a/funasr/train_utils/trainer.py
+++ b/funasr/train_utils/trainer.py
@@ -239,6 +239,8 @@
         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()
 
@@ -248,8 +250,14 @@
         optim.zero_grad()
         speed_stats = {}
         time5 = time.perf_counter()
-        
+        # iterator_stop = torch.tensor(0).to(self.device)
+
+        dataloader_train.batch_sampler.set_epoch(epoch)
         for batch_idx, batch in enumerate(dataloader_train):
+            # if self.use_ddp or self.use_fsdp:
+            #     dist.all_reduce(iterator_stop, dist.ReduceOp.SUM)
+            #     if iterator_stop > 0:
+            #         break
             self.batch_total += 1
             time1 = time.perf_counter()
             speed_stats["data_load"] = f"{time1-time5:0.3f}"
@@ -356,7 +364,11 @@
             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)
 
-        
+        # else:
+        #     if self.use_ddp or self.use_fsdp:
+        #         iterator_stop.fill_(1)
+        #         dist.all_reduce(iterator_stop, dist.ReduceOp.SUM)
+                
         if self.use_ddp or self.use_fsdp:
             dist.barrier()
         
@@ -376,6 +388,8 @@
         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()
         
@@ -383,7 +397,15 @@
             
             speed_stats = {}
             time5 = time.perf_counter()
+            # iterator_stop = torch.tensor(0).to(self.device)
+
             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()
                 speed_stats["data_load"] = f"{time1 - time5:0.3f}"
                 batch = to_device(batch, self.device)
@@ -397,7 +419,7 @@
                     # 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
@@ -433,9 +455,14 @@
                          tag="val",
                          )
 
+            # else:
+            #     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)
         model.train()
-        
+
         if self.use_ddp or self.use_fsdp:
             dist.barrier()
         

--
Gitblit v1.9.1