[Advanced DreamBooth LoRA SDXL] Support EDM-style training (follow up of #7126) (#7182)

* add edm style training

* style

* finish adding edm training feature

* import fix

* fix latents mean

* minor adjustments

* add edm to readme

* style

* fix autocast and scheduler config issues when using edm

* style

---------

Co-authored-by: Sayak Paul <[email protected]>

Author: linoytsaban

examples/advanced_diffusion_training/README.md

@@ -259,6 +259,50 @@ pip install git+https://github.com/huggingface/peft.git
 **Inference** 
 The inference is the same as if you train a regular LoRA 🤗
 
+## Conducting EDM-style training
+
+It's now possible to perform EDM-style training as proposed in [Elucidating the Design Space of Diffusion-Based Generative Models](https://arxiv.org/abs/2206.00364). 
+
+simply set:
+
+```diff
++  --do_edm_style_training \
+```
+
+Other SDXL-like models that use the EDM formulation, such as [playgroundai/playground-v2.5-1024px-aesthetic](https://huggingface.co/playgroundai/playground-v2.5-1024px-aesthetic), can also be DreamBooth'd with the script. Below is an example command:
+
+```bash
+accelerate launch train_dreambooth_lora_sdxl_advanced.py \
+  --pretrained_model_name_or_path="playgroundai/playground-v2.5-1024px-aesthetic"  \
+  --dataset_name="linoyts/3d_icon" \
+  --instance_prompt="3d icon in the style of TOK" \
+  --validation_prompt="a TOK icon of an astronaut riding a horse, in the style of TOK" \
+  --output_dir="3d-icon-SDXL-LoRA" \
+  --do_edm_style_training \
+  --caption_column="prompt" \
+  --mixed_precision="bf16" \
+  --resolution=1024 \
+  --train_batch_size=3 \
+  --repeats=1 \
+  --report_to="wandb"\
+  --gradient_accumulation_steps=1 \
+  --gradient_checkpointing \
+  --learning_rate=1.0 \
+  --text_encoder_lr=1.0 \
+  --optimizer="prodigy"\
+  --train_text_encoder_ti\
+  --train_text_encoder_ti_frac=0.5\
+  --lr_scheduler="constant" \
+  --lr_warmup_steps=0 \
+  --rank=8 \
+  --max_train_steps=1000 \
+  --checkpointing_steps=2000 \
+  --seed="0" \
+  --push_to_hub
+```
+
+> [!CAUTION]
+> Min-SNR gamma is not supported with the EDM-style training yet. When training with the PlaygroundAI model, it's recommended to not pass any "variant".
 
 ### Tips and Tricks
 Check out [these recommended practices](https://huggingface.co/blog/sdxl_lora_advanced_script#additional-good-practices)

examples/advanced_diffusion_training/train_dreambooth_lora_sdxl_advanced.py

@@ -14,9 +14,11 @@
 # See the License for the specific language governing permissions and
 
 import argparse
+import contextlib
 import gc
 import hashlib
 import itertools
+import json
 import logging
 import math
 import os
@@ -37,7 +39,7 @@
 from accelerate import Accelerator
 from accelerate.logging import get_logger
 from accelerate.utils import DistributedDataParallelKwargs, ProjectConfiguration, set_seed
-from huggingface_hub import create_repo, upload_folder
+from huggingface_hub import create_repo, hf_hub_download, upload_folder
 from packaging import version
 from peft import LoraConfig, set_peft_model_state_dict
 from peft.utils import get_peft_model_state_dict
@@ -55,6 +57,8 @@
     AutoencoderKL,
     DDPMScheduler,
     DPMSolverMultistepScheduler,
+    EDMEulerScheduler,
+    EulerDiscreteScheduler,
     StableDiffusionXLPipeline,
     UNet2DConditionModel,
 )
@@ -79,6 +83,20 @@
 logger = get_logger(__name__)
 
 
+def determine_scheduler_type(pretrained_model_name_or_path, revision):
+    model_index_filename = "model_index.json"
+    if os.path.isdir(pretrained_model_name_or_path):
+        model_index = os.path.join(pretrained_model_name_or_path, model_index_filename)
+    else:
+        model_index = hf_hub_download(
+            repo_id=pretrained_model_name_or_path, filename=model_index_filename, revision=revision
+        )
+
+    with open(model_index, "r") as f:
+        scheduler_type = json.load(f)["scheduler"][1]
+    return scheduler_type
+
+
 def save_model_card(
     repo_id: str,
     use_dora: bool,
@@ -370,6 +388,11 @@ def parse_args(input_args=None):
             " `args.validation_prompt` multiple times: `args.num_validation_images`."
         ),
     )
+    parser.add_argument(
+        "--do_edm_style_training",
+        action="store_true",
+        help="Flag to conduct training using the EDM formulation as introduced in https://arxiv.org/abs/2206.00364.",
+    )
     parser.add_argument(
         "--with_prior_preservation",
         default=False,
@@ -1117,6 +1140,8 @@ def main(args):
             "You cannot use both --report_to=wandb and --hub_token due to a security risk of exposing your token."
             " Please use `huggingface-cli login` to authenticate with the Hub."
         )
+    if args.do_edm_style_training and args.snr_gamma is not None:
+        raise ValueError("Min-SNR formulation is not supported when conducting EDM-style training.")
 
     logging_dir = Path(args.output_dir, args.logging_dir)
 
@@ -1234,7 +1259,19 @@ def main(args):
     )
 
     # Load scheduler and models
-    noise_scheduler = DDPMScheduler.from_pretrained(args.pretrained_model_name_or_path, subfolder="scheduler")
+    scheduler_type = determine_scheduler_type(args.pretrained_model_name_or_path, args.revision)
+    if "EDM" in scheduler_type:
+        args.do_edm_style_training = True
+        noise_scheduler = EDMEulerScheduler.from_pretrained(args.pretrained_model_name_or_path, subfolder="scheduler")
+        logger.info("Performing EDM-style training!")
+    elif args.do_edm_style_training:
+        noise_scheduler = EulerDiscreteScheduler.from_pretrained(
+            args.pretrained_model_name_or_path, subfolder="scheduler"
+        )
+        logger.info("Performing EDM-style training!")
+    else:
+        noise_scheduler = DDPMScheduler.from_pretrained(args.pretrained_model_name_or_path, subfolder="scheduler")
+
     text_encoder_one = text_encoder_cls_one.from_pretrained(
         args.pretrained_model_name_or_path, subfolder="text_encoder", revision=args.revision, variant=args.variant
     )
@@ -1252,7 +1289,12 @@ def main(args):
         revision=args.revision,
         variant=args.variant,
     )
-    vae_scaling_factor = vae.config.scaling_factor
+    latents_mean = latents_std = None
+    if hasattr(vae.config, "latents_mean") and vae.config.latents_mean is not None:
+        latents_mean = torch.tensor(vae.config.latents_mean).view(1, 4, 1, 1)
+    if hasattr(vae.config, "latents_std") and vae.config.latents_std is not None:
+        latents_std = torch.tensor(vae.config.latents_std).view(1, 4, 1, 1)
+
     unet = UNet2DConditionModel.from_pretrained(
         args.pretrained_model_name_or_path, subfolder="unet", revision=args.revision, variant=args.variant
     )
@@ -1790,6 +1832,19 @@ def compute_text_embeddings(prompt, text_encoders, tokenizers):
         disable=not accelerator.is_local_main_process,
     )
 
+    def get_sigmas(timesteps, n_dim=4, dtype=torch.float32):
+        # TODO: revisit other sampling algorithms
+        sigmas = noise_scheduler.sigmas.to(device=accelerator.device, dtype=dtype)
+        schedule_timesteps = noise_scheduler.timesteps.to(accelerator.device)
+        timesteps = timesteps.to(accelerator.device)
+
+        step_indices = [(schedule_timesteps == t).nonzero().item() for t in timesteps]
+
+        sigma = sigmas[step_indices].flatten()
+        while len(sigma.shape) < n_dim:
+            sigma = sigma.unsqueeze(-1)
+        return sigma
+
     if args.train_text_encoder:
         num_train_epochs_text_encoder = int(args.train_text_encoder_frac * args.num_train_epochs)
     elif args.train_text_encoder_ti:  # args.train_text_encoder_ti
@@ -1841,9 +1896,15 @@ def compute_text_embeddings(prompt, text_encoders, tokenizers):
                     pixel_values = batch["pixel_values"].to(dtype=vae.dtype)
                     model_input = vae.encode(pixel_values).latent_dist.sample()
 
-                model_input = model_input * vae_scaling_factor
-                if args.pretrained_vae_model_name_or_path is None:
-                    model_input = model_input.to(weight_dtype)
+                if latents_mean is None and latents_std is None:
+                    model_input = model_input * vae.config.scaling_factor
+                    if args.pretrained_vae_model_name_or_path is None:
+                        model_input = model_input.to(weight_dtype)
+                else:
+                    latents_mean = latents_mean.to(device=model_input.device, dtype=model_input.dtype)
+                    latents_std = latents_std.to(device=model_input.device, dtype=model_input.dtype)
+                    model_input = (model_input - latents_mean) * vae.config.scaling_factor / latents_std
+                    model_input = model_input.to(dtype=weight_dtype)
 
                 # Sample noise that we'll add to the latents
                 noise = torch.randn_like(model_input)
@@ -1854,15 +1915,32 @@ def compute_text_embeddings(prompt, text_encoders, tokenizers):
                     )
 
                 bsz = model_input.shape[0]
+
                 # Sample a random timestep for each image
-                timesteps = torch.randint(
-                    0, noise_scheduler.config.num_train_timesteps, (bsz,), device=model_input.device
-                )
-                timesteps = timesteps.long()
+                if not args.do_edm_style_training:
+                    timesteps = torch.randint(
+                        0, noise_scheduler.config.num_train_timesteps, (bsz,), device=model_input.device
+                    )
+                    timesteps = timesteps.long()
+                else:
+                    # in EDM formulation, the model is conditioned on the pre-conditioned noise levels
+                    # instead of discrete timesteps, so here we sample indices to get the noise levels
+                    # from `scheduler.timesteps`
+                    indices = torch.randint(0, noise_scheduler.config.num_train_timesteps, (bsz,))
+                    timesteps = noise_scheduler.timesteps[indices].to(device=model_input.device)
 
                 # Add noise to the model input according to the noise magnitude at each timestep
                 # (this is the forward diffusion process)
                 noisy_model_input = noise_scheduler.add_noise(model_input, noise, timesteps)
+                # For EDM-style training, we first obtain the sigmas based on the continuous timesteps.
+                # We then precondition the final model inputs based on these sigmas instead of the timesteps.
+                # Follow: Section 5 of https://arxiv.org/abs/2206.00364.
+                if args.do_edm_style_training:
+                    sigmas = get_sigmas(timesteps, len(noisy_model_input.shape), noisy_model_input.dtype)
+                    if "EDM" in scheduler_type:
+                        inp_noisy_latents = noise_scheduler.precondition_inputs(noisy_model_input, sigmas)
+                    else:
+                        inp_noisy_latents = noisy_model_input / ((sigmas**2 + 1) ** 0.5)
 
                 # time ids
                 add_time_ids = torch.cat(
@@ -1888,7 +1966,7 @@ def compute_text_embeddings(prompt, text_encoders, tokenizers):
                     }
                     prompt_embeds_input = prompt_embeds.repeat(elems_to_repeat_text_embeds, 1, 1)
                     model_pred = unet(
-                        noisy_model_input,
+                        inp_noisy_latents if args.do_edm_style_training else noisy_model_input,
                         timesteps,
                         prompt_embeds_input,
                         added_cond_kwargs=unet_added_conditions,
@@ -1906,14 +1984,42 @@ def compute_text_embeddings(prompt, text_encoders, tokenizers):
                     )
                     prompt_embeds_input = prompt_embeds.repeat(elems_to_repeat_text_embeds, 1, 1)
                     model_pred = unet(
-                        noisy_model_input, timesteps, prompt_embeds_input, added_cond_kwargs=unet_added_conditions
+                        inp_noisy_latents if args.do_edm_style_training else noisy_model_input,
+                        timesteps,
+                        prompt_embeds_input,
+                        added_cond_kwargs=unet_added_conditions,
                     ).sample
 
+                weighting = None
+                if args.do_edm_style_training:
+                    # Similar to the input preconditioning, the model predictions are also preconditioned
+                    # on noised model inputs (before preconditioning) and the sigmas.
+                    # Follow: Section 5 of https://arxiv.org/abs/2206.00364.
+                    if "EDM" in scheduler_type:
+                        model_pred = noise_scheduler.precondition_outputs(noisy_model_input, model_pred, sigmas)
+                    else:
+                        if noise_scheduler.config.prediction_type == "epsilon":
+                            model_pred = model_pred * (-sigmas) + noisy_model_input
+                        elif noise_scheduler.config.prediction_type == "v_prediction":
+                            model_pred = model_pred * (-sigmas / (sigmas**2 + 1) ** 0.5) + (
+                                noisy_model_input / (sigmas**2 + 1)
+                            )
+                    # We are not doing weighting here because it tends result in numerical problems.
+                    # See: https://github.com/huggingface/diffusers/pull/7126#issuecomment-1968523051
+                    # There might be other alternatives for weighting as well:
+                    # https://github.com/huggingface/diffusers/pull/7126#discussion_r1505404686
+                    if "EDM" not in scheduler_type:
+                        weighting = (sigmas**-2.0).float()
+
                 # Get the target for loss depending on the prediction type
                 if noise_scheduler.config.prediction_type == "epsilon":
-                    target = noise
+                    target = model_input if args.do_edm_style_training else noise
                 elif noise_scheduler.config.prediction_type == "v_prediction":
-                    target = noise_scheduler.get_velocity(model_input, noise, timesteps)
+                    target = (
+                        model_input
+                        if args.do_edm_style_training
+                        else noise_scheduler.get_velocity(model_input, noise, timesteps)
+                    )
                 else:
                     raise ValueError(f"Unknown prediction type {noise_scheduler.config.prediction_type}")
 
@@ -1923,10 +2029,28 @@ def compute_text_embeddings(prompt, text_encoders, tokenizers):
                     target, target_prior = torch.chunk(target, 2, dim=0)
 
                     # Compute prior loss
-                    prior_loss = F.mse_loss(model_pred_prior.float(), target_prior.float(), reduction="mean")
+                    if weighting is not None:
+                        prior_loss = torch.mean(
+                            (weighting.float() * (model_pred_prior.float() - target_prior.float()) ** 2).reshape(
+                                target_prior.shape[0], -1
+                            ),
+                            1,
+                        )
+                        prior_loss = prior_loss.mean()
+                    else:
+                        prior_loss = F.mse_loss(model_pred_prior.float(), target_prior.float(), reduction="mean")
 
                 if args.snr_gamma is None:
-                    loss = F.mse_loss(model_pred.float(), target.float(), reduction="mean")
+                    if weighting is not None:
+                        loss = torch.mean(
+                            (weighting.float() * (model_pred.float() - target.float()) ** 2).reshape(
+                                target.shape[0], -1
+                            ),
+                            1,
+                        )
+                        loss = loss.mean()
+                    else:
+                        loss = F.mse_loss(model_pred.float(), target.float(), reduction="mean")
                 else:
                     # Compute loss-weights as per Section 3.4 of https://arxiv.org/abs/2303.09556.
                     # Since we predict the noise instead of x_0, the original formulation is slightly changed.
@@ -2049,26 +2173,32 @@ def compute_text_embeddings(prompt, text_encoders, tokenizers):
                 # We train on the simplified learning objective. If we were previously predicting a variance, we need the scheduler to ignore it
                 scheduler_args = {}
 
-                if "variance_type" in pipeline.scheduler.config:
-                    variance_type = pipeline.scheduler.config.variance_type
+                if not args.do_edm_style_training:
+                    if "variance_type" in pipeline.scheduler.config:
+                        variance_type = pipeline.scheduler.config.variance_type
 
-                    if variance_type in ["learned", "learned_range"]:
-                        variance_type = "fixed_small"
+                        if variance_type in ["learned", "learned_range"]:
+                            variance_type = "fixed_small"
 
-                    scheduler_args["variance_type"] = variance_type
+                        scheduler_args["variance_type"] = variance_type
 
-                pipeline.scheduler = DPMSolverMultistepScheduler.from_config(
-                    pipeline.scheduler.config, **scheduler_args
-                )
+                    pipeline.scheduler = DPMSolverMultistepScheduler.from_config(
+                        pipeline.scheduler.config, **scheduler_args
+                    )
 
                 pipeline = pipeline.to(accelerator.device)
                 pipeline.set_progress_bar_config(disable=True)
 
                 # run inference
                 generator = torch.Generator(device=accelerator.device).manual_seed(args.seed) if args.seed else None
                 pipeline_args = {"prompt": args.validation_prompt}
+                inference_ctx = (
+                    contextlib.nullcontext()
+                    if "playground" in args.pretrained_model_name_or_path
+                    else torch.cuda.amp.autocast()
+                )
 
-                with torch.cuda.amp.autocast():
+                with inference_ctx:
                     images = [
                         pipeline(**pipeline_args, generator=generator).images[0]
                         for _ in range(args.num_validation_images)
@@ -2144,15 +2274,18 @@ def compute_text_embeddings(prompt, text_encoders, tokenizers):
             # We train on the simplified learning objective. If we were previously predicting a variance, we need the scheduler to ignore it
             scheduler_args = {}
 
-            if "variance_type" in pipeline.scheduler.config:
-                variance_type = pipeline.scheduler.config.variance_type
+            if not args.do_edm_style_training:
+                if "variance_type" in pipeline.scheduler.config:
+                    variance_type = pipeline.scheduler.config.variance_type
 
-                if variance_type in ["learned", "learned_range"]:
-                    variance_type = "fixed_small"
+                    if variance_type in ["learned", "learned_range"]:
+                        variance_type = "fixed_small"
 
-                scheduler_args["variance_type"] = variance_type
+                    scheduler_args["variance_type"] = variance_type
 
-            pipeline.scheduler = DPMSolverMultistepScheduler.from_config(pipeline.scheduler.config, **scheduler_args)
+                pipeline.scheduler = DPMSolverMultistepScheduler.from_config(
+                    pipeline.scheduler.config, **scheduler_args
+                )
 
             # load attention processors
             pipeline.load_lora_weights(args.output_dir)