Add LossRecorder and use moving average in all places

fine_tune.py

@@ -295,7 +295,7 @@ def train(args):
         for m in training_models:
             m.train()
 
-        loss_total = 0
+        loss_recorder = train_util.LossRecorder()
         for step, batch in enumerate(train_dataloader):
             current_step.value = global_step
             with accelerator.accumulate(training_models[0]):  # 複数モデルに対応していない模様だがとりあえずこうしておく
@@ -405,9 +405,8 @@ def train(args):
                     )
                 accelerator.log(logs, step=global_step)
 
-            # TODO moving averageにする
+            loss_recorder.add(epoch=epoch, step=step, loss=current_loss)
-            loss_total += current_loss
+            avr_loss: float = loss_recorder.get_moving_average()
-            avr_loss = loss_total / (step + 1)
             logs = {"loss": avr_loss}  # , "lr": lr_scheduler.get_last_lr()[0]}
             progress_bar.set_postfix(**logs)
 
@@ -415,7 +414,7 @@ def train(args):
                 break
 
         if args.logging_dir is not None:
-            logs = {"loss/epoch": loss_total / len(train_dataloader)}
+            logs = {"loss/epoch": loss_recorder.get_moving_average()}
             accelerator.log(logs, step=epoch + 1)
 
         accelerator.wait_for_everyone()

library/train_util.py

@@ -4685,3 +4685,20 @@ class collator_class:
         dataset.set_current_epoch(self.current_epoch.value)
         dataset.set_current_step(self.current_step.value)
         return examples[0]
+
+
+class LossRecorder:
+    def __init__(self):
+        self.loss_list: List[float] = []
+        self.loss_total: float = 0.0
+
+    def add(self, *, epoch:int, step: int, loss: float) -> None:
+        if epoch == 0:
+            self.loss_list.append(loss)
+        else:
+            self.loss_total -= self.loss_list[step]
+            self.loss_list[step] = loss
+        self.loss_total += loss
+
+    def get_moving_average(self) -> float:
+        return self.loss_total / len(self.loss_list)

sdxl_train.py

@@ -459,7 +459,7 @@ def train(args):
         for m in training_models:
             m.train()
 
-        loss_total = 0
+        loss_recorder = train_util.LossRecorder()
         for step, batch in enumerate(train_dataloader):
             current_step.value = global_step
             with accelerator.accumulate(training_models[0]):  # 複数モデルに対応していない模様だがとりあえずこうしておく
@@ -632,9 +632,8 @@ def train(args):
 
                 accelerator.log(logs, step=global_step)
 
-            # TODO moving averageにする
+            loss_recorder.add(epoch=epoch, step=step, loss=current_loss)
-            loss_total += current_loss
+            avr_loss: float = loss_recorder.get_moving_average()
-            avr_loss = loss_total / (step + 1)
             logs = {"loss": avr_loss}  # , "lr": lr_scheduler.get_last_lr()[0]}
             progress_bar.set_postfix(**logs)
 
@@ -642,7 +641,7 @@ def train(args):
                 break
 
         if args.logging_dir is not None:
-            logs = {"loss/epoch": loss_total / len(train_dataloader)}
+            logs = {"loss/epoch": loss_recorder.get_moving_average()}
             accelerator.log(logs, step=epoch + 1)
 
         accelerator.wait_for_everyone()

train_db.py

@@ -264,8 +264,7 @@ def train(args):
             init_kwargs = toml.load(args.log_tracker_config)
         accelerator.init_trackers("dreambooth" if args.log_tracker_name is None else args.log_tracker_name, init_kwargs=init_kwargs)
 
-    loss_list = []
+    loss_recorder = train_util.LossRecorder()
-    loss_total = 0.0
     for epoch in range(num_train_epochs):
         accelerator.print(f"\nepoch {epoch+1}/{num_train_epochs}")
         current_epoch.value = epoch + 1
@@ -392,13 +391,8 @@ def train(args):
                     )
                 accelerator.log(logs, step=global_step)
 
-            if epoch == 0:
+            loss_recorder.add(epoch=epoch, step=step, loss=current_loss)
-                loss_list.append(current_loss)
+            avr_loss: float = loss_recorder.get_moving_average()
-            else:
-                loss_total -= loss_list[step]
-                loss_list[step] = current_loss
-            loss_total += current_loss
-            avr_loss = loss_total / len(loss_list)
             logs = {"loss": avr_loss}  # , "lr": lr_scheduler.get_last_lr()[0]}
             progress_bar.set_postfix(**logs)
 
@@ -406,7 +400,7 @@ def train(args):
                 break
 
         if args.logging_dir is not None:
-            logs = {"loss/epoch": loss_total / len(loss_list)}
+            logs = {"loss/epoch": loss_recorder.get_moving_average()}
             accelerator.log(logs, step=epoch + 1)
 
         accelerator.wait_for_everyone()

train_network.py

@@ -703,8 +703,7 @@ class NetworkTrainer:
                 "network_train" if args.log_tracker_name is None else args.log_tracker_name, init_kwargs=init_kwargs
             )
 
-        loss_list = []
+        loss_recorder = train_util.LossRecorder()
-        loss_total = 0.0
         del train_dataset_group
 
         # callback for step start
@@ -854,13 +853,8 @@ class NetworkTrainer:
                                 remove_model(remove_ckpt_name)
 
                 current_loss = loss.detach().item()
-                if epoch == 0:
+                loss_recorder.add(epoch=epoch, step=step, loss=current_loss)
-                    loss_list.append(current_loss)
+                avr_loss: float = loss_recorder.get_moving_average()
-                else:
-                    loss_total -= loss_list[step]
-                    loss_list[step] = current_loss
-                loss_total += current_loss
-                avr_loss = loss_total / len(loss_list)
                 logs = {"loss": avr_loss}  # , "lr": lr_scheduler.get_last_lr()[0]}
                 progress_bar.set_postfix(**logs)
 
@@ -875,7 +869,7 @@ class NetworkTrainer:
                     break
 
             if args.logging_dir is not None:
-                logs = {"loss/epoch": loss_total / len(loss_list)}
+                logs = {"loss/epoch": loss_recorder.get_moving_average()}
                 accelerator.log(logs, step=epoch + 1)
 
             accelerator.wait_for_everyone()