Merge pull request #1359 from kohya-ss/train_resume_step

Train resume step

Author: kohya-ss

README.md

@@ -178,6 +178,12 @@ The majority of scripts is licensed under ASL 2.0 (including codes from Diffuser
 
 - The ControlNet training script `train_controlnet.py` for SD1.5/2.x was not working, but it has been fixed. PR [#1284](https://github.com/kohya-ss/sd-scripts/pull/1284) Thanks to sdbds!
 
+- `train_network.py` and `sdxl_train_network.py` now restore the order/position of data loading from DataSet when resuming training. PR [#1353](https://github.com/kohya-ss/sd-scripts/pull/1353) [#1359](https://github.com/kohya-ss/sd-scripts/pull/1359) Thanks to KohakuBlueleaf!
+  - This resolves the issue where the order of data loading from DataSet changes when resuming training.
+  - Specify the `--skip_until_initial_step` option to skip data loading until the specified step. If not specified, data loading starts from the beginning of the DataSet (same as before).
+  - If `--resume` is specified, the step saved in the state is used.
+  - Specify the `--initial_step` or `--initial_epoch` option to skip data loading until the specified step or epoch. Use these options in conjunction with `--skip_until_initial_step`. These options can be used without `--resume` (use them when resuming training with `--network_weights`).
+
 - An option `--disable_mmap_load_safetensors` is added to disable memory mapping when loading the model's .safetensors in SDXL. PR [#1266](https://github.com/kohya-ss/sd-scripts/pull/1266) Thanks to Zovjsra!
   - It seems that the model file loading is faster in the WSL environment etc.
   - Available in `sdxl_train.py`, `sdxl_train_network.py`, `sdxl_train_textual_inversion.py`, and `sdxl_train_control_net_lllite.py`.
@@ -235,6 +241,12 @@ https://github.com/kohya-ss/sd-scripts/pull/1290) Thanks to frodo821!
 
 - SD1.5/2.x 用の ControlNet 学習スクリプト `train_controlnet.py` が動作しなくなっていたのが修正されました。PR [#1284](https://github.com/kohya-ss/sd-scripts/pull/1284) sdbds 氏に感謝します。
 
+- `train_network.py` および `sdxl_train_network.py` で、学習再開時に DataSet の読み込み順についても復元できるようになりました。PR [#1353](https://github.com/kohya-ss/sd-scripts/pull/1353) [#1359](https://github.com/kohya-ss/sd-scripts/pull/1359) KohakuBlueleaf 氏に感謝します。
+  - これにより、学習再開時に DataSet の読み込み順が変わってしまう問題が解消されます。
+  - `--skip_until_initial_step` オプションを指定すると、指定したステップまで DataSet 読み込みをスキップします。指定しない場合の動作は変わりません（DataSet の最初から読み込みます）
+  - `--resume` オプションを指定すると、state に保存されたステップ数が使用されます。
+  - `--initial_step` または `--initial_epoch` オプションを指定すると、指定したステップまたはエポックまで DataSet 読み込みをスキップします。これらのオプションは `--skip_until_initial_step` と併用してください。またこれらのオプションは `--resume` と併用しなくても使えます（`--network_weights` を用いた学習再開時などにお使いください ）。
+
 - SDXL でモデルの .safetensors を読み込む際にメモリマッピングを無効化するオプション `--disable_mmap_load_safetensors` が追加されました。PR [#1266](https://github.com/kohya-ss/sd-scripts/pull/1266) Zovjsra 氏に感謝します。
   - WSL 環境等でモデルファイルの読み込みが高速化されるようです。
   - `sdxl_train.py`、`sdxl_train_network.py`、`sdxl_train_textual_inversion.py`、`sdxl_train_control_net_lllite.py` で使用可能です。

library/train_util.py

@@ -657,8 +657,16 @@ def set_caching_mode(self, mode):
 
     def set_current_epoch(self, epoch):
         if not self.current_epoch == epoch:  # epochが切り替わったらバケツをシャッフルする
-            self.shuffle_buckets()
-        self.current_epoch = epoch
+            if epoch > self.current_epoch:
+                logger.info("epoch is incremented. current_epoch: {}, epoch: {}".format(self.current_epoch, epoch))
+                num_epochs = epoch - self.current_epoch
+                for _ in range(num_epochs):
+                    self.current_epoch += 1
+                    self.shuffle_buckets()
+                # self.current_epoch seem to be set to 0 again in the next epoch. it may be caused by skipped_dataloader?
+            else:
+                logger.warning("epoch is not incremented. current_epoch: {}, epoch: {}".format(self.current_epoch, epoch))
+                self.current_epoch = epoch
 
     def set_current_step(self, step):
         self.current_step = step
@@ -5553,6 +5561,8 @@ def add(self, *, epoch: int, step: int, loss: float) -> None:
         if epoch == 0:
             self.loss_list.append(loss)
         else:
+            while len(self.loss_list) <= step:
+                self.loss_list.append(0.0)
             self.loss_total -= self.loss_list[step]
             self.loss_list[step] = loss
         self.loss_total += loss

train_network.py

@@ -504,6 +504,15 @@ def save_model_hook(models, weights, output_dir):
                         weights.pop(i)
                 # print(f"save model hook: {len(weights)} weights will be saved")
 
+            # save current ecpoch and step
+            train_state_file = os.path.join(output_dir, "train_state.json")
+            # +1 is needed because the state is saved before current_step is set from global_step
+            logger.info(f"save train state to {train_state_file} at epoch {current_epoch.value} step {current_step.value+1}")
+            with open(train_state_file, "w", encoding="utf-8") as f:
+                json.dump({"current_epoch": current_epoch.value, "current_step": current_step.value + 1}, f)
+
+        steps_from_state = None
+
         def load_model_hook(models, input_dir):
             # remove models except network
             remove_indices = []
@@ -514,6 +523,15 @@ def load_model_hook(models, input_dir):
                 models.pop(i)
             # print(f"load model hook: {len(models)} models will be loaded")
 
+            # load current epoch and step to
+            nonlocal steps_from_state
+            train_state_file = os.path.join(input_dir, "train_state.json")
+            if os.path.exists(train_state_file):
+                with open(train_state_file, "r", encoding="utf-8") as f:
+                    data = json.load(f)
+                steps_from_state = data["current_step"]
+                logger.info(f"load train state from {train_state_file}: {data}")
+
         accelerator.register_save_state_pre_hook(save_model_hook)
         accelerator.register_load_state_pre_hook(load_model_hook)
 
@@ -757,7 +775,54 @@ def load_model_hook(models, input_dir):
             if key in metadata:
                 minimum_metadata[key] = metadata[key]
 
-        progress_bar = tqdm(range(args.max_train_steps), smoothing=0, disable=not accelerator.is_local_main_process, desc="steps")
+        # calculate steps to skip when resuming or starting from a specific step
+        initial_step = 0
+        if args.initial_epoch is not None or args.initial_step is not None:
+            # if initial_epoch or initial_step is specified, steps_from_state is ignored even when resuming
+            if steps_from_state is not None:
+                logger.warning(
+                    "steps from the state is ignored because initial_step is specified / initial_stepが指定されているため、stateからのステップ数は無視されます"
+                )
+            if args.initial_step is not None:
+                initial_step = args.initial_step
+            else:
+                # num steps per epoch is calculated by num_processes and gradient_accumulation_steps
+                initial_step = (args.initial_epoch - 1) * math.ceil(
+                    len(train_dataloader) / accelerator.num_processes / args.gradient_accumulation_steps
+                )
+        else:
+            # if initial_epoch and initial_step are not specified, steps_from_state is used when resuming
+            if steps_from_state is not None:
+                initial_step = steps_from_state
+                steps_from_state = None
+
+        if initial_step > 0:
+            assert (
+                args.max_train_steps > initial_step
+            ), f"max_train_steps should be greater than initial step / max_train_stepsは初期ステップより大きい必要があります: {args.max_train_steps} vs {initial_step}"
+
+        progress_bar = tqdm(
+            range(args.max_train_steps - initial_step), smoothing=0, disable=not accelerator.is_local_main_process, desc="steps"
+        )
+
+        epoch_to_start = 0
+        if initial_step > 0:
+            if args.skip_until_initial_step:
+                # if skip_until_initial_step is specified, load data and discard it to ensure the same data is used
+                if not args.resume:
+                    logger.info(
+                        f"initial_step is specified but not resuming. lr scheduler will be started from the beginning / initial_stepが指定されていますがresumeしていないため、lr schedulerは最初から始まります"
+                    )
+                logger.info(f"skipping {initial_step} steps / {initial_step}ステップをスキップします")
+                initial_step *= args.gradient_accumulation_steps
+
+                # set epoch to start to make initial_step less than len(train_dataloader)
+                epoch_to_start = initial_step // math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
+            else:
+                # if not, only epoch no is skipped for informative purpose
+                epoch_to_start = initial_step // math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
+                initial_step = 0  # do not skip
+
         global_step = 0
 
         noise_scheduler = DDPMScheduler(
@@ -816,16 +881,31 @@ def remove_model(old_ckpt_name):
         self.sample_images(accelerator, args, 0, global_step, accelerator.device, vae, tokenizer, text_encoder, unet)
 
         # training loop
-        for epoch in range(num_train_epochs):
+        if initial_step > 0:  # only if skip_until_initial_step is specified
+            for skip_epoch in range(epoch_to_start):  # skip epochs
+                logger.info(f"skipping epoch {skip_epoch+1} because initial_step (multiplied) is {initial_step}")
+                initial_step -= len(train_dataloader)
+            global_step = initial_step
+
+        for epoch in range(epoch_to_start, num_train_epochs):
             accelerator.print(f"\nepoch {epoch+1}/{num_train_epochs}")
             current_epoch.value = epoch + 1
 
             metadata["ss_epoch"] = str(epoch + 1)
 
             accelerator.unwrap_model(network).on_epoch_start(text_encoder, unet)
 
-            for step, batch in enumerate(train_dataloader):
+            skipped_dataloader = None
+            if initial_step > 0:
+                skipped_dataloader = accelerator.skip_first_batches(train_dataloader, initial_step - 1)
+                initial_step = 1
+
+            for step, batch in enumerate(skipped_dataloader or train_dataloader):
                 current_step.value = global_step
+                if initial_step > 0:
+                    initial_step -= 1
+                    continue
+
                 with accelerator.accumulate(training_model):
                     on_step_start(text_encoder, unet)
 
@@ -1126,6 +1206,25 @@ def setup_parser() -> argparse.ArgumentParser:
         action="store_true",
         help="do not use fp16/bf16 VAE in mixed precision (use float VAE) / mixed precisionでも fp16/bf16 VAEを使わずfloat VAEを使う",
     )
+    parser.add_argument(
+        "--skip_until_initial_step",
+        action="store_true",
+        help="skip training until initial_step is reached / initial_stepに到達するまで学習をスキップする",
+    )
+    parser.add_argument(
+        "--initial_epoch",
+        type=int,
+        default=None,
+        help="initial epoch number, 1 means first epoch (same as not specifying). NOTE: initial_epoch/step doesn't affect to lr scheduler. Which means lr scheduler will start from 0 without `--resume`."
+        + " / 初期エポック数、1で最初のエポック（未指定時と同じ）。注意：initial_epoch/stepはlr schedulerに影響しないため、`--resume`しない場合はlr schedulerは0から始まる",
+    )
+    parser.add_argument(
+        "--initial_step",
+        type=int,
+        default=None,
+        help="initial step number including all epochs, 0 means first step (same as not specifying). overwrites initial_epoch."
+        + " / 初期ステップ数、全エポックを含むステップ数、0で最初のステップ（未指定時と同じ）。initial_epochを上書きする",
+    )
     # parser.add_argument("--loraplus_lr_ratio", default=None, type=float, help="LoRA+ learning rate ratio")
     # parser.add_argument("--loraplus_unet_lr_ratio", default=None, type=float, help="LoRA+ UNet learning rate ratio")
     # parser.add_argument("--loraplus_text_encoder_lr_ratio", default=None, type=float, help="LoRA+ text encoder learning rate ratio")