Support for more Dadaptation (#455)

* Update train_util.py for add DAdaptAdan and DAdaptSGD

* Update train_util.py for DAdaptadam

* Update train_network.py for dadapt

* Update train_README-ja.md for DAdapt

* Update train_util.py for DAdapt

* Update train_network.py for DAdaptAdaGrad

* Update train_db.py for DAdapt

* Update fine_tune.py for DAdapt

* Update train_textual_inversion.py for DAdapt

* Update train_textual_inversion_XTI.py for DAdapt

fine_tune.py

@@ -381,7 +381,7 @@ def train(args):
             current_loss = loss.detach().item()  # 平均なのでbatch sizeは関係ないはず
             if args.logging_dir is not None:
                 logs = {"loss": current_loss, "lr": float(lr_scheduler.get_last_lr()[0])}
-                if args.optimizer_type.lower() == "DAdaptation".lower():  # tracking d*lr value
+                if args.optimizer_type.lower() == "DAdaptation".lower() or args.optimizer_type.lower() == "DAdaptAdam".lower() or args.optimizer_type.lower() == "DAdaptAdaGrad".lower() or args.optimizer_type.lower() == "DAdaptAdan".lower() or args.optimizer_type.lower() == "DAdaptSGD".lower():  # tracking d*lr value
                     logs["lr/d*lr"] = (
                         lr_scheduler.optimizers[0].param_groups[0]["d"] * lr_scheduler.optimizers[0].param_groups[0]["lr"]
                     )

library/train_util.py

@@ -1885,7 +1885,7 @@ def add_optimizer_arguments(parser: argparse.ArgumentParser):
         "--optimizer_type",
         type=str,
         default="",
-        help="Optimizer to use / オプティマイザの種類: AdamW (default), AdamW8bit, Lion, Lion8bit,SGDNesterov, SGDNesterov8bit, DAdaptation, AdaFactor",
+        help="Optimizer to use / オプティマイザの種類: AdamW (default), AdamW8bit, Lion, SGDNesterov, SGDNesterov8bit, DAdaptation(DAdaptAdam), DAdaptAdaGrad, DAdaptAdan, DAdaptSGD, AdaFactor",
     )
 
     # backward compatibility
@@ -2467,7 +2467,7 @@ def resume_from_local_or_hf_if_specified(accelerator, args):
 
 
 def get_optimizer(args, trainable_params):
-    # "Optimizer to use: AdamW, AdamW8bit, Lion, Lion8bit, SGDNesterov, SGDNesterov8bit, DAdaptation, Adafactor"
+    # "Optimizer to use: AdamW, AdamW8bit, Lion, SGDNesterov, SGDNesterov8bit, DAdaptation, DAdaptation(DAdaptAdam), DAdaptAdaGrad, DAdaptAdan, DAdaptSGD, Adafactor"
 
     optimizer_type = args.optimizer_type
     if args.use_8bit_adam:
@@ -2570,7 +2570,7 @@ def get_optimizer(args, trainable_params):
         optimizer_class = torch.optim.SGD
         optimizer = optimizer_class(trainable_params, lr=lr, nesterov=True, **optimizer_kwargs)
 
-    elif optimizer_type == "DAdaptation".lower():
+    elif optimizer_type == "DAdaptation".lower() or optimizer_type == "DAdaptAdam".lower():
         try:
             import dadaptation
         except ImportError:
@@ -2599,6 +2599,93 @@ def get_optimizer(args, trainable_params):
         optimizer_class = dadaptation.DAdaptAdam
         optimizer = optimizer_class(trainable_params, lr=lr, **optimizer_kwargs)
         
+    elif optimizer_type == "DAdaptAdaGrad".lower():
+        try:
+            import dadaptation
+        except ImportError:
+            raise ImportError("No dadaptation / dadaptation がインストールされていないようです")
+        print(f"use D-Adaptation AdaGrad optimizer | {optimizer_kwargs}")
+
+        actual_lr = lr
+        lr_count = 1
+        if type(trainable_params) == list and type(trainable_params[0]) == dict:
+            lrs = set()
+            actual_lr = trainable_params[0].get("lr", actual_lr)
+            for group in trainable_params:
+                lrs.add(group.get("lr", actual_lr))
+            lr_count = len(lrs)
+
+        if actual_lr <= 0.1:
+            print(
+                f"learning rate is too low. If using dadaptation, set learning rate around 1.0 / 学習率が低すぎるようです。1.0前後の値を指定してください: lr={actual_lr}"
+            )
+            print("recommend option: lr=1.0 / 推奨は1.0です")
+        if lr_count > 1:
+            print(
+                f"when multiple learning rates are specified with dadaptation (e.g. for Text Encoder and U-Net), only the first one will take effect / D-Adaptationで複数の学習率を指定した場合（Text EncoderとU-Netなど）、最初の学習率のみが有効になります: lr={actual_lr}"
+            )
+
+        optimizer_class = dadaptation.DAdaptAdaGrad
+        optimizer = optimizer_class(trainable_params, lr=lr, **optimizer_kwargs)
+    
+    elif optimizer_type == "DAdaptAdan".lower():
+        try:
+            import dadaptation
+        except ImportError:
+            raise ImportError("No dadaptation / dadaptation がインストールされていないようです")
+        print(f"use D-Adaptation Adan optimizer | {optimizer_kwargs}")
+
+        actual_lr = lr
+        lr_count = 1
+        if type(trainable_params) == list and type(trainable_params[0]) == dict:
+            lrs = set()
+            actual_lr = trainable_params[0].get("lr", actual_lr)
+            for group in trainable_params:
+                lrs.add(group.get("lr", actual_lr))
+            lr_count = len(lrs)
+
+        if actual_lr <= 0.1:
+            print(
+                f"learning rate is too low. If using dadaptation, set learning rate around 1.0 / 学習率が低すぎるようです。1.0前後の値を指定してください: lr={actual_lr}"
+            )
+            print("recommend option: lr=1.0 / 推奨は1.0です")
+        if lr_count > 1:
+            print(
+                f"when multiple learning rates are specified with dadaptation (e.g. for Text Encoder and U-Net), only the first one will take effect / D-Adaptationで複数の学習率を指定した場合（Text EncoderとU-Netなど）、最初の学習率のみが有効になります: lr={actual_lr}"
+            )
+
+        optimizer_class = dadaptation.DAdaptAdan
+        optimizer = optimizer_class(trainable_params, lr=lr, **optimizer_kwargs)
+        
+    elif optimizer_type == "DAdaptSGD".lower():
+        try:
+            import dadaptation
+        except ImportError:
+            raise ImportError("No dadaptation / dadaptation がインストールされていないようです")
+        print(f"use D-Adaptation SGD optimizer | {optimizer_kwargs}")
+
+        actual_lr = lr
+        lr_count = 1
+        if type(trainable_params) == list and type(trainable_params[0]) == dict:
+            lrs = set()
+            actual_lr = trainable_params[0].get("lr", actual_lr)
+            for group in trainable_params:
+                lrs.add(group.get("lr", actual_lr))
+            lr_count = len(lrs)
+
+        if actual_lr <= 0.1:
+            print(
+                f"learning rate is too low. If using dadaptation, set learning rate around 1.0 / 学習率が低すぎるようです。1.0前後の値を指定してください: lr={actual_lr}"
+            )
+            print("recommend option: lr=1.0 / 推奨は1.0です")
+        if lr_count > 1:
+            print(
+                f"when multiple learning rates are specified with dadaptation (e.g. for Text Encoder and U-Net), only the first one will take effect / D-Adaptationで複数の学習率を指定した場合（Text EncoderとU-Netなど）、最初の学習率のみが有効になります: lr={actual_lr}"
+            )
+
+        optimizer_class = dadaptation.DAdaptSGD
+        optimizer = optimizer_class(trainable_params, lr=lr, **optimizer_kwargs)
+    
     elif optimizer_type == "Adafactor".lower():
         # 引数を確認して適宜補正する
         if "relative_step" not in optimizer_kwargs:

train_README-ja.md

@@ -566,7 +566,10 @@ masterpiece, best quality, 1boy, in business suit, standing at street, looking b
     - Lion8bit : 引数は同上
     - SGDNesterov : [torch.optim.SGD](https://pytorch.org/docs/stable/generated/torch.optim.SGD.html), nesterov=True
     - SGDNesterov8bit : 引数は同上
-    - DAdaptation : https://github.com/facebookresearch/dadaptation
+    - DAdaptation(DAdaptAdam) : https://github.com/facebookresearch/dadaptation
+    - DAdaptAdaGrad : 引数は同上
+    - DAdaptAdan : 引数は同上
+    - DAdaptSGD : 引数は同上
     - AdaFactor : [Transformers AdaFactor](https://huggingface.co/docs/transformers/main_classes/optimizer_schedules)
     - 任意のオプティマイザ
 

train_db.py

@@ -367,7 +367,7 @@ def train(args):
             current_loss = loss.detach().item()
             if args.logging_dir is not None:
                 logs = {"loss": current_loss, "lr": float(lr_scheduler.get_last_lr()[0])}
-                if args.optimizer_type.lower() == "DAdaptation".lower():  # tracking d*lr value
+                if args.optimizer_type.lower() == "DAdaptation".lower() or args.optimizer_type.lower() == "DAdaptAdam".lower() or args.optimizer_type.lower() == "DAdaptAdaGrad".lower() or args.optimizer_type.lower() == "DAdaptAdan".lower() or args.optimizer_type.lower() == "DAdaptSGD".lower():  # tracking d*lr value
                     logs["lr/d*lr"] = (
                         lr_scheduler.optimizers[0].param_groups[0]["d"] * lr_scheduler.optimizers[0].param_groups[0]["lr"]
                     )

train_network.py

@@ -43,7 +43,7 @@ def generate_step_logs(args: argparse.Namespace, current_loss, avr_loss, lr_sche
             logs["lr/textencoder"] = float(lrs[0])
             logs["lr/unet"] = float(lrs[-1])  # may be same to textencoder
 
-        if args.optimizer_type.lower() == "DAdaptation".lower():  # tracking d*lr value of unet.
+        if args.optimizer_type.lower() == "DAdaptation".lower() or args.optimizer_type.lower() == "DAdaptAdam".lower() or args.optimizer_type.lower() == "DAdaptAdaGrad".lower() or args.optimizer_type.lower() == "DAdaptAdan".lower() or args.optimizer_type.lower() == "DAdaptSGD".lower():  # tracking d*lr value of unet.
             logs["lr/d*lr"] = lr_scheduler.optimizers[-1].param_groups[0]["d"] * lr_scheduler.optimizers[-1].param_groups[0]["lr"]
     else:
         idx = 0
@@ -53,7 +53,7 @@ def generate_step_logs(args: argparse.Namespace, current_loss, avr_loss, lr_sche
 
         for i in range(idx, len(lrs)):
             logs[f"lr/group{i}"] = float(lrs[i])
-            if args.optimizer_type.lower() == "DAdaptation".lower():
+            if args.optimizer_type.lower() == "DAdaptation".lower() or args.optimizer_type.lower() == "DAdaptAdam".lower() or args.optimizer_type.lower() == "DAdaptAdaGrad".lower() or args.optimizer_type.lower() == "DAdaptAdan".lower() or args.optimizer_type.lower() == "DAdaptSGD".lower():
                 logs[f"lr/d*lr/group{i}"] = (
                     lr_scheduler.optimizers[-1].param_groups[i]["d"] * lr_scheduler.optimizers[-1].param_groups[i]["lr"]
                 )

train_textual_inversion.py

@@ -465,7 +465,7 @@ def train(args):
             current_loss = loss.detach().item()
             if args.logging_dir is not None:
                 logs = {"loss": current_loss, "lr": float(lr_scheduler.get_last_lr()[0])}
-                if args.optimizer_type.lower() == "DAdaptation".lower():  # tracking d*lr value
+                if args.optimizer_type.lower() == "DAdaptation".lower() or args.optimizer_type.lower() == "DAdaptAdam".lower() or args.optimizer_type.lower() == "DAdaptAdaGrad".lower() or args.optimizer_type.lower() == "DAdaptAdan".lower() or args.optimizer_type.lower() == "DAdaptSGD".lower():  # tracking d*lr value
                     logs["lr/d*lr"] = (
                         lr_scheduler.optimizers[0].param_groups[0]["d"] * lr_scheduler.optimizers[0].param_groups[0]["lr"]
                     )

train_textual_inversion_XTI.py

@@ -504,7 +504,7 @@ def train(args):
             current_loss = loss.detach().item()
             if args.logging_dir is not None:
                 logs = {"loss": current_loss, "lr": float(lr_scheduler.get_last_lr()[0])}
-                if args.optimizer_type.lower() == "DAdaptation".lower():  # tracking d*lr value
+                if args.optimizer_type.lower() == "DAdaptation".lower() or args.optimizer_type.lower() == "DAdaptAdam".lower() or args.optimizer_type.lower() == "DAdaptAdaGrad".lower() or args.optimizer_type.lower() == "DAdaptAdan".lower() or args.optimizer_type.lower() == "DAdaptSGD".lower():  # tracking d*lr value
                     logs["lr/d*lr"] = (
                         lr_scheduler.optimizers[0].param_groups[0]["d"] * lr_scheduler.optimizers[0].param_groups[0]["lr"]
                     )