Use accelerate save & loading hooks to have better checkpoint structure (huggingface#2048)

* better accelerated saving

* up

* finish

* finish

* uP

* up

* up

* fix

* Apply suggestions from code review

* correct ema

* Remove @

* up

* Apply suggestions from code review

Co-authored-by: Pedro Cuenca <[email protected]>

* Update docs/source/en/training/dreambooth.mdx

Co-authored-by: Pedro Cuenca <[email protected]>

---------

Co-authored-by: Pedro Cuenca <[email protected]>

Author: patrickvonplaten

docs/source/en/training/dreambooth.mdx

@@ -127,7 +127,30 @@ This would be a good opportunity to tweak some of your hyperparameters if you wi
 
 Saved checkpoints are stored in a format suitable for resuming training. They not only include the model weights, but also the state of the optimizer, data loaders and learning rate.
 
-You can use a checkpoint for inference, but first you need to convert it to an inference pipeline. This is how you could do it:
+**Note**: If you have installed `"accelerate>=0.16.0"` you can use the following code to run 
+inference from an intermediate checkpoint.
+
+```python
+from diffusers import DiffusionPipeline, UNet2DConditionModel
+from transformers import CLIPTextModel
+import torch
+
+# Load the pipeline with the same arguments (model, revision) that were used for training
+model_id = "CompVis/stable-diffusion-v1-4"
+
+unet = UNet2DConditionModel.from_pretrained("/sddata/dreambooth/daruma-v2-1/checkpoint-100/unet")
+
+# if you have trained with `--args.train_text_encoder` make sure to also load the text encoder
+text_encoder = CLIPTextModel.from_pretrained("/sddata/dreambooth/daruma-v2-1/checkpoint-100/text_encoder")
+
+pipeline = DiffusionPipeline.from_pretrained(model_id, unet=unet, text_encoder=text_encoder, dtype=torch.float16)
+pipeline.to("cuda")
+
+# Perform inference, or save, or push to the hub
+pipeline.save_pretrained("dreambooth-pipeline")
+```
+
+If you have installed `"accelerate<0.16.0"` you need to first convert it to an inference pipeline. This is how you could do it:
 
 ```python
 from accelerate import Accelerator
@@ -271,6 +294,10 @@ accelerate launch train_dreambooth.py \
 
 Once you have trained a model, inference can be done using the `StableDiffusionPipeline`, by simply indicating the path where the model was saved. Make sure that your prompts include the special `identifier` used during training (`sks` in the previous examples).
 
+**Note**: If you have installed `"accelerate>=0.16.0"` you can use the following code to run 
+inference from an intermediate checkpoint.
+
+
 ```python
 from diffusers import StableDiffusionPipeline
 import torch
@@ -284,4 +311,4 @@ image = pipe(prompt, num_inference_steps=50, guidance_scale=7.5).images[0]
 image.save("dog-bucket.png")
 ```
 
-You may also run inference from [any of the saved training checkpoints](#performing-inference-using-a-saved-checkpoint).
+You may also run inference from [any of the saved training checkpoints](#performing-inference-using-a-saved-checkpoint).

examples/dreambooth/train_dreambooth.py

@@ -28,6 +28,7 @@
 import torch.utils.checkpoint
 from torch.utils.data import Dataset
 
+import accelerate
 import diffusers
 import transformers
 from accelerate import Accelerator
@@ -38,6 +39,7 @@
 from diffusers.utils import check_min_version
 from diffusers.utils.import_utils import is_xformers_available
 from huggingface_hub import HfFolder, Repository, create_repo, whoami
+from packaging import version
 from PIL import Image
 from torchvision import transforms
 from tqdm.auto import tqdm
@@ -606,6 +608,37 @@ def main(args):
         args.pretrained_model_name_or_path, subfolder="unet", revision=args.revision
     )
 
+    # `accelerate` 0.16.0 will have better support for customized saving
+    if version.parse(accelerate.__version__) >= version.parse("0.16.0"):
+        # create custom saving & loading hooks so that `accelerator.save_state(...)` serializes in a nice format
+        def save_model_hook(models, weights, output_dir):
+            for model in models:
+                sub_dir = "unet" if type(model) == type(unet) else "text_encoder"
+                model.save_pretrained(os.path.join(output_dir, sub_dir))
+
+                # make sure to pop weight so that corresponding model is not saved again
+                weights.pop()
+
+        def load_model_hook(models, input_dir):
+            while len(models) > 0:
+                # pop models so that they are not loaded again
+                model = models.pop()
+
+                if type(model) == type(text_encoder):
+                    # load transformers style into model
+                    load_model = text_encoder_cls.from_pretrained(input_dir, subfolder="text_encoder")
+                    model.config = load_model.config
+                else:
+                    # load diffusers style into model
+                    load_model = UNet2DConditionModel.from_pretrained(input_dir, subfolder="unet")
+                    model.register_to_config(**load_model.config)
+
+                model.load_state_dict(load_model.state_dict())
+                del load_model
+
+        accelerator.register_save_state_pre_hook(save_model_hook)
+        accelerator.register_load_state_pre_hook(load_model_hook)
+
     vae.requires_grad_(False)
     if not args.train_text_encoder:
         text_encoder.requires_grad_(False)

examples/text_to_image/train_text_to_image.py

@@ -26,6 +26,7 @@
 import torch.nn.functional as F
 import torch.utils.checkpoint
 
+import accelerate
 import datasets
 import diffusers
 import transformers
@@ -36,9 +37,10 @@
 from diffusers import AutoencoderKL, DDPMScheduler, StableDiffusionPipeline, UNet2DConditionModel
 from diffusers.optimization import get_scheduler
 from diffusers.training_utils import EMAModel
-from diffusers.utils import check_min_version
+from diffusers.utils import check_min_version, deprecate
 from diffusers.utils.import_utils import is_xformers_available
 from huggingface_hub import HfFolder, Repository, create_repo, whoami
+from packaging import version
 from torchvision import transforms
 from tqdm.auto import tqdm
 from transformers import CLIPTextModel, CLIPTokenizer
@@ -319,6 +321,16 @@ def get_full_repo_name(model_id: str, organization: Optional[str] = None, token:
 
 def main():
     args = parse_args()
+
+    if args.non_ema_revision is not None:
+        deprecate(
+            "non_ema_revision!=None",
+            "0.15.0",
+            message=(
+                "Downloading 'non_ema' weights from revision branches of the Hub is deprecated. Please make sure to"
+                " use `--variant=non_ema` instead."
+            ),
+        )
     logging_dir = os.path.join(args.output_dir, args.logging_dir)
 
     accelerator = Accelerator(
@@ -396,6 +408,39 @@ def main():
         else:
             raise ValueError("xformers is not available. Make sure it is installed correctly")
 
+    # `accelerate` 0.16.0 will have better support for customized saving
+    if version.parse(accelerate.__version__) >= version.parse("0.16.0"):
+        # create custom saving & loading hooks so that `accelerator.save_state(...)` serializes in a nice format
+        def save_model_hook(models, weights, output_dir):
+            if args.use_ema:
+                ema_unet.save_pretrained(os.path.join(output_dir, "unet_ema"))
+
+            for i, model in enumerate(models):
+                model.save_pretrained(os.path.join(output_dir, "unet"))
+
+                # make sure to pop weight so that corresponding model is not saved again
+                weights.pop()
+
+        def load_model_hook(models, input_dir):
+            if args.use_ema:
+                load_model = EMAModel.from_pretrained(os.path.join(input_dir, "unet_ema"), UNet2DConditionModel)
+                ema_unet.load_state_dict(load_model.state_dict())
+                del load_model
+
+            for i in range(len(models)):
+                # pop models so that they are not loaded again
+                model = models.pop()
+
+                # load diffusers style into model
+                load_model = UNet2DConditionModel.from_pretrained(input_dir, subfolder="unet")
+                model.register_to_config(**load_model.config)
+
+                model.load_state_dict(load_model.state_dict())
+                del load_model
+
+        accelerator.register_save_state_pre_hook(save_model_hook)
+        accelerator.register_load_state_pre_hook(load_model_hook)
+
     if args.gradient_checkpointing:
         unet.enable_gradient_checkpointing()
 
@@ -552,8 +597,9 @@ def collate_fn(examples):
     unet, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
         unet, optimizer, train_dataloader, lr_scheduler
     )
+
     if args.use_ema:
-        accelerator.register_for_checkpointing(ema_unet)
+        ema_unet.to(accelerator.device)
 
     # For mixed precision training we cast the text_encoder and vae weights to half-precision
     # as these models are only used for inference, keeping weights in full precision is not required.
@@ -566,8 +612,6 @@ def collate_fn(examples):
     # Move text_encode and vae to gpu and cast to weight_dtype
     text_encoder.to(accelerator.device, dtype=weight_dtype)
     vae.to(accelerator.device, dtype=weight_dtype)
-    if args.use_ema:
-        ema_unet.to(accelerator.device)
 
     # We need to recalculate our total training steps as the size of the training dataloader may have changed.
     num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)

examples/unconditional_image_generation/train_unconditional.py

@@ -9,6 +9,7 @@
 import torch
 import torch.nn.functional as F
 
+import accelerate
 import datasets
 import diffusers
 from accelerate import Accelerator
@@ -19,6 +20,7 @@
 from diffusers.training_utils import EMAModel
 from diffusers.utils import check_min_version
 from huggingface_hub import HfFolder, Repository, create_repo, whoami
+from packaging import version
 from torchvision import transforms
 from tqdm.auto import tqdm
 
@@ -271,6 +273,40 @@ def main(args):
         logging_dir=logging_dir,
     )
 
+    # `accelerate` 0.16.0 will have better support for customized saving
+    if version.parse(accelerate.__version__) >= version.parse("0.16.0"):
+        # create custom saving & loading hooks so that `accelerator.save_state(...)` serializes in a nice format
+        def save_model_hook(models, weights, output_dir):
+            if args.use_ema:
+                ema_model.save_pretrained(os.path.join(output_dir, "unet_ema"))
+
+            for i, model in enumerate(models):
+                model.save_pretrained(os.path.join(output_dir, "unet"))
+
+                # make sure to pop weight so that corresponding model is not saved again
+                weights.pop()
+
+        def load_model_hook(models, input_dir):
+            if args.use_ema:
+                load_model = EMAModel.from_pretrained(os.path.join(input_dir, "unet_ema"), UNet2DModel)
+                ema_model.load_state_dict(load_model.state_dict())
+                ema_model.to(accelerator.device)
+                del load_model
+
+            for i in range(len(models)):
+                # pop models so that they are not loaded again
+                model = models.pop()
+
+                # load diffusers style into model
+                load_model = UNet2DModel.from_pretrained(input_dir, subfolder="unet")
+                model.register_to_config(**load_model.config)
+
+                model.load_state_dict(load_model.state_dict())
+                del load_model
+
+        accelerator.register_save_state_pre_hook(save_model_hook)
+        accelerator.register_load_state_pre_hook(load_model_hook)
+
     # Make one log on every process with the configuration for debugging.
     logging.basicConfig(
         format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
@@ -336,6 +372,8 @@ def main(args):
             use_ema_warmup=True,
             inv_gamma=args.ema_inv_gamma,
             power=args.ema_power,
+            model_cls=UNet2DModel,
+            model_config=model.config,
         )
 
     # Initialize the scheduler
@@ -411,7 +449,6 @@ def transform_images(examples):
     )
 
     if args.use_ema:
-        accelerator.register_for_checkpointing(ema_model)
         ema_model.to(accelerator.device)
 
     # We need to initialize the trackers we use, and also store our configuration.

src/diffusers/training_utils.py

@@ -1,7 +1,7 @@
 import copy
 import os
 import random
-from typing import Iterable, Union
+from typing import Any, Dict, Iterable, Optional, Union
 
 import numpy as np
 import torch
@@ -57,6 +57,8 @@ def __init__(
         use_ema_warmup: bool = False,
         inv_gamma: Union[float, int] = 1.0,
         power: Union[float, int] = 2 / 3,
+        model_cls: Optional[Any] = None,
+        model_config: Dict[str, Any] = None,
         **kwargs,
     ):
         """
@@ -123,6 +125,35 @@ def __init__(
         self.power = power
         self.optimization_step = 0
 
+        self.model_cls = model_cls
+        self.model_config = model_config
+
+    @classmethod
+    def from_pretrained(cls, path, model_cls) -> "EMAModel":
+        _, ema_kwargs = model_cls.load_config(path, return_unused_kwargs=True)
+        model = model_cls.from_pretrained(path)
+
+        ema_model = cls(model.parameters(), model_cls=model_cls, model_config=model.config)
+
+        ema_model.load_state_dict(ema_kwargs)
+        return ema_model
+
+    def save_pretrained(self, path):
+        if self.model_cls is None:
+            raise ValueError("`save_pretrained` can only be used if `model_cls` was defined at __init__.")
+
+        if self.model_config is None:
+            raise ValueError("`save_pretrained` can only be used if `model_config` was defined at __init__.")
+
+        model = self.model_cls.from_config(self.model_config)
+        state_dict = self.state_dict()
+        state_dict.pop("shadow_params", None)
+        state_dict.pop("collected_params", None)
+
+        model.register_to_config(**state_dict)
+        self.copy_to(model.parameters())
+        model.save_pretrained(path)
+
     def get_decay(self, optimization_step: int) -> float:
         """
         Compute the decay factor for the exponential moving average.
@@ -184,7 +215,7 @@ def copy_to(self, parameters: Iterable[torch.nn.Parameter]) -> None:
         """
         parameters = list(parameters)
         for s_param, param in zip(self.shadow_params, parameters):
-            param.data.copy_(s_param.data)
+            param.data.copy_(s_param.to(param.device).data)
 
     def to(self, device=None, dtype=None) -> None:
         r"""Move internal buffers of the ExponentialMovingAverage to `device`.
@@ -257,13 +288,15 @@ def load_state_dict(self, state_dict: dict) -> None:
         if not isinstance(self.power, (float, int)):
             raise ValueError("Invalid power")
 
-        self.shadow_params = state_dict["shadow_params"]
-        if not isinstance(self.shadow_params, list):
-            raise ValueError("shadow_params must be a list")
-        if not all(isinstance(p, torch.Tensor) for p in self.shadow_params):
-            raise ValueError("shadow_params must all be Tensors")
+        shadow_params = state_dict.get("shadow_params", None)
+        if shadow_params is not None:
+            self.shadow_params = shadow_params
+            if not isinstance(self.shadow_params, list):
+                raise ValueError("shadow_params must be a list")
+            if not all(isinstance(p, torch.Tensor) for p in self.shadow_params):
+                raise ValueError("shadow_params must all be Tensors")
 
-        self.collected_params = state_dict["collected_params"]
+        self.collected_params = state_dict.get("collected_params", None)
         if self.collected_params is not None:
             if not isinstance(self.collected_params, list):
                 raise ValueError("collected_params must be a list")