From d80ac2fd2df4e7fb8a28acfa512bb11472b5cc99 Mon Sep 17 00:00:00 2001
From: liugz18 <57401541+liugz18@users.noreply.github.com>
Date: 星期四, 18 七月 2024 21:34:55 +0800
Subject: [PATCH] Rename 'res' in line 514 to avoid with naming conflict with line 365

---
 funasr/train_utils/trainer.py |   94 +++++++++++++++++++++++++++--------------------
 1 files changed, 54 insertions(+), 40 deletions(-)

diff --git a/funasr/train_utils/trainer.py b/funasr/train_utils/trainer.py
index dd0ac7a..665a7af 100644
--- a/funasr/train_utils/trainer.py
+++ b/funasr/train_utils/trainer.py
@@ -85,7 +85,12 @@
         self.batch_total = 0
         self.use_fp16 = use_fp16
         self.save_checkpoint_interval = kwargs.get("save_checkpoint_interval", 5000)
-        self.validate_interval = kwargs.get("validate_interval", 5000)
+        self.validate_interval = kwargs.get("validate_interval", -1)
+        if self.validate_interval < 0:
+            self.validate_interval = self.save_checkpoint_interval
+        assert (
+            self.save_checkpoint_interval == self.validate_interval
+        ), f"save_checkpoint_interval must equal to validate_interval"
         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)
@@ -308,6 +313,7 @@
                     checkpoint["step_in_epoch"] if "step_in_epoch" in checkpoint else 0
                 )
                 self.step_in_epoch = 0 if self.step_in_epoch is None else self.step_in_epoch
+                print(checkpoint["train_acc_avg"])
                 self.train_acc_avg = (
                     checkpoint["train_acc_avg"] if "train_acc_avg" in checkpoint else 0
                 )
@@ -356,10 +362,10 @@
         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)
-                if iterator_stop > 0:
-                    break
+            # 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
             self.step_in_epoch += 1
             time1 = time.perf_counter()
@@ -375,14 +381,12 @@
                 with maybe_autocast(self.use_fp16):
                     retval = model(**batch)
 
-                    if (
-                        self.reset_gpu_cache
-                        and (torch.cuda.max_memory_reserved() / 1024 / 1024 / 1024) > 70
-                    ):
-                        torch.cuda.empty_cache()
+                    # if (
+                    #     self.reset_gpu_cache
+                    #     and (torch.cuda.max_memory_reserved() / 1024 / 1024 / 1024) > 70
+                    # ):
+                    #     torch.cuda.empty_cache()
 
-                time3 = time.perf_counter()
-                speed_stats["forward_time"] = f"{time3 - time2:0.3f}"
                 loss, stats, weight = retval
                 stats = {k: v for k, v in stats.items() if v is not None}
                 if self.use_ddp or self.use_fsdp:
@@ -397,34 +401,28 @@
                     # Multiply world_size because DistributedDataParallel
                     # automatically normalizes the gradient by world_size.
                     loss *= self.world_size
+                # loss *= self.world_size
                 # Scale the loss since we're not updating for every mini-batch
                 loss = loss / accum_grad
+
+                time3 = time.perf_counter()
+                speed_stats["forward_time"] = f"{time3 - time2:0.3f}"
                 if self.use_fp16:
                     scaler.scale(loss).backward()
                 else:
                     loss.backward()
                 time4 = time.perf_counter()
-                speed_stats["backward_time"] = f"{time4 - time3:0.3f}"
+                speed_stats["backward_and_AllReaduce_time"] = f"{time4 - time3:0.3f}"
 
                 self.train_loss_avg = (
-                    self.train_loss_avg * (self.step_in_epoch - 1) + loss.detach().cpu().item()
-                ) / self.step_in_epoch
+                    self.train_loss_avg * (batch_idx + kwargs.get("start_step", 0))
+                    + loss.detach().cpu().item()
+                ) / (batch_idx + kwargs.get("start_step", 0) + 1)
                 if "acc" in stats:
                     self.train_acc_avg = (
-                        self.train_acc_avg * (self.step_in_epoch - 1)
+                        self.train_acc_avg * (batch_idx + kwargs.get("start_step", 0))
                         + stats["acc"].detach().cpu().item()
-                    ) / self.step_in_epoch
-                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
+                    ) / (batch_idx + kwargs.get("start_step", 0) + 1)
 
             # Perform an optimizer step only after accumulating enough gradients
             if (batch_idx + 1) % accum_grad == 0:
@@ -453,8 +451,22 @@
                 scheduler.step()
                 # Clear gradients for the next accumulation stage
                 optim.zero_grad(set_to_none=True)
-                total_time = f"{time.perf_counter() - time5:0.3f}"
+
+                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
+
+                total_time = f"{(time.perf_counter() - time5)/accum_grad:0.3f}"
                 time5 = time.perf_counter()
+
                 speed_stats["optim_time"] = f"{time5 - time4:0.3f}"
 
                 speed_stats["total_time"] = total_time
@@ -464,11 +476,12 @@
                     batch_num_epoch = len(dataloader_train)
                 self.log(
                     epoch,
-                    batch_idx + kwargs.get("start_step", 0),
+                    batch_idx,
+                    log_step=batch_idx + kwargs.get("start_step", 0),
                     step_in_epoch=self.step_in_epoch,
                     batch_num_epoch=batch_num_epoch,
                     lr=lr,
-                    loss=loss.detach().cpu().item(),
+                    loss=accum_grad * loss.detach().cpu().item(),
                     speed_stats=speed_stats,
                     stats=stats,
                     writer=writer,
@@ -503,14 +516,14 @@
                 )
 
             time_beg = time.perf_counter()
-        else:
-            if self.use_ddp or self.use_fsdp:
-                iterator_stop.fill_(1)
-                dist.all_reduce(iterator_stop, dist.ReduceOp.SUM)
+        # 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()
-            iterator_stop = torch.tensor(0).to(self.device)
+            # iterator_stop = torch.tensor(0).to(self.device)
 
     def validate_epoch(
         self,
@@ -633,11 +646,12 @@
         tag="train",
         data_split_i=0,
         data_split_num=1,
+        log_step=None,
         **kwargs,
     ):
 
         if (batch_idx + 1) % self.log_interval == 0:
-
+            batch_idx = log_step if log_step is not None else batch_idx
             gpu_info = (
                 "GPU, memory: usage: {:.3f} GB, "
                 "peak: {:.3f} GB, "
@@ -659,9 +673,9 @@
                 f"data_slice: {data_split_i}/{data_split_num}, "
                 f"step_in_slice: {batch_idx + 1}/{batch_num_epoch}, step_in_epoch: {step_in_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):.3e}), "
-                f"(acc_avg_epoch: {acc_avg_epoch:.3f}), "
+                f"(loss_avg_slice: {loss_avg_epoch:.3f}), "
+                f"(ppl_avg_slice: {math.exp(loss_avg_epoch):.3e}), "
+                f"(acc_avg_slice: {acc_avg_epoch:.3f}), "
                 f"(lr: {lr:.3e}), "
                 f"{[(k, round(v.detach().cpu().item(), 3)) for k, v in stats.items()]}, "
                 f"{speed_stats}, "

--
Gitblit v1.9.1