simplyfy code by updating accelerate to 0.30.0

* init

* use no schedule

* fix typo

* update for eval()

* fix typo

.github/workflows/typos.yml

@@ -18,4 +18,4 @@ jobs:
       - uses: actions/checkout@v4
 
       - name: typos-action
-        uses: crate-ci/typos@v1.17.2
+        uses: crate-ci/typos@v1.19.0

README.md

@@ -137,6 +137,12 @@ The majority of scripts is licensed under ASL 2.0 (including codes from Diffuser
 
 ## Change History
 
+### Apr 7, 2024 / 2024-04-07: v0.8.7
+
+- The default value of `huber_schedule` in Scheduled Huber Loss is changed from `exponential` to `snr`, which is expected to give better results.
+
+- Scheduled Huber Loss の `huber_schedule` のデフォルト値を `exponential` から、より良い結果が期待できる `snr` に変更しました。
+
 ### Apr 7, 2024 / 2024-04-07: v0.8.6
 
 #### Highlights
@@ -159,7 +165,7 @@ The majority of scripts is licensed under ASL 2.0 (including codes from Diffuser
 - Fixed a bug that U-Net and Text Encoders are included in the state in `train_network.py` and `sdxl_train_network.py`. The saving and loading of the state are faster, the file size is smaller, and the memory usage when loading is reduced.
 - DeepSpeed is supported. PR [#1101](https://github.com/kohya-ss/sd-scripts/pull/1101)  and [#1139](https://github.com/kohya-ss/sd-scripts/pull/1139) Thanks to BootsofLagrangian! See PR [#1101](https://github.com/kohya-ss/sd-scripts/pull/1101) for details.
 - The masked loss is supported in each training script. PR [#1207](https://github.com/kohya-ss/sd-scripts/pull/1207) See [Masked loss](#about-masked-loss) for details.
-- Scheduled Huber Loss has been introduced to each training scripts. PR [#1228](https://github.com/kohya-ss/sd-scripts/pull/1228/) Thanks to kabachuha for the PR and cheald, drhead, and others for the discussion! See [Scheduled Huber Loss](#about-scheduled-huber-loss) for details.
+- Scheduled Huber Loss has been introduced to each training scripts. PR [#1228](https://github.com/kohya-ss/sd-scripts/pull/1228/) Thanks to kabachuha for the PR and cheald, drhead, and others for the discussion! See the PR and [Scheduled Huber Loss](#about-scheduled-huber-loss) for details.
 - The options `--noise_offset_random_strength` and `--ip_noise_gamma_random_strength` are added to each training script. These options can be used to vary the noise offset and ip noise gamma in the range of 0 to the specified value. PR [#1177](https://github.com/kohya-ss/sd-scripts/pull/1177) Thanks to KohakuBlueleaf!
 - The options `--save_state_on_train_end` are added to each training script. PR [#1168](https://github.com/kohya-ss/sd-scripts/pull/1168) Thanks to gesen2egee!
 - The options `--sample_every_n_epochs` and `--sample_every_n_steps` in each training script now display a warning and ignore them when a number less than or equal to `0` is specified. Thanks to S-Del for raising the issue.
@@ -216,9 +222,11 @@ The newly added arguments loss_type, huber_schedule, and huber_c allow for the s
 See PR [#1228](https://github.com/kohya-ss/sd-scripts/pull/1228/) for details.
 
 - `loss_type`: Specify the loss function type. Choose `huber` for Huber loss, `smooth_l1` for smooth L1 loss, and `l2` for MSE loss. The default is `l2`, which is the same as before.
-- `huber_schedule`: Specify the scheduling method. Choose `exponential`, `constant`, or `SNR`. The default is `exponential`.
+- `huber_schedule`: Specify the scheduling method. Choose `exponential`, `constant`, or `snr`. The default is `snr`.
 - `huber_c`: Specify the Huber's parameter. The default is `0.1`.
 
+Please read [Releases](https://github.com/kohya-ss/sd-scripts/releases) for recent updates.
+
 #### 主要な変更点
 
 - 依存ライブラリが更新されました。[アップグレード](./README-ja.md#アップグレード) を参照しライブラリを更新してください。
@@ -239,7 +247,7 @@ See PR [#1228](https://github.com/kohya-ss/sd-scripts/pull/1228/) for details.
 - `train_network.py` および `sdxl_train_network.py` で、state に U-Net および Text Encoder が含まれる不具合を修正しました。state の保存、読み込みが高速化され、ファイルサイズも小さくなり、また読み込み時のメモリ使用量も削減されます。
 - DeepSpeed がサポートされました。PR [#1101](https://github.com/kohya-ss/sd-scripts/pull/1101) 、[#1139](https://github.com/kohya-ss/sd-scripts/pull/1139) BootsofLagrangian 氏に感謝します。詳細は PR [#1101](https://github.com/kohya-ss/sd-scripts/pull/1101) をご覧ください。
 - 各学習スクリプトでマスクロスをサポートしました。PR [#1207](https://github.com/kohya-ss/sd-scripts/pull/1207) 詳細は [マスクロスについて](#マスクロスについて) をご覧ください。
-- 各学習スクリプトに Scheduled Huber Loss を追加しました。PR [#1228](https://github.com/kohya-ss/sd-scripts/pull/1228/) ご提案いただいた kabachuha 氏、および議論を深めてくださった cheald 氏、drhead 氏を始めとする諸氏に感謝します。詳細は [Scheduled Huber Loss について](#scheduled-huber-loss-について) をご覧ください。
+- 各学習スクリプトに Scheduled Huber Loss を追加しました。PR [#1228](https://github.com/kohya-ss/sd-scripts/pull/1228/) ご提案いただいた kabachuha 氏、および議論を深めてくださった cheald 氏、drhead 氏を始めとする諸氏に感謝します。詳細は当該 PR および [Scheduled Huber Loss について](#scheduled-huber-loss-について) をご覧ください。
 - 各学習スクリプトに、noise offset、ip noise gammaを、それぞれ 0~指定した値の範囲で変動させるオプション `--noise_offset_random_strength` および `--ip_noise_gamma_random_strength` が追加されました。 PR [#1177](https://github.com/kohya-ss/sd-scripts/pull/1177) KohakuBlueleaf 氏に感謝します。
 - 各学習スクリプトに、学習終了時に state を保存する `--save_state_on_train_end` オプションが追加されました。 PR [#1168](https://github.com/kohya-ss/sd-scripts/pull/1168) gesen2egee 氏に感謝します。
 - 各学習スクリプトで `--sample_every_n_epochs` および `--sample_every_n_steps` オプションに `0` 以下の数値を指定した時、警告を表示するとともにそれらを無視するよう変更しました。問題提起していただいた S-Del 氏に感謝します。
@@ -280,9 +288,6 @@ See PR [#1228](https://github.com/kohya-ss/sd-scripts/pull/1228/) for details.
 
 マスクの指定には ControlNet データセットを使用します。マスク画像は RGB 画像である必要があります。R チャンネルのピクセル値 255 がロス計算対象、0 がロス計算対象外になります。0-255 の値は、0-1 の範囲に変換されます（つまりピクセル値 128 の部分はロスの重みが半分になります）。データセットの詳細は [LLLite ドキュメント](./docs/train_lllite_README-ja.md#データセットの準備) をご覧ください。
 
-Please read [Releases](https://github.com/kohya-ss/sd-scripts/releases) for recent updates.
-最近の更新情報は [Release](https://github.com/kohya-ss/sd-scripts/releases) をご覧ください。
-
 #### Scheduled Huber Loss について
 
 各学習スクリプトに、学習データ中の異常値や外れ値（data corruption）への耐性を高めるための手法、Scheduled Huber Lossが導入されました。
@@ -298,11 +303,13 @@ Please read [Releases](https://github.com/kohya-ss/sd-scripts/releases) for rece
 詳細は PR [#1228](https://github.com/kohya-ss/sd-scripts/pull/1228/) をご覧ください。
 
 - `loss_type` : 損失関数の種類を指定します。`huber` で Huber損失、`smooth_l1` で smooth L1 損失、`l2` で MSE 損失を選択します。デフォルトは `l2` で、従来と同様です。
-- `huber_schedule` : スケジューリング方法を指定します。`exponential` で指数関数的、`constant` で一定、`snr` で信号対雑音比に基づくスケジューリングを選択します。デフォルトは `exponential` です。
+- `huber_schedule` : スケジューリング方法を指定します。`exponential` で指数関数的、`constant` で一定、`snr` で信号対雑音比に基づくスケジューリングを選択します。デフォルトは `snr` です。
 - `huber_c` : Huber損失のパラメータを指定します。デフォルトは `0.1` です。
 
 PR 内でいくつかの比較が共有されています。この機能を試す場合、最初は `--loss_type smooth_l1 --huber_schedule snr --huber_c 0.1` などで試してみるとよいかもしれません。
 
+最近の更新情報は [Release](https://github.com/kohya-ss/sd-scripts/releases) をご覧ください。
+
 ## Additional Information
 
 ### Naming of LoRA

fine_tune.py

@@ -250,23 +250,32 @@ def train(args):
         unet.to(weight_dtype)
         text_encoder.to(weight_dtype)
 
+    use_schedule_free_optimizer = args.optimizer_type.lower().endswith("schedulefree")
     if args.deepspeed:
         if args.train_text_encoder:
             ds_model = deepspeed_utils.prepare_deepspeed_model(args, unet=unet, text_encoder=text_encoder)
         else:
             ds_model = deepspeed_utils.prepare_deepspeed_model(args, unet=unet)
-        ds_model, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
-            ds_model, optimizer, train_dataloader, lr_scheduler
-        )
+        ds_model, optimizer, train_dataloader = accelerator.prepare(ds_model, optimizer, train_dataloader)
+        if not use_schedule_free_optimizer:
+            lr_scheduler = accelerator.prepare(lr_scheduler)
         training_models = [ds_model]
     else:
         # acceleratorがなんかよろしくやってくれるらしい
         if args.train_text_encoder:
-            unet, text_encoder, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
-                unet, text_encoder, optimizer, train_dataloader, lr_scheduler
-            )
+            unet, text_encoder, optimizer, train_dataloader = accelerator.prepare(unet, text_encoder, optimizer, train_dataloader)
         else:
-            unet, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(unet, optimizer, train_dataloader, lr_scheduler)
+            unet, optimizer, train_dataloader = accelerator.prepare(unet, optimizer, train_dataloader)
+        if not use_schedule_free_optimizer:
+            lr_scheduler = accelerator.prepare(lr_scheduler)
+
+    # make lambda function for calling optimizer.train() and optimizer.eval() if schedule-free optimizer is used
+    if use_schedule_free_optimizer:
+        optimizer_train_if_needed = lambda: optimizer.train()
+        optimizer_eval_if_needed = lambda: optimizer.eval()
+    else:
+        optimizer_train_if_needed = lambda: None
+        optimizer_eval_if_needed = lambda: None
 
     # 実験的機能：勾配も含めたfp16学習を行う　PyTorchにパッチを当ててfp16でのgrad scaleを有効にする
     if args.full_fp16:
@@ -324,6 +333,7 @@ def train(args):
             m.train()
 
         for step, batch in enumerate(train_dataloader):
+            optimizer_train_if_needed()
             current_step.value = global_step
             with accelerator.accumulate(*training_models):
                 with torch.no_grad():
@@ -354,7 +364,9 @@ def train(args):
 
                 # Sample noise, sample a random timestep for each image, and add noise to the latents,
                 # with noise offset and/or multires noise if specified
-                noise, noisy_latents, timesteps, huber_c = train_util.get_noise_noisy_latents_and_timesteps(args, noise_scheduler, latents)
+                noise, noisy_latents, timesteps, huber_c = train_util.get_noise_noisy_latents_and_timesteps(
+                    args, noise_scheduler, latents
+                )
 
                 # Predict the noise residual
                 with accelerator.autocast():
@@ -368,7 +380,9 @@ def train(args):
 
                 if args.min_snr_gamma or args.scale_v_pred_loss_like_noise_pred or args.debiased_estimation_loss:
                     # do not mean over batch dimension for snr weight or scale v-pred loss
-                    loss = train_util.conditional_loss(noise_pred.float(), target.float(), reduction="none", loss_type=args.loss_type, huber_c=huber_c)
+                    loss = train_util.conditional_loss(
+                        noise_pred.float(), target.float(), reduction="none", loss_type=args.loss_type, huber_c=huber_c
+                    )
                     loss = loss.mean([1, 2, 3])
 
                     if args.min_snr_gamma:
@@ -380,7 +394,9 @@ def train(args):
 
                     loss = loss.mean()  # mean over batch dimension
                 else:
-                    loss = train_util.conditional_loss(noise_pred.float(), target.float(), reduction="mean", loss_type=args.loss_type, huber_c=huber_c)
+                    loss = train_util.conditional_loss(
+                        noise_pred.float(), target.float(), reduction="mean", loss_type=args.loss_type, huber_c=huber_c
+                    )
 
                 accelerator.backward(loss)
                 if accelerator.sync_gradients and args.max_grad_norm != 0.0:
@@ -390,9 +406,11 @@ def train(args):
                     accelerator.clip_grad_norm_(params_to_clip, args.max_grad_norm)
 
                 optimizer.step()
-                lr_scheduler.step()
+                lr_scheduler.step()  # if schedule-free optimizer is used, this is a no-op
                 optimizer.zero_grad(set_to_none=True)
 
+            optimizer_eval_if_needed()
+
             # Checks if the accelerator has performed an optimization step behind the scenes
             if accelerator.sync_gradients:
                 progress_bar.update(1)
@@ -471,7 +489,7 @@ def train(args):
 
     accelerator.end_training()
 
-    if is_main_process and (args.save_state or args.save_state_on_train_end):        
+    if is_main_process and (args.save_state or args.save_state_on_train_end):
         train_util.save_state_on_train_end(args, accelerator)
 
     del accelerator  # この後メモリを使うのでこれは消す

library/train_util.py

@@ -3087,7 +3087,7 @@ def add_training_arguments(parser: argparse.ArgumentParser, support_dreambooth:
     )
     parser.add_argument("--seed", type=int, default=None, help="random seed for training / 学習時の乱数のseed")
     parser.add_argument(
-        "--gradient_checkpointing", action="store_true", help="enable gradient checkpointing / grandient checkpointingを有効にする"
+        "--gradient_checkpointing", action="store_true", help="enable gradient checkpointing / gradient checkpointingを有効にする"
     )
     parser.add_argument(
         "--gradient_accumulation_steps",
@@ -3246,10 +3246,10 @@ def add_training_arguments(parser: argparse.ArgumentParser, support_dreambooth:
     parser.add_argument(
         "--huber_schedule",
         type=str,
-        default="exponential",
+        default="snr",
         choices=["constant", "exponential", "snr"],
-        help="The scheduling method for Huber loss (constant, exponential, or SNR-based). Only used when loss_type is 'huber' or 'smooth_l1'. default is exponential"
-        + " / Huber損失のスケジューリング方法（constant、exponential、またはSNRベース）。loss_typeが'huber'または'smooth_l1'の場合に有効、デフォルトはexponential",
+        help="The scheduling method for Huber loss (constant, exponential, or SNR-based). Only used when loss_type is 'huber' or 'smooth_l1'. default is snr"
+        + " / Huber損失のスケジューリング方法（constant、exponential、またはSNRベース）。loss_typeが'huber'または'smooth_l1'の場合に有効、デフォルトは snr",
     )
     parser.add_argument(
         "--huber_c",
@@ -4088,6 +4088,21 @@ def get_optimizer(args, trainable_params):
         optimizer_class = torch.optim.AdamW
         optimizer = optimizer_class(trainable_params, lr=lr, **optimizer_kwargs)
 
+    elif optimizer_type.endswith("schedulefree".lower()):
+        try:
+            import schedulefree as sf
+        except ImportError:
+            raise ImportError("No schedulefree / schedulefreeがインストールされていないようです")
+        if optimizer_type == "AdamWScheduleFree".lower():
+            optimizer_class = sf.AdamWScheduleFree
+            logger.info(f"use AdamWScheduleFree optimizer | {optimizer_kwargs}")
+        elif optimizer_type == "SGDScheduleFree".lower():
+            optimizer_class = sf.SGDScheduleFree
+            logger.info(f"use SGDScheduleFree optimizer | {optimizer_kwargs}")
+        else:
+            raise ValueError(f"Unknown optimizer type: {optimizer_type}")
+        optimizer = optimizer_class(trainable_params, lr=lr, **optimizer_kwargs)
+
     if optimizer is None:
         # 任意のoptimizerを使う
         optimizer_type = args.optimizer_type  # lowerでないやつ（微妙）
@@ -4116,6 +4131,14 @@ def get_scheduler_fix(args, optimizer: Optimizer, num_processes: int):
     """
     Unified API to get any scheduler from its name.
     """
+    # supports schedule free optimizer
+    if args.optimizer_type.lower().endswith("schedulefree"):
+        # return dummy scheduler: it has 'step' method but does nothing
+        logger.info("use dummy scheduler for schedule free optimizer / schedule free optimizer用のダミースケジューラを使用します")
+        lr_scheduler = TYPE_TO_SCHEDULER_FUNCTION[SchedulerType.CONSTANT](optimizer)
+        lr_scheduler.step = lambda: None
+        return lr_scheduler
+
     name = args.lr_scheduler
     num_warmup_steps: Optional[int] = args.lr_warmup_steps
     num_training_steps = args.max_train_steps * num_processes  # * args.gradient_accumulation_steps
@@ -4250,7 +4273,7 @@ def load_tokenizer(args: argparse.Namespace):
     return tokenizer
 
 
-def prepare_accelerator(args: argparse.Namespace):
+def prepare_accelerator(args: argparse.Namespace) -> Accelerator:
     """
     this function also prepares deepspeed plugin
     """

requirements.txt

@@ -1,4 +1,4 @@
-accelerate==0.25.0
+accelerate==0.30.0
 transformers==4.36.2
 diffusers[torch]==0.25.0
 ftfy==6.1.1
@@ -9,6 +9,7 @@ pytorch-lightning==1.9.0
 bitsandbytes==0.43.0
 prodigyopt==1.0
 lion-pytorch==0.0.6
+schedulefree==1.2.5
 tensorboard
 safetensors==0.4.2
 # gradio==3.16.2

sdxl_train.py

@@ -407,6 +407,7 @@ def train(args):
         text_encoder1.text_model.encoder.layers[-1].requires_grad_(False)
         text_encoder1.text_model.final_layer_norm.requires_grad_(False)
 
+    use_schedule_free_optimizer = args.optimizer_type.lower().endswith("schedulefree")
     if args.deepspeed:
         ds_model = deepspeed_utils.prepare_deepspeed_model(
             args,
@@ -415,9 +416,9 @@ def train(args):
             text_encoder2=text_encoder2 if train_text_encoder2 else None,
         )
         # most of ZeRO stage uses optimizer partitioning, so we have to prepare optimizer and ds_model at the same time. # pull/1139#issuecomment-1986790007
-        ds_model, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
-            ds_model, optimizer, train_dataloader, lr_scheduler
-        )
+        ds_model, optimizer, train_dataloader = accelerator.prepare(ds_model, optimizer, train_dataloader)
+        if not use_schedule_free_optimizer:
+            lr_scheduler = accelerator.prepare(lr_scheduler)
         training_models = [ds_model]
 
     else:
@@ -428,7 +429,17 @@ def train(args):
             text_encoder1 = accelerator.prepare(text_encoder1)
         if train_text_encoder2:
             text_encoder2 = accelerator.prepare(text_encoder2)
-        optimizer, train_dataloader, lr_scheduler = accelerator.prepare(optimizer, train_dataloader, lr_scheduler)
+        if not use_schedule_free_optimizer:
+            lr_scheduler = accelerator.prepare(lr_scheduler)
+        optimizer, train_dataloader = accelerator.prepare(optimizer, train_dataloader)
+
+    # make lambda function for calling optimizer.train() and optimizer.eval() if schedule-free optimizer is used
+    if use_schedule_free_optimizer:
+        optimizer_train_if_needed = lambda: optimizer.train()
+        optimizer_eval_if_needed = lambda: optimizer.eval()
+    else:
+        optimizer_train_if_needed = lambda: None
+        optimizer_eval_if_needed = lambda: None
 
     # TextEncoderの出力をキャッシュするときにはCPUへ移動する
     if args.cache_text_encoder_outputs:
@@ -503,6 +514,7 @@ def train(args):
             m.train()
 
         for step, batch in enumerate(train_dataloader):
+            optimizer_train_if_needed()
             current_step.value = global_step
             with accelerator.accumulate(*training_models):
                 if "latents" in batch and batch["latents"] is not None:
@@ -582,7 +594,9 @@ def train(args):
 
                 # Sample noise, sample a random timestep for each image, and add noise to the latents,
                 # with noise offset and/or multires noise if specified
-                noise, noisy_latents, timesteps, huber_c = train_util.get_noise_noisy_latents_and_timesteps(args, noise_scheduler, latents)
+                noise, noisy_latents, timesteps, huber_c = train_util.get_noise_noisy_latents_and_timesteps(
+                    args, noise_scheduler, latents
+                )
 
                 noisy_latents = noisy_latents.to(weight_dtype)  # TODO check why noisy_latents is not weight_dtype
 
@@ -600,7 +614,9 @@ def train(args):
                     or args.masked_loss
                 ):
                     # do not mean over batch dimension for snr weight or scale v-pred loss
-                    loss = train_util.conditional_loss(noise_pred.float(), target.float(), reduction="none", loss_type=args.loss_type, huber_c=huber_c)
+                    loss = train_util.conditional_loss(
+                        noise_pred.float(), target.float(), reduction="none", loss_type=args.loss_type, huber_c=huber_c
+                    )
                     if args.masked_loss:
                         loss = apply_masked_loss(loss, batch)
                     loss = loss.mean([1, 2, 3])
@@ -616,7 +632,9 @@ def train(args):
 
                     loss = loss.mean()  # mean over batch dimension
                 else:
-                    loss = train_util.conditional_loss(noise_pred.float(), target.float(), reduction="mean", loss_type=args.loss_type, huber_c=huber_c)
+                    loss = train_util.conditional_loss(
+                        noise_pred.float(), target.float(), reduction="mean", loss_type=args.loss_type, huber_c=huber_c
+                    )
 
                 accelerator.backward(loss)
                 if accelerator.sync_gradients and args.max_grad_norm != 0.0:
@@ -626,9 +644,11 @@ def train(args):
                     accelerator.clip_grad_norm_(params_to_clip, args.max_grad_norm)
 
                 optimizer.step()
-                lr_scheduler.step()
+                lr_scheduler.step()  # if schedule-free optimizer is used, this is a no-op
                 optimizer.zero_grad(set_to_none=True)
 
+            optimizer_eval_if_needed()
+
             # Checks if the accelerator has performed an optimization step behind the scenes
             if accelerator.sync_gradients:
                 progress_bar.update(1)
@@ -736,7 +756,7 @@ def train(args):
 
     accelerator.end_training()
 
-    if args.save_state or args.save_state_on_train_end:        
+    if args.save_state or args.save_state_on_train_end:
         train_util.save_state_on_train_end(args, accelerator)
 
     del accelerator  # この後メモリを使うのでこれは消す

sdxl_train_control_net_lllite.py

@@ -15,6 +15,7 @@ from tqdm import tqdm
 
 import torch
 from library.device_utils import init_ipex, clean_memory_on_device
+
 init_ipex()
 
 from torch.nn.parallel import DistributedDataParallel as DDP
@@ -286,7 +287,18 @@ def train(args):
     unet.to(weight_dtype)
 
     # acceleratorがなんかよろしくやってくれるらしい
-    unet, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(unet, optimizer, train_dataloader, lr_scheduler)
+    use_schedule_free_optimizer = args.optimizer_type.lower().endswith("schedulefree")
+    unet, optimizer, train_dataloader = accelerator.prepare(unet, optimizer, train_dataloader)
+    if not use_schedule_free_optimizer:
+        lr_scheduler = accelerator.prepare(lr_scheduler)
+
+    # make lambda function for calling optimizer.train() and optimizer.eval() if schedule-free optimizer is used
+    if use_schedule_free_optimizer:
+        optimizer_train_if_needed = lambda: optimizer.train()
+        optimizer_eval_if_needed = lambda: optimizer.eval()
+    else:
+        optimizer_train_if_needed = lambda: None
+        optimizer_eval_if_needed = lambda: None
 
     if args.gradient_checkpointing:
         unet.train()  # according to TI example in Diffusers, train is required -> これオリジナルのU-Netしたので本当は外せる
@@ -390,6 +402,7 @@ def train(args):
         current_epoch.value = epoch + 1
 
         for step, batch in enumerate(train_dataloader):
+            optimizer_train_if_needed()
             current_step.value = global_step
             with accelerator.accumulate(unet):
                 with torch.no_grad():
@@ -439,7 +452,9 @@ def train(args):
 
                 # Sample noise, sample a random timestep for each image, and add noise to the latents,
                 # with noise offset and/or multires noise if specified
-                noise, noisy_latents, timesteps, huber_c = train_util.get_noise_noisy_latents_and_timesteps(args, noise_scheduler, latents)
+                noise, noisy_latents, timesteps, huber_c = train_util.get_noise_noisy_latents_and_timesteps(
+                    args, noise_scheduler, latents
+                )
 
                 noisy_latents = noisy_latents.to(weight_dtype)  # TODO check why noisy_latents is not weight_dtype
 
@@ -458,7 +473,9 @@ def train(args):
                 else:
                     target = noise
 
-                loss = train_util.conditional_loss(noise_pred.float(), target.float(), reduction="none", loss_type=args.loss_type, huber_c=huber_c)
+                loss = train_util.conditional_loss(
+                    noise_pred.float(), target.float(), reduction="none", loss_type=args.loss_type, huber_c=huber_c
+                )
                 loss = loss.mean([1, 2, 3])
 
                 loss_weights = batch["loss_weights"]  # 各sampleごとのweight
@@ -484,6 +501,8 @@ def train(args):
                 lr_scheduler.step()
                 optimizer.zero_grad(set_to_none=True)
 
+            optimizer_eval_if_needed()
+
             # Checks if the accelerator has performed an optimization step behind the scenes
             if accelerator.sync_gradients:
                 progress_bar.update(1)

sdxl_train_control_net_lllite_old.py

@@ -12,6 +12,7 @@ from tqdm import tqdm
 
 import torch
 from library.device_utils import init_ipex, clean_memory_on_device
+
 init_ipex()
 
 from torch.nn.parallel import DistributedDataParallel as DDP
@@ -254,9 +255,19 @@ def train(args):
         network.to(weight_dtype)
 
     # acceleratorがなんかよろしくやってくれるらしい
-    unet, network, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
-        unet, network, optimizer, train_dataloader, lr_scheduler
-    )
+    use_schedule_free_optimizer = args.optimizer_type.lower().endswith("schedulefree")
+    unet, network, optimizer, train_dataloader = accelerator.prepare(unet, network, optimizer, train_dataloader)
+    if not use_schedule_free_optimizer:
+        lr_scheduler = accelerator.prepare(lr_scheduler)
+
+    # make lambda function for calling optimizer.train() and optimizer.eval() if schedule-free optimizer is used
+    if use_schedule_free_optimizer:
+        optimizer_train_if_needed = lambda: optimizer.train()
+        optimizer_eval_if_needed = lambda: optimizer.eval()
+    else:
+        optimizer_train_if_needed = lambda: None
+        optimizer_eval_if_needed = lambda: None
+
     network: control_net_lllite.ControlNetLLLite
 
     if args.gradient_checkpointing:
@@ -357,6 +368,7 @@ def train(args):
         network.on_epoch_start()  # train()
 
         for step, batch in enumerate(train_dataloader):
+            optimizer_train_if_needed()
             current_step.value = global_step
             with accelerator.accumulate(network):
                 with torch.no_grad():
@@ -406,7 +418,9 @@ def train(args):
 
                 # Sample noise, sample a random timestep for each image, and add noise to the latents,
                 # with noise offset and/or multires noise if specified
-                noise, noisy_latents, timesteps, huber_c = train_util.get_noise_noisy_latents_and_timesteps(args, noise_scheduler, latents)
+                noise, noisy_latents, timesteps, huber_c = train_util.get_noise_noisy_latents_and_timesteps(
+                    args, noise_scheduler, latents
+                )
 
                 noisy_latents = noisy_latents.to(weight_dtype)  # TODO check why noisy_latents is not weight_dtype
 
@@ -426,7 +440,9 @@ def train(args):
                 else:
                     target = noise
 
-                loss = train_util.conditional_loss(noise_pred.float(), target.float(), reduction="none", loss_type=args.loss_type, huber_c=huber_c)
+                loss = train_util.conditional_loss(
+                    noise_pred.float(), target.float(), reduction="none", loss_type=args.loss_type, huber_c=huber_c
+                )
                 loss = loss.mean([1, 2, 3])
 
                 loss_weights = batch["loss_weights"]  # 各sampleごとのweight
@@ -452,6 +468,8 @@ def train(args):
                 lr_scheduler.step()
                 optimizer.zero_grad(set_to_none=True)
 
+            optimizer_eval_if_needed()
+
             # Checks if the accelerator has performed an optimization step behind the scenes
             if accelerator.sync_gradients:
                 progress_bar.update(1)

train_controlnet.py

@@ -13,6 +13,7 @@ from tqdm import tqdm
 import torch
 from library import deepspeed_utils
 from library.device_utils import init_ipex, clean_memory_on_device
+
 init_ipex()
 
 from torch.nn.parallel import DistributedDataParallel as DDP
@@ -226,7 +227,7 @@ def train(args):
             )
         vae.to("cpu")
         clean_memory_on_device(accelerator.device)
-        
+
         accelerator.wait_for_everyone()
 
     if args.gradient_checkpointing:
@@ -276,9 +277,18 @@ def train(args):
         controlnet.to(weight_dtype)
 
     # acceleratorがなんかよろしくやってくれるらしい
-    controlnet, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
-        controlnet, optimizer, train_dataloader, lr_scheduler
-    )
+    use_schedule_free_optimizer = args.optimizer_type.lower().endswith("schedulefree")
+    controlnet, optimizer, train_dataloader = accelerator.prepare(controlnet, optimizer, train_dataloader)
+    if not use_schedule_free_optimizer:
+        lr_scheduler = accelerator.prepare(lr_scheduler)
+
+    # make lambda function for calling optimizer.train() and optimizer.eval() if schedule-free optimizer is used
+    if use_schedule_free_optimizer:
+        optimizer_train_if_needed = lambda: optimizer.train()
+        optimizer_eval_if_needed = lambda: optimizer.eval()
+    else:
+        optimizer_train_if_needed = lambda: None
+        optimizer_eval_if_needed = lambda: None
 
     unet.requires_grad_(False)
     text_encoder.requires_grad_(False)
@@ -393,6 +403,7 @@ def train(args):
         current_epoch.value = epoch + 1
 
         for step, batch in enumerate(train_dataloader):
+            optimizer_train_if_needed()
             current_step.value = global_step
             with accelerator.accumulate(controlnet):
                 with torch.no_grad():
@@ -420,7 +431,9 @@ def train(args):
                     )
 
                 # Sample a random timestep for each image
-                timesteps, huber_c = train_util.get_timesteps_and_huber_c(args, 0, noise_scheduler.config.num_train_timesteps, noise_scheduler, b_size, latents.device)
+                timesteps, huber_c = train_util.get_timesteps_and_huber_c(
+                    args, 0, noise_scheduler.config.num_train_timesteps, noise_scheduler, b_size, latents.device
+                )
 
                 # Add noise to the latents according to the noise magnitude at each timestep
                 # (this is the forward diffusion process)
@@ -452,7 +465,9 @@ def train(args):
                 else:
                     target = noise
 
-                loss = train_util.conditional_loss(noise_pred.float(), target.float(), reduction="none", loss_type=args.loss_type, huber_c=huber_c)
+                loss = train_util.conditional_loss(
+                    noise_pred.float(), target.float(), reduction="none", loss_type=args.loss_type, huber_c=huber_c
+                )
                 loss = loss.mean([1, 2, 3])
 
                 loss_weights = batch["loss_weights"]  # 各sampleごとのweight
@@ -472,6 +487,8 @@ def train(args):
                 lr_scheduler.step()
                 optimizer.zero_grad(set_to_none=True)
 
+            optimizer_eval_if_needed()
+
             # Checks if the accelerator has performed an optimization step behind the scenes
             if accelerator.sync_gradients:
                 progress_bar.update(1)

train_db.py

@@ -224,25 +224,34 @@ def train(args):
         text_encoder.to(weight_dtype)
 
     # acceleratorがなんかよろしくやってくれるらしい
+    use_schedule_free_optimizer = args.optimizer_type.lower().endswith("schedulefree")
     if args.deepspeed:
         if args.train_text_encoder:
             ds_model = deepspeed_utils.prepare_deepspeed_model(args, unet=unet, text_encoder=text_encoder)
         else:
             ds_model = deepspeed_utils.prepare_deepspeed_model(args, unet=unet)
-        ds_model, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
-            ds_model, optimizer, train_dataloader, lr_scheduler
-        )
+        ds_model, optimizer, train_dataloader = accelerator.prepare(ds_model, optimizer, train_dataloader)
+        if not use_schedule_free_optimizer:
+            lr_scheduler = accelerator.prepare(lr_scheduler)
         training_models = [ds_model]
 
     else:
         if train_text_encoder:
-            unet, text_encoder, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
-                unet, text_encoder, optimizer, train_dataloader, lr_scheduler
-            )
+            unet, text_encoder, optimizer, train_dataloader = accelerator.prepare(unet, text_encoder, optimizer, train_dataloader)
             training_models = [unet, text_encoder]
         else:
-            unet, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(unet, optimizer, train_dataloader, lr_scheduler)
+            unet, optimizer, train_dataloader = accelerator.prepare(unet, optimizer, train_dataloader)
             training_models = [unet]
+        if not use_schedule_free_optimizer:
+            lr_scheduler = accelerator.prepare(lr_scheduler)
+
+    # make lambda function for calling optimizer.train() and optimizer.eval() if schedule-free optimizer is used
+    if use_schedule_free_optimizer:
+        optimizer_train_if_needed = lambda: optimizer.train()
+        optimizer_eval_if_needed = lambda: optimizer.eval()
+    else:
+        optimizer_train_if_needed = lambda: None
+        optimizer_eval_if_needed = lambda: None
 
     if not train_text_encoder:
         text_encoder.to(accelerator.device, dtype=weight_dtype)  # to avoid 'cpu' vs 'cuda' error
@@ -307,6 +316,7 @@ def train(args):
             text_encoder.train()
 
         for step, batch in enumerate(train_dataloader):
+            optimizer_train_if_needed()
             current_step.value = global_step
             # 指定したステップ数でText Encoderの学習を止める
             if global_step == args.stop_text_encoder_training:
@@ -346,7 +356,9 @@ def train(args):
 
                 # Sample noise, sample a random timestep for each image, and add noise to the latents,
                 # with noise offset and/or multires noise if specified
-                noise, noisy_latents, timesteps, huber_c = train_util.get_noise_noisy_latents_and_timesteps(args, noise_scheduler, latents)
+                noise, noisy_latents, timesteps, huber_c = train_util.get_noise_noisy_latents_and_timesteps(
+                    args, noise_scheduler, latents
+                )
 
                 # Predict the noise residual
                 with accelerator.autocast():
@@ -358,7 +370,9 @@ def train(args):
                 else:
                     target = noise
 
-                loss = train_util.conditional_loss(noise_pred.float(), target.float(), reduction="none", loss_type=args.loss_type, huber_c=huber_c)
+                loss = train_util.conditional_loss(
+                    noise_pred.float(), target.float(), reduction="none", loss_type=args.loss_type, huber_c=huber_c
+                )
                 if args.masked_loss:
                     loss = apply_masked_loss(loss, batch)
                 loss = loss.mean([1, 2, 3])
@@ -387,6 +401,8 @@ def train(args):
                 lr_scheduler.step()
                 optimizer.zero_grad(set_to_none=True)
 
+            optimizer_eval_if_needed()
+
             # Checks if the accelerator has performed an optimization step behind the scenes
             if accelerator.sync_gradients:
                 progress_bar.update(1)

train_network.py

@@ -412,6 +412,7 @@ class NetworkTrainer:
                 t_enc.text_model.embeddings.to(dtype=(weight_dtype if te_weight_dtype != weight_dtype else te_weight_dtype))
 
         # acceleratorがなんかよろしくやってくれるらしい / accelerator will do something good
+        use_schedule_free_optimizer = args.optimizer_type.lower().endswith("schedulefree")
         if args.deepspeed:
             ds_model = deepspeed_utils.prepare_deepspeed_model(
                 args,
@@ -420,9 +421,9 @@ class NetworkTrainer:
                 text_encoder2=text_encoders[1] if train_text_encoder and len(text_encoders) > 1 else None,
                 network=network,
             )
-            ds_model, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
-                ds_model, optimizer, train_dataloader, lr_scheduler
-            )
+            ds_model, optimizer, train_dataloader = accelerator.prepare(ds_model, optimizer, train_dataloader)
+            if not use_schedule_free_optimizer:
+                lr_scheduler = accelerator.prepare(lr_scheduler)
             training_model = ds_model
         else:
             if train_unet:
@@ -438,14 +439,23 @@ class NetworkTrainer:
             else:
                 pass  # if text_encoder is not trained, no need to prepare. and device and dtype are already set
 
-            network, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
-                network, optimizer, train_dataloader, lr_scheduler
-            )
+            network, optimizer, train_dataloader = accelerator.prepare(network, optimizer, train_dataloader)
+            if not use_schedule_free_optimizer:
+                lr_scheduler = accelerator.prepare(lr_scheduler)
             training_model = network
 
+        # make lambda function for calling optimizer.train() and optimizer.eval() if schedule-free optimizer is used
+        if use_schedule_free_optimizer:
+            optimizer_train_if_needed = lambda: (optimizer.optimizer if hasattr(optimizer, "optimizer") else optimizer).train()
+            optimizer_eval_if_needed = lambda: (optimizer.optimizer if hasattr(optimizer, "optimizer") else optimizer).eval()
+        else:
+            optimizer_train_if_needed = lambda: None
+            optimizer_eval_if_needed = lambda: None
+
         if args.gradient_checkpointing:
             # according to TI example in Diffusers, train is required
             unet.train()
+
             for t_enc in text_encoders:
                 t_enc.train()
 
@@ -474,13 +484,15 @@ class NetworkTrainer:
         # before resuming make hook for saving/loading to save/load the network weights only
         def save_model_hook(models, weights, output_dir):
             # pop weights of other models than network to save only network weights
-            if accelerator.is_main_process:
+            # only main process or deepspeed https://github.com/huggingface/diffusers/issues/2606
+            if accelerator.is_main_process or args.deepspeed:
                 remove_indices = []
                 for i, model in enumerate(models):
                     if not isinstance(model, type(accelerator.unwrap_model(network))):
                         remove_indices.append(i)
                 for i in reversed(remove_indices):
-                    weights.pop(i)
+                    if len(weights) > i:
+                        weights.pop(i)
                 # print(f"save model hook: {len(weights)} weights will be saved")
 
         def load_model_hook(models, input_dir):
@@ -802,6 +814,7 @@ class NetworkTrainer:
             accelerator.unwrap_model(network).on_epoch_start(text_encoder, unet)
 
             for step, batch in enumerate(train_dataloader):
+                optimizer_train_if_needed()
                 current_step.value = global_step
                 with accelerator.accumulate(training_model):
                     on_step_start(text_encoder, unet)
@@ -918,6 +931,8 @@ class NetworkTrainer:
                 else:
                     keys_scaled, mean_norm, maximum_norm = None, None, None
 
+                optimizer_eval_if_needed()
+
                 # Checks if the accelerator has performed an optimization step behind the scenes
                 if accelerator.sync_gradients:
                     progress_bar.update(1)

train_textual_inversion.py

@@ -415,20 +415,28 @@ class TextualInversionTrainer:
         lr_scheduler = train_util.get_scheduler_fix(args, optimizer, accelerator.num_processes)
 
         # acceleratorがなんかよろしくやってくれるらしい
+        use_schedule_free_optimizer = args.optimizer_type.lower().endswith("schedulefree")
         if len(text_encoders) == 1:
-            text_encoder_or_list, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
-                text_encoder_or_list, optimizer, train_dataloader, lr_scheduler
+            text_encoder_or_list, optimizer, train_dataloader = accelerator.preparet(
+                text_encoder_or_list, optimizer, train_dataloader
             )
-
         elif len(text_encoders) == 2:
-            text_encoder1, text_encoder2, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
-                text_encoders[0], text_encoders[1], optimizer, train_dataloader, lr_scheduler
+            text_encoder1, text_encoder2, optimizer, train_dataloader = accelerator.prepare(
+                text_encoders[0], text_encoders[1], optimizer, train_dataloader
             )
-
             text_encoder_or_list = text_encoders = [text_encoder1, text_encoder2]
-
         else:
             raise NotImplementedError()
+        if not use_schedule_free_optimizer:
+            optimizer, lr_scheduler = accelerator.prepare(optimizer, lr_scheduler)
+
+        # make lambda function for calling optimizer.train() and optimizer.eval() if schedule-free optimizer is used
+        if use_schedule_free_optimizer:
+            optimizer_train_if_needed = lambda: (optimizer.optimizer if hasattr(optimizer, "optimizer") else optimizer).train()
+            optimizer_eval_if_needed = lambda: (optimizer.optimizer if hasattr(optimizer, "optimizer") else optimizer).eval()
+        else:
+            optimizer_train_if_needed = lambda: None
+            optimizer_eval_if_needed = lambda: None
 
         index_no_updates_list = []
         orig_embeds_params_list = []
@@ -557,6 +565,7 @@ class TextualInversionTrainer:
             loss_total = 0
 
             for step, batch in enumerate(train_dataloader):
+                optimizer_train_if_needed()
                 current_step.value = global_step
                 with accelerator.accumulate(text_encoders[0]):
                     with torch.no_grad():
@@ -588,7 +597,9 @@ class TextualInversionTrainer:
                     else:
                         target = noise
 
-                    loss = train_util.conditional_loss(noise_pred.float(), target.float(), reduction="none", loss_type=args.loss_type, huber_c=huber_c)
+                    loss = train_util.conditional_loss(
+                        noise_pred.float(), target.float(), reduction="none", loss_type=args.loss_type, huber_c=huber_c
+                    )
                     if args.masked_loss:
                         loss = apply_masked_loss(loss, batch)
                     loss = loss.mean([1, 2, 3])
@@ -627,6 +638,8 @@ class TextualInversionTrainer:
                                 index_no_updates
                             ]
 
+                optimizer_eval_if_needed()
+
                 # Checks if the accelerator has performed an optimization step behind the scenes
                 if accelerator.sync_gradients:
                     progress_bar.update(1)

train_textual_inversion_XTI.py

@@ -335,9 +335,18 @@ def train(args):
     lr_scheduler = train_util.get_scheduler_fix(args, optimizer, accelerator.num_processes)
 
     # acceleratorがなんかよろしくやってくれるらしい
-    text_encoder, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
-        text_encoder, optimizer, train_dataloader, lr_scheduler
-    )
+    use_schedule_free_optimizer = args.optimizer_type.lower().endswith("schedulefree")
+    text_encoder, optimizer, train_dataloader = accelerator.prepare(text_encoder, optimizer, train_dataloader)
+    if not use_schedule_free_optimizer:
+        lr_scheduler = accelerator.prepare(lr_scheduler)
+
+    # make lambda function for calling optimizer.train() and optimizer.eval() if schedule-free optimizer is used
+    if use_schedule_free_optimizer:
+        optimizer_train_if_needed = lambda: optimizer.train()
+        optimizer_eval_if_needed = lambda: optimizer.eval()
+    else:
+        optimizer_train_if_needed = lambda: None
+        optimizer_eval_if_needed = lambda: None
 
     index_no_updates = torch.arange(len(tokenizer)) < token_ids_XTI[0]
     # logger.info(len(index_no_updates), torch.sum(index_no_updates))
@@ -438,6 +447,7 @@ def train(args):
         loss_total = 0
 
         for step, batch in enumerate(train_dataloader):
+            optimizer_train_if_needed()
             current_step.value = global_step
             with accelerator.accumulate(text_encoder):
                 with torch.no_grad():
@@ -461,7 +471,9 @@ def train(args):
 
                 # Sample noise, sample a random timestep for each image, and add noise to the latents,
                 # with noise offset and/or multires noise if specified
-                noise, noisy_latents, timesteps, huber_c = train_util.get_noise_noisy_latents_and_timesteps(args, noise_scheduler, latents)
+                noise, noisy_latents, timesteps, huber_c = train_util.get_noise_noisy_latents_and_timesteps(
+                    args, noise_scheduler, latents
+                )
 
                 # Predict the noise residual
                 with accelerator.autocast():
@@ -473,7 +485,9 @@ def train(args):
                 else:
                     target = noise
 
-                loss = train_util.conditional_loss(noise_pred.float(), target.float(), reduction="none", loss_type=args.loss_type, huber_c=huber_c)
+                loss = train_util.conditional_loss(
+                    noise_pred.float(), target.float(), reduction="none", loss_type=args.loss_type, huber_c=huber_c
+                )
                 if args.masked_loss:
                     loss = apply_masked_loss(loss, batch)
                 loss = loss.mean([1, 2, 3])
@@ -505,6 +519,8 @@ def train(args):
                         index_no_updates
                     ]
 
+            optimizer_eval_if_needed()
+
             # Checks if the accelerator has performed an optimization step behind the scenes
             if accelerator.sync_gradients:
                 progress_bar.update(1)