Fix EMA and make it compatible with deepspeed.

examples/text_to_image/train_text_to_image.py

@@ -1,11 +1,10 @@
 import argparse
-import copy
 import logging
 import math
 import os
 import random
 from pathlib import Path
-from typing import Optional
+from typing import Iterable, Optional
 
 import numpy as np
 import torch
@@ -234,25 +233,17 @@ def get_full_repo_name(model_id: str, organization: Optional[str] = None, token:
 }
 
 
+# Adapted from torch-ema https://github.com/fadel/pytorch_ema/blob/master/torch_ema/ema.py#L14
 class EMAModel:
     """
     Exponential Moving Average of models weights
     """
 
-    def __init__(
-        self,
-        model,
-        decay=0.9999,
-        device=None,
-    ):
-        self.averaged_model = copy.deepcopy(model).eval()
-        self.averaged_model.requires_grad_(False)
+    def __init__(self, parameters: Iterable[torch.nn.Parameter], decay=0.9999):
+        parameters = list(parameters)
+        self.shadow_params = [p.clone().detach() for p in parameters]
 
         self.decay = decay
-
-        if device is not None:
-            self.averaged_model = self.averaged_model.to(device=device)
-
         self.optimization_step = 0
 
     def get_decay(self, optimization_step):
@@ -263,34 +254,47 @@ def get_decay(self, optimization_step):
         return 1 - min(self.decay, value)
 
     @torch.no_grad()
-    def step(self, new_model):
-        ema_state_dict = self.averaged_model.state_dict()
+    def step(self, parameters):
+        parameters = list(parameters)
 
         self.optimization_step += 1
         self.decay = self.get_decay(self.optimization_step)
 
-        for key, param in new_model.named_parameters():
-            if isinstance(param, dict):
-                continue
-            try:
-                ema_param = ema_state_dict[key]
-            except KeyError:
-                ema_param = param.float().clone() if param.ndim == 1 else copy.deepcopy(param)
-                ema_state_dict[key] = ema_param
-
-            param = param.clone().detach().to(ema_param.dtype).to(ema_param.device)
-
+        for s_param, param in zip(self.shadow_params, parameters):
             if param.requires_grad:
-                ema_state_dict[key].sub_(self.decay * (ema_param - param))
+                tmp = self.decay * (s_param - param)
+                s_param.sub_(tmp)
             else:
-                ema_state_dict[key].copy_(param)
-
-        for key, param in new_model.named_buffers():
-            ema_state_dict[key] = param
+                s_param.copy_(param)
 
-        self.averaged_model.load_state_dict(ema_state_dict, strict=False)
         torch.cuda.empty_cache()
 
+    def copy_to(self, parameters: Iterable[torch.nn.Parameter]) -> None:
+        """
+        Copy current averaged parameters into given collection of parameters.
+
+        Args:
+            parameters: Iterable of `torch.nn.Parameter`; the parameters to be
+                updated with the stored moving averages. If `None`, the
+                parameters with which this `ExponentialMovingAverage` was
+                initialized will be used.
+        """
+        parameters = list(parameters)
+        for s_param, param in zip(self.shadow_params, parameters):
+            param.data.copy_(s_param.data)
+
+    def to(self, device=None, dtype=None) -> None:
+        r"""Move internal buffers of the ExponentialMovingAverage to `device`.
+
+        Args:
+            device: like `device` argument to `torch.Tensor.to`
+        """
+        # .to() on the tensors handles None correctly
+        self.shadow_params = [
+            p.to(device=device, dtype=dtype) if p.is_floating_point() else p.to(device=device)
+            for p in self.shadow_params
+        ]
+
 
 def main():
     args = parse_args()
@@ -336,9 +340,6 @@ def main():
     vae = AutoencoderKL.from_pretrained(args.pretrained_model_name_or_path, subfolder="vae")
     unet = UNet2DConditionModel.from_pretrained(args.pretrained_model_name_or_path, subfolder="unet")
 
-    if args.use_ema:
-        ema_unet = EMAModel(unet)
-
     # Freeze vae and text_encoder
     vae.requires_grad_(False)
     text_encoder.requires_grad_(False)
@@ -510,8 +511,9 @@ def collate_fn(examples):
     text_encoder.to(accelerator.device, dtype=weight_dtype)
     vae.to(accelerator.device, dtype=weight_dtype)
 
-    # Move the ema_unet to gpu.
-    ema_unet.averaged_model.to(accelerator.device)
+    # Create EMA for the unet.
+    if args.use_ema:
+        ema_unet = EMAModel(unet.parameters())
 
     # 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)
@@ -583,7 +585,7 @@ def collate_fn(examples):
             # Checks if the accelerator has performed an optimization step behind the scenes
             if accelerator.sync_gradients:
                 if args.use_ema:
-                    ema_unet.step(unet)
+                    ema_unet.step(unet.parameters())
                 progress_bar.update(1)
                 global_step += 1
                 accelerator.log({"train_loss": train_loss}, step=global_step)
@@ -598,10 +600,14 @@ def collate_fn(examples):
     # Create the pipeline using the trained modules and save it.
     accelerator.wait_for_everyone()
     if accelerator.is_main_process:
+        unet = accelerator.unwrap_model(unet)
+        if args.use_ema:
+            ema_unet.copy_to(unet.parameters())
+
         pipeline = StableDiffusionPipeline(
             text_encoder=text_encoder,
             vae=vae,
-            unet=accelerator.unwrap_model(ema_unet.averaged_model if args.use_ema else unet),
+            unet=unet,
             tokenizer=tokenizer,
             scheduler=PNDMScheduler(
                 beta_start=0.00085, beta_end=0.012, beta_schedule="scaled_linear", skip_prk_steps=True