Add DREAM training (#6381)

A new function compute_dream_and_update_latents has been added to the
training utilities that allows you to do DREAM rectified training in line
with the paper https://arxiv.org/abs/2312.00210. The method can be used
with an extra argument in the train_text_to_image.py script.

Co-authored-by: Jimmy <39@🇺🇸.com>

examples/text_to_image/README.md

@@ -170,6 +170,11 @@ For our small Pokemons dataset, the effects of Min-SNR weighting strategy might
 
 Also, note that in this example, we either predict `epsilon` (i.e., the noise) or the `v_prediction`. For both of these cases, the formulation of the Min-SNR weighting strategy that we have used holds.
 
+#### Training with DREAM
+
+We support training epsilon (noise) prediction models using the [DREAM (Diffusion Rectification and Estimation-Adaptive Models) strategy](https://arxiv.org/abs/2312.00210). DREAM claims to increase model fidelity for the performance cost of an extra grad-less unet `forward` step in the training loop.  You can turn on DREAM training by using the `--dream_training` argument. The `--dream_detail_preservation` argument controls the detail preservation variable p and is the default of 1 from the paper.
+
+
 ## Training with LoRA
 
 Low-Rank Adaption of Large Language Models was first introduced by Microsoft in [LoRA: Low-Rank Adaptation of Large Language Models](https://arxiv.org/abs/2106.09685) by *Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen*.

examples/text_to_image/train_text_to_image.py

@@ -45,7 +45,7 @@ from transformers.utils import ContextManagers
 import diffusers
 from diffusers import AutoencoderKL, DDPMScheduler, StableDiffusionPipeline, UNet2DConditionModel
 from diffusers.optimization import get_scheduler
-from diffusers.training_utils import EMAModel, compute_snr
+from diffusers.training_utils import EMAModel, compute_dream_and_update_latents, compute_snr
 from diffusers.utils import check_min_version, deprecate, is_wandb_available, make_image_grid
 from diffusers.utils.hub_utils import load_or_create_model_card, populate_model_card
 from diffusers.utils.import_utils import is_xformers_available
@@ -361,6 +361,20 @@ def parse_args():
         help="SNR weighting gamma to be used if rebalancing the loss. Recommended value is 5.0. "
         "More details here: https://arxiv.org/abs/2303.09556.",
     )
+    parser.add_argument(
+        "--dream_training",
+        action="store_true",
+        help=(
+            "Use the DREAM training method, which makes training more efficient and accurate at the ",
+            "expense of doing an extra forward pass. See: https://arxiv.org/abs/2312.00210",
+        ),
+    )
+    parser.add_argument(
+        "--dream_detail_preservation",
+        type=float,
+        default=1.0,
+        help="Dream detail preservation factor p (should be greater than 0; default=1.0, as suggested in the paper)",
+    )
     parser.add_argument(
         "--use_8bit_adam", action="store_true", help="Whether or not to use 8-bit Adam from bitsandbytes."
     )
@@ -948,6 +962,18 @@ def main():
                 else:
                     raise ValueError(f"Unknown prediction type {noise_scheduler.config.prediction_type}")
 
+                if args.dream_training:
+                    noisy_latents, target = compute_dream_and_update_latents(
+                        unet,
+                        noise_scheduler,
+                        timesteps,
+                        noise,
+                        noisy_latents,
+                        target,
+                        encoder_hidden_states,
+                        args.dream_detail_preservation,
+                    )
+
                 # Predict the noise residual and compute loss
                 model_pred = unet(noisy_latents, timesteps, encoder_hidden_states, return_dict=False)[0]
 

src/diffusers/training_utils.py

@@ -1,12 +1,13 @@
 import contextlib
 import copy
 import random
-from typing import Any, Dict, Iterable, List, Optional, Union
+from typing import Any, Dict, Iterable, List, Optional, Tuple, Union
 
 import numpy as np
 import torch
 
 from .models import UNet2DConditionModel
+from .schedulers import SchedulerMixin
 from .utils import (
     convert_state_dict_to_diffusers,
     convert_state_dict_to_peft,
@@ -117,6 +118,60 @@ def resolve_interpolation_mode(interpolation_type: str):
     return interpolation_mode
 
 
+def compute_dream_and_update_latents(
+    unet: UNet2DConditionModel,
+    noise_scheduler: SchedulerMixin,
+    timesteps: torch.Tensor,
+    noise: torch.Tensor,
+    noisy_latents: torch.Tensor,
+    target: torch.Tensor,
+    encoder_hidden_states: torch.Tensor,
+    dream_detail_preservation: float = 1.0,
+) -> Tuple[Optional[torch.Tensor], Optional[torch.Tensor]]:
+    """
+    Implements "DREAM (Diffusion Rectification and Estimation-Adaptive Models)" from http://arxiv.org/abs/2312.00210.
+    DREAM helps align training with sampling to help training be more efficient and accurate at the cost of an extra
+    forward step without gradients.
+
+    Args:
+        `unet`: The state unet to use to make a prediction.
+        `noise_scheduler`: The noise scheduler used to add noise for the given timestep.
+        `timesteps`: The timesteps for the noise_scheduler to user.
+        `noise`: A tensor of noise in the shape of noisy_latents.
+        `noisy_latents`: Previously noise latents from the training loop.
+        `target`: The ground-truth tensor to predict after eps is removed.
+        `encoder_hidden_states`: Text embeddings from the text model.
+        `dream_detail_preservation`: A float value that indicates detail preservation level.
+          See reference.
+
+    Returns:
+        `tuple[torch.Tensor, torch.Tensor]`: Adjusted noisy_latents and target.
+    """
+    alphas_cumprod = noise_scheduler.alphas_cumprod.to(timesteps.device)[timesteps, None, None, None]
+    sqrt_one_minus_alphas_cumprod = (1.0 - alphas_cumprod) ** 0.5
+
+    # The paper uses lambda = sqrt(1 - alpha) ** p, with p = 1 in their experiments.
+    dream_lambda = sqrt_one_minus_alphas_cumprod**dream_detail_preservation
+
+    pred = None
+    with torch.no_grad():
+        pred = unet(noisy_latents, timesteps, encoder_hidden_states).sample
+
+    noisy_latents, target = (None, None)
+    if noise_scheduler.config.prediction_type == "epsilon":
+        predicted_noise = pred
+        delta_noise = (noise - predicted_noise).detach()
+        delta_noise.mul_(dream_lambda)
+        noisy_latents = noisy_latents.add(sqrt_one_minus_alphas_cumprod * delta_noise)
+        target = target.add(delta_noise)
+    elif noise_scheduler.config.prediction_type == "v_prediction":
+        raise NotImplementedError("DREAM has not been implemented for v-prediction")
+    else:
+        raise ValueError(f"Unknown prediction type {noise_scheduler.config.prediction_type}")
+
+    return noisy_latents, target
+
+
 def unet_lora_state_dict(unet: UNet2DConditionModel) -> Dict[str, torch.Tensor]:
     r"""
     Returns: