add EMA during training

examples/train_unconditional.py

@@ -9,14 +9,14 @@ from accelerate import Accelerator
 from datasets import load_dataset
 from diffusers import DDPM, DDPMScheduler, UNetModel
 from diffusers.hub_utils import init_git_repo, push_to_hub
-from diffusers.modeling_utils import unwrap_model
 from diffusers.optimization import get_scheduler
+from diffusers.training_utils import EMAModel
 from diffusers.utils import logging
 from torchvision.transforms import (
     CenterCrop,
     Compose,
     InterpolationMode,
-    Lambda,
+    Normalize,
     RandomHorizontalFlip,
     Resize,
     ToTensor,
@@ -48,7 +48,7 @@ def main(args):
             CenterCrop(args.resolution),
             RandomHorizontalFlip(),
             ToTensor(),
-            Lambda(lambda x: x * 2 - 1),
+            Normalize([0.5], [0.5]),
         ]
     )
     dataset = load_dataset(args.dataset, split="train")
@@ -71,6 +71,8 @@ def main(args):
         model, optimizer, train_dataloader, lr_scheduler
     )
 
+    ema_model = EMAModel(model, inv_gamma=1.0, power=3 / 4)
+
     if args.push_to_hub:
         repo = init_git_repo(args, at_init=True)
 
@@ -87,6 +89,7 @@ def main(args):
     logger.info(f"  Gradient Accumulation steps = {args.gradient_accumulation_steps}")
     logger.info(f"  Total optimization steps = {max_steps}")
 
+    global_step = 0
     for epoch in range(args.num_epochs):
         model.train()
         with tqdm(total=len(train_dataloader), unit="ba") as pbar:
@@ -117,19 +120,22 @@ def main(args):
                     torch.nn.utils.clip_grad_norm_(model.parameters(), 1.0)
                     optimizer.step()
                     lr_scheduler.step()
+                    ema_model.step(model, global_step)
                     optimizer.zero_grad()
                 pbar.update(1)
-                pbar.set_postfix(loss=loss.detach().item(), lr=optimizer.param_groups[0]["lr"])
+                pbar.set_postfix(
+                    loss=loss.detach().item(), lr=optimizer.param_groups[0]["lr"], ema_decay=ema_model.decay
+                )
+                global_step += 1
 
-                optimizer.step()
-        if is_distributed:
-            torch.distributed.barrier()
+        accelerator.wait_for_everyone()
 
         # Generate a sample image for visual inspection
-        if args.local_rank in [-1, 0]:
-            model.eval()
+        if accelerator.is_main_process:
             with torch.no_grad():
-                pipeline = DDPM(unet=unwrap_model(model), noise_scheduler=noise_scheduler)
+                pipeline = DDPM(
+                    unet=accelerator.unwrap_model(ema_model.averaged_model), noise_scheduler=noise_scheduler
+                )
 
                 generator = torch.manual_seed(0)
                 # run pipeline in inference (sample random noise and denoise)
@@ -151,8 +157,7 @@ def main(args):
                 push_to_hub(args, pipeline, repo, commit_message=f"Epoch {epoch}", blocking=False)
             else:
                 pipeline.save_pretrained(args.output_dir)
-        if is_distributed:
-            torch.distributed.barrier()
+        accelerator.wait_for_everyone()
 
 
 if __name__ == "__main__":

src/diffusers/training_utils.py

+import copy
+
+import torch
+
+
+class EMAModel:
+    """
+    Exponential Moving Average of models weights
+    """
+
+    def __init__(
+        self,
+        model,
+        update_after_step=0,
+        inv_gamma=1.0,
+        power=2 / 3,
+        min_value=0.0,
+        max_value=0.9999,
+        device=None,
+    ):
+        """
+        @crowsonkb's notes on EMA Warmup:
+            If gamma=1 and power=1, implements a simple average. gamma=1, power=2/3 are
+            good values for models you plan to train for a million or more steps (reaches decay
+            factor 0.999 at 31.6K steps, 0.9999 at 1M steps), gamma=1, power=3/4 for models
+            you plan to train for less (reaches decay factor 0.999 at 10K steps, 0.9999 at
+            215.4k steps).
+        Args:
+            inv_gamma (float): Inverse multiplicative factor of EMA warmup. Default: 1.
+            power (float): Exponential factor of EMA warmup. Default: 2/3.
+            min_value (float): The minimum EMA decay rate. Default: 0.
+        """
+
+        self.averaged_model = copy.deepcopy(model)
+        self.averaged_model.requires_grad_(False)
+
+        self.update_after_step = update_after_step
+        self.inv_gamma = inv_gamma
+        self.power = power
+        self.min_value = min_value
+        self.max_value = max_value
+
+        if device is not None:
+            self.averaged_model = self.averaged_model.to(device=device)
+
+        self.decay = 0.0
+
+    def get_decay(self, optimization_step):
+        """
+        Compute the decay factor for the exponential moving average.
+        """
+        step = max(0, optimization_step - self.update_after_step - 1)
+        value = 1 - (1 + step / self.inv_gamma) ** -self.power
+
+        if step <= 0:
+            return 0.0
+
+        return max(self.min_value, min(value, self.max_value))
+
+    @torch.no_grad()
+    def step(self, new_model, optimization_step):
+        ema_state_dict = {}
+        ema_params = self.averaged_model.state_dict()
+
+        self.decay = self.get_decay(optimization_step)
+
+        for key, param in new_model.named_parameters():
+            if isinstance(param, dict):
+                continue
+            try:
+                ema_param = ema_params[key]
+            except KeyError:
+                ema_param = param.float().clone() if param.ndim == 1 else copy.deepcopy(param)
+                ema_params[key] = ema_param
+
+            if not param.requires_grad:
+                ema_params[key].copy_(param.to(dtype=ema_param.dtype).data)
+                ema_param = ema_params[key]
+            else:
+                ema_param.mul_(self.decay)
+                ema_param.add_(param.data.to(dtype=ema_param.dtype), alpha=1 - self.decay)
+
+            ema_state_dict[key] = ema_param
+
+        for key, param in new_model.named_buffers():
+            ema_state_dict[key] = param
+
+        self.averaged_model.load_state_dict(ema_state_dict, strict=False)

tests/test_modeling_utils.py

@@ -25,10 +25,10 @@ from diffusers import (
     BDDM,
     DDIM,
     DDPM,
-    Glide,
     PNDM,
     DDIMScheduler,
     DDPMScheduler,
+    Glide,
     GlideSuperResUNetModel,
     GlideTextToImageUNetModel,
     GradTTS,