AI/diffusers
1
0
Fork 0
mirror of https://github.com/huggingface/diffusers.git synced 2026-01-27 17:22:53 +03:00
Code Activity

[Examples] add compute_snr() to training utils. (#5188)

Browse Source 
add compute_snr() to training utils.
This commit is contained in:
Sayak Paul
2023-09-27 21:42:20 +05:30
committed by GitHub
parent ba59e92fb0
commit cdcc01be0e
11 changed files with 46 additions and 256 deletions
Download Patch File Download Diff File Expand all files Collapse all files

27
examples/dreambooth/train_dreambooth.py
View File

@@ -52,6 +52,7 @@ from diffusers import (
UNet2DConditionModel,
)
from diffusers.optimization import get_scheduler
from diffusers.training_utils import compute_snr
from diffusers.utils import check_min_version, is_wandb_available
from diffusers.utils.import_utils import is_xformers_available
@@ -224,30 +225,6 @@ def import_model_class_from_model_name_or_path(pretrained_model_name_or_path: st
raise ValueError(f"{model_class} is not supported.")
def compute_snr(timesteps, noise_scheduler):
"""
Computes SNR as per https://github.com/TiankaiHang/Min-SNR-Diffusion-Training/blob/521b624bd70c67cee4bdf49225915f5945a872e3/guided_diffusion/gaussian_diffusion.py#L847-L849
"""
alphas_cumprod = noise_scheduler.alphas_cumprod
sqrt_alphas_cumprod = alphas_cumprod**0.5
sqrt_one_minus_alphas_cumprod = (1.0 - alphas_cumprod) ** 0.5
# Expand the tensors.
# Adapted from https://github.com/TiankaiHang/Min-SNR-Diffusion-Training/blob/521b624bd70c67cee4bdf49225915f5945a872e3/guided_diffusion/gaussian_diffusion.py#L1026
sqrt_alphas_cumprod = sqrt_alphas_cumprod.to(device=timesteps.device)[timesteps].float()
while len(sqrt_alphas_cumprod.shape) < len(timesteps.shape):
sqrt_alphas_cumprod = sqrt_alphas_cumprod[..., None]
alpha = sqrt_alphas_cumprod.expand(timesteps.shape)
sqrt_one_minus_alphas_cumprod = sqrt_one_minus_alphas_cumprod.to(device=timesteps.device)[timesteps].float()
while len(sqrt_one_minus_alphas_cumprod.shape) < len(timesteps.shape):
sqrt_one_minus_alphas_cumprod = sqrt_one_minus_alphas_cumprod[..., None]
sigma = sqrt_one_minus_alphas_cumprod.expand(timesteps.shape)
# Compute SNR
snr = (alpha / sigma) ** 2
return snr
def parse_args(input_args=None):
parser = argparse.ArgumentParser(description="Simple example of a training script.")
parser.add_argument(
@@ -1302,7 +1279,7 @@ def main(args):
# 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.
# This is discussed in Section 4.2 of the same paper.
snr = compute_snr(timesteps, noise_scheduler)
snr = compute_snr(noise_scheduler, timesteps)
base_weight = (
torch.stack([snr, args.snr_gamma * torch.ones_like(timesteps)], dim=1).min(dim=1)[0] / snr
)

28
examples/kandinsky2_2/text_to_image/train_text_to_image_decoder.py
View File

@@ -42,7 +42,7 @@ from transformers.utils import ContextManagers
import diffusers
from diffusers import AutoPipelineForText2Image, DDPMScheduler, UNet2DConditionModel, VQModel
from diffusers.optimization import get_scheduler
from diffusers.training_utils import EMAModel
from diffusers.training_utils import EMAModel, compute_snr
from diffusers.utils import check_min_version, is_wandb_available, make_image_grid
from diffusers.utils.import_utils import is_xformers_available
@@ -530,30 +530,6 @@ def main():
else:
raise ValueError("xformers is not available. Make sure it is installed correctly")
def compute_snr(timesteps):
"""
Computes SNR as per https://github.com/TiankaiHang/Min-SNR-Diffusion-Training/blob/521b624bd70c67cee4bdf49225915f5945a872e3/guided_diffusion/gaussian_diffusion.py#L847-L849
"""
alphas_cumprod = noise_scheduler.alphas_cumprod
sqrt_alphas_cumprod = alphas_cumprod**0.5
sqrt_one_minus_alphas_cumprod = (1.0 - alphas_cumprod) ** 0.5
# Expand the tensors.
# Adapted from https://github.com/TiankaiHang/Min-SNR-Diffusion-Training/blob/521b624bd70c67cee4bdf49225915f5945a872e3/guided_diffusion/gaussian_diffusion.py#L1026
sqrt_alphas_cumprod = sqrt_alphas_cumprod.to(device=timesteps.device)[timesteps].float()
while len(sqrt_alphas_cumprod.shape) < len(timesteps.shape):
sqrt_alphas_cumprod = sqrt_alphas_cumprod[..., None]
alpha = sqrt_alphas_cumprod.expand(timesteps.shape)
sqrt_one_minus_alphas_cumprod = sqrt_one_minus_alphas_cumprod.to(device=timesteps.device)[timesteps].float()
while len(sqrt_one_minus_alphas_cumprod.shape) < len(timesteps.shape):
sqrt_one_minus_alphas_cumprod = sqrt_one_minus_alphas_cumprod[..., None]
sigma = sqrt_one_minus_alphas_cumprod.expand(timesteps.shape)
# Compute SNR.
snr = (alpha / sigma) ** 2
return snr
# `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
@@ -800,7 +776,7 @@ def main():
# 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.
# This is discussed in Section 4.2 of the same paper.
snr = compute_snr(timesteps)
snr = compute_snr(noise_scheduler, timesteps)
base_weight = (
torch.stack([snr, args.snr_gamma * torch.ones_like(timesteps)], dim=1).min(dim=1)[0] / snr
)

27
examples/kandinsky2_2/text_to_image/train_text_to_image_lora_decoder.py
View File

@@ -41,6 +41,7 @@ from diffusers import AutoPipelineForText2Image, DDPMScheduler, UNet2DConditionM
from diffusers.loaders import AttnProcsLayers
from diffusers.models.attention_processor import LoRAAttnAddedKVProcessor
from diffusers.optimization import get_scheduler
from diffusers.training_utils import compute_snr
from diffusers.utils import check_min_version, is_wandb_available
@@ -419,30 +420,6 @@ def main():
unet.set_attn_processor(lora_attn_procs)
def compute_snr(timesteps):
"""
Computes SNR as per https://github.com/TiankaiHang/Min-SNR-Diffusion-Training/blob/521b624bd70c67cee4bdf49225915f5945a872e3/guided_diffusion/gaussian_diffusion.py#L847-L849
"""
alphas_cumprod = noise_scheduler.alphas_cumprod
sqrt_alphas_cumprod = alphas_cumprod**0.5
sqrt_one_minus_alphas_cumprod = (1.0 - alphas_cumprod) ** 0.5
# Expand the tensors.
# Adapted from https://github.com/TiankaiHang/Min-SNR-Diffusion-Training/blob/521b624bd70c67cee4bdf49225915f5945a872e3/guided_diffusion/gaussian_diffusion.py#L1026
sqrt_alphas_cumprod = sqrt_alphas_cumprod.to(device=timesteps.device)[timesteps].float()
while len(sqrt_alphas_cumprod.shape) < len(timesteps.shape):
sqrt_alphas_cumprod = sqrt_alphas_cumprod[..., None]
alpha = sqrt_alphas_cumprod.expand(timesteps.shape)
sqrt_one_minus_alphas_cumprod = sqrt_one_minus_alphas_cumprod.to(device=timesteps.device)[timesteps].float()
while len(sqrt_one_minus_alphas_cumprod.shape) < len(timesteps.shape):
sqrt_one_minus_alphas_cumprod = sqrt_one_minus_alphas_cumprod[..., None]
sigma = sqrt_one_minus_alphas_cumprod.expand(timesteps.shape)
# Compute SNR.
snr = (alpha / sigma) ** 2
return snr
lora_layers = AttnProcsLayers(unet.attn_processors)
if args.allow_tf32:
@@ -653,7 +630,7 @@ def main():
# 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.
# This is discussed in Section 4.2 of the same paper.
snr = compute_snr(timesteps)
snr = compute_snr(noise_scheduler, timesteps)
base_weight = (
torch.stack([snr, args.snr_gamma * torch.ones_like(timesteps)], dim=1).min(dim=1)[0] / snr
)

28
examples/kandinsky2_2/text_to_image/train_text_to_image_lora_prior.py
View File

@@ -41,6 +41,7 @@ from diffusers import AutoPipelineForText2Image, DDPMScheduler, PriorTransformer
from diffusers.loaders import AttnProcsLayers
from diffusers.models.attention_processor import LoRAAttnProcessor
from diffusers.optimization import get_scheduler
from diffusers.training_utils import compute_snr
from diffusers.utils import check_min_version, is_wandb_available
@@ -413,31 +414,6 @@ def main():
lora_attn_procs[name] = LoRAAttnProcessor(hidden_size=2048, rank=args.rank)
prior.set_attn_processor(lora_attn_procs)
def compute_snr(timesteps):
"""
Computes SNR as per https://github.com/TiankaiHang/Min-SNR-Diffusion-Training/blob/521b624bd70c67cee4bdf49225915f5945a872e3/guided_diffusion/gaussian_diffusion.py#L847-L849
"""
alphas_cumprod = noise_scheduler.alphas_cumprod
sqrt_alphas_cumprod = alphas_cumprod**0.5
sqrt_one_minus_alphas_cumprod = (1.0 - alphas_cumprod) ** 0.5
# Expand the tensors.
# Adapted from https://github.com/TiankaiHang/Min-SNR-Diffusion-Training/blob/521b624bd70c67cee4bdf49225915f5945a872e3/guided_diffusion/gaussian_diffusion.py#L1026
sqrt_alphas_cumprod = sqrt_alphas_cumprod.to(device=timesteps.device)[timesteps].float()
while len(sqrt_alphas_cumprod.shape) < len(timesteps.shape):
sqrt_alphas_cumprod = sqrt_alphas_cumprod[..., None]
alpha = sqrt_alphas_cumprod.expand(timesteps.shape)
sqrt_one_minus_alphas_cumprod = sqrt_one_minus_alphas_cumprod.to(device=timesteps.device)[timesteps].float()
while len(sqrt_one_minus_alphas_cumprod.shape) < len(timesteps.shape):
sqrt_one_minus_alphas_cumprod = sqrt_one_minus_alphas_cumprod[..., None]
sigma = sqrt_one_minus_alphas_cumprod.expand(timesteps.shape)
# Compute SNR.
snr = (alpha / sigma) ** 2
return snr
lora_layers = AttnProcsLayers(prior.attn_processors)
if args.allow_tf32:
@@ -684,7 +660,7 @@ def main():
# 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.
# This is discussed in Section 4.2 of the same paper.
snr = compute_snr(timesteps)
snr = compute_snr(noise_scheduler, timesteps)
base_weight = (
torch.stack([snr, args.snr_gamma * torch.ones_like(timesteps)], dim=1).min(dim=1)[0] / snr
)

28
examples/kandinsky2_2/text_to_image/train_text_to_image_prior.py
View File

@@ -42,7 +42,7 @@ from transformers.utils import ContextManagers
import diffusers
from diffusers import AutoPipelineForText2Image, DDPMScheduler, PriorTransformer
from diffusers.optimization import get_scheduler
from diffusers.training_utils import EMAModel
from diffusers.training_utils import EMAModel, compute_snr
from diffusers.utils import check_min_version, is_wandb_available, make_image_grid
@@ -523,30 +523,6 @@ def main():
ema_prior = EMAModel(ema_prior.parameters(), model_cls=PriorTransformer, model_config=ema_prior.config)
ema_prior.to(accelerator.device)
def compute_snr(timesteps):
"""
Computes SNR as per https://github.com/TiankaiHang/Min-SNR-Diffusion-Training/blob/521b624bd70c67cee4bdf49225915f5945a872e3/guided_diffusion/gaussian_diffusion.py#L847-L849
"""
alphas_cumprod = noise_scheduler.alphas_cumprod
sqrt_alphas_cumprod = alphas_cumprod**0.5
sqrt_one_minus_alphas_cumprod = (1.0 - alphas_cumprod) ** 0.5
# Expand the tensors.
# Adapted from https://github.com/TiankaiHang/Min-SNR-Diffusion-Training/blob/521b624bd70c67cee4bdf49225915f5945a872e3/guided_diffusion/gaussian_diffusion.py#L1026
sqrt_alphas_cumprod = sqrt_alphas_cumprod.to(device=timesteps.device)[timesteps].float()
while len(sqrt_alphas_cumprod.shape) < len(timesteps.shape):
sqrt_alphas_cumprod = sqrt_alphas_cumprod[..., None]
alpha = sqrt_alphas_cumprod.expand(timesteps.shape)
sqrt_one_minus_alphas_cumprod = sqrt_one_minus_alphas_cumprod.to(device=timesteps.device)[timesteps].float()
while len(sqrt_one_minus_alphas_cumprod.shape) < len(timesteps.shape):
sqrt_one_minus_alphas_cumprod = sqrt_one_minus_alphas_cumprod[..., None]
sigma = sqrt_one_minus_alphas_cumprod.expand(timesteps.shape)
# Compute SNR.
snr = (alpha / sigma) ** 2
return snr
# `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
@@ -832,7 +808,7 @@ def main():
# 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.
# This is discussed in Section 4.2 of the same paper.
snr = compute_snr(timesteps)
snr = compute_snr(noise_scheduler, timesteps)
base_weight = (
torch.stack([snr, args.snr_gamma * torch.ones_like(timesteps)], dim=1).min(dim=1)[0] / snr
)

28
examples/research_projects/onnxruntime/text_to_image/train_text_to_image.py
View File

@@ -44,7 +44,7 @@ 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
from diffusers.training_utils import EMAModel, compute_snr
from diffusers.utils import check_min_version, deprecate, is_wandb_available
from diffusers.utils.import_utils import is_xformers_available
@@ -524,30 +524,6 @@ def main():
else:
raise ValueError("xformers is not available. Make sure it is installed correctly")
def compute_snr(timesteps):
"""
Computes SNR as per https://github.com/TiankaiHang/Min-SNR-Diffusion-Training/blob/521b624bd70c67cee4bdf49225915f5945a872e3/guided_diffusion/gaussian_diffusion.py#L847-L849
"""
alphas_cumprod = noise_scheduler.alphas_cumprod
sqrt_alphas_cumprod = alphas_cumprod**0.5
sqrt_one_minus_alphas_cumprod = (1.0 - alphas_cumprod) ** 0.5
# Expand the tensors.
# Adapted from https://github.com/TiankaiHang/Min-SNR-Diffusion-Training/blob/521b624bd70c67cee4bdf49225915f5945a872e3/guided_diffusion/gaussian_diffusion.py#L1026
sqrt_alphas_cumprod = sqrt_alphas_cumprod.to(device=timesteps.device)[timesteps].float()
while len(sqrt_alphas_cumprod.shape) < len(timesteps.shape):
sqrt_alphas_cumprod = sqrt_alphas_cumprod[..., None]
alpha = sqrt_alphas_cumprod.expand(timesteps.shape)
sqrt_one_minus_alphas_cumprod = sqrt_one_minus_alphas_cumprod.to(device=timesteps.device)[timesteps].float()
while len(sqrt_one_minus_alphas_cumprod.shape) < len(timesteps.shape):
sqrt_one_minus_alphas_cumprod = sqrt_one_minus_alphas_cumprod[..., None]
sigma = sqrt_one_minus_alphas_cumprod.expand(timesteps.shape)
# Compute SNR.
snr = (alpha / sigma) ** 2
return snr
# `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
@@ -871,7 +847,7 @@ def main():
# 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.
# This is discussed in Section 4.2 of the same paper.
snr = compute_snr(timesteps)
snr = compute_snr(noise_scheduler, timesteps)
base_weight = (
torch.stack([snr, args.snr_gamma * torch.ones_like(timesteps)], dim=1).min(dim=1)[0] / snr
)

28
examples/text_to_image/train_text_to_image.py
View File

@@ -43,7 +43,7 @@ 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
from diffusers.training_utils import EMAModel, compute_snr
from diffusers.utils import check_min_version, deprecate, is_wandb_available, make_image_grid
from diffusers.utils.import_utils import is_xformers_available
@@ -601,30 +601,6 @@ def main():
else:
raise ValueError("xformers is not available. Make sure it is installed correctly")
def compute_snr(timesteps):
"""
Computes SNR as per https://github.com/TiankaiHang/Min-SNR-Diffusion-Training/blob/521b624bd70c67cee4bdf49225915f5945a872e3/guided_diffusion/gaussian_diffusion.py#L847-L849
"""
alphas_cumprod = noise_scheduler.alphas_cumprod
sqrt_alphas_cumprod = alphas_cumprod**0.5
sqrt_one_minus_alphas_cumprod = (1.0 - alphas_cumprod) ** 0.5
# Expand the tensors.
# Adapted from https://github.com/TiankaiHang/Min-SNR-Diffusion-Training/blob/521b624bd70c67cee4bdf49225915f5945a872e3/guided_diffusion/gaussian_diffusion.py#L1026
sqrt_alphas_cumprod = sqrt_alphas_cumprod.to(device=timesteps.device)[timesteps].float()
while len(sqrt_alphas_cumprod.shape) < len(timesteps.shape):
sqrt_alphas_cumprod = sqrt_alphas_cumprod[..., None]
alpha = sqrt_alphas_cumprod.expand(timesteps.shape)
sqrt_one_minus_alphas_cumprod = sqrt_one_minus_alphas_cumprod.to(device=timesteps.device)[timesteps].float()
while len(sqrt_one_minus_alphas_cumprod.shape) < len(timesteps.shape):
sqrt_one_minus_alphas_cumprod = sqrt_one_minus_alphas_cumprod[..., None]
sigma = sqrt_one_minus_alphas_cumprod.expand(timesteps.shape)
# Compute SNR.
snr = (alpha / sigma) ** 2
return snr
# `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
@@ -951,7 +927,7 @@ def main():
# 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.
# This is discussed in Section 4.2 of the same paper.
snr = compute_snr(timesteps)
snr = compute_snr(noise_scheduler, timesteps)
base_weight = (
torch.stack([snr, args.snr_gamma * torch.ones_like(timesteps)], dim=1).min(dim=1)[0] / snr
)

27
examples/text_to_image/train_text_to_image_lora.py
View File

@@ -43,6 +43,7 @@ from diffusers import AutoencoderKL, DDPMScheduler, DiffusionPipeline, UNet2DCon
from diffusers.loaders import AttnProcsLayers
from diffusers.models.attention_processor import LoRAAttnProcessor
from diffusers.optimization import get_scheduler
from diffusers.training_utils import compute_snr
from diffusers.utils import check_min_version, is_wandb_available
from diffusers.utils.import_utils import is_xformers_available
@@ -491,30 +492,6 @@ def main():
else:
raise ValueError("xformers is not available. Make sure it is installed correctly")
def compute_snr(timesteps):
"""
Computes SNR as per https://github.com/TiankaiHang/Min-SNR-Diffusion-Training/blob/521b624bd70c67cee4bdf49225915f5945a872e3/guided_diffusion/gaussian_diffusion.py#L847-L849
"""
alphas_cumprod = noise_scheduler.alphas_cumprod
sqrt_alphas_cumprod = alphas_cumprod**0.5
sqrt_one_minus_alphas_cumprod = (1.0 - alphas_cumprod) ** 0.5
# Expand the tensors.
# Adapted from https://github.com/TiankaiHang/Min-SNR-Diffusion-Training/blob/521b624bd70c67cee4bdf49225915f5945a872e3/guided_diffusion/gaussian_diffusion.py#L1026
sqrt_alphas_cumprod = sqrt_alphas_cumprod.to(device=timesteps.device)[timesteps].float()
while len(sqrt_alphas_cumprod.shape) < len(timesteps.shape):
sqrt_alphas_cumprod = sqrt_alphas_cumprod[..., None]
alpha = sqrt_alphas_cumprod.expand(timesteps.shape)
sqrt_one_minus_alphas_cumprod = sqrt_one_minus_alphas_cumprod.to(device=timesteps.device)[timesteps].float()
while len(sqrt_one_minus_alphas_cumprod.shape) < len(timesteps.shape):
sqrt_one_minus_alphas_cumprod = sqrt_one_minus_alphas_cumprod[..., None]
sigma = sqrt_one_minus_alphas_cumprod.expand(timesteps.shape)
# Compute SNR.
snr = (alpha / sigma) ** 2
return snr
lora_layers = AttnProcsLayers(unet.attn_processors)
# Enable TF32 for faster training on Ampere GPUs,
@@ -782,7 +759,7 @@ def main():
# 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.
# This is discussed in Section 4.2 of the same paper.
snr = compute_snr(timesteps)
snr = compute_snr(noise_scheduler, timesteps)
base_weight = (
torch.stack([snr, args.snr_gamma * torch.ones_like(timesteps)], dim=1).min(dim=1)[0] / snr
)

27
examples/text_to_image/train_text_to_image_lora_sdxl.py
View File

@@ -52,6 +52,7 @@ from diffusers import (
from diffusers.loaders import LoraLoaderMixin, text_encoder_lora_state_dict
from diffusers.models.attention_processor import LoRAAttnProcessor, LoRAAttnProcessor2_0
from diffusers.optimization import get_scheduler
from diffusers.training_utils import compute_snr
from diffusers.utils import check_min_version, is_wandb_available
from diffusers.utils.import_utils import is_xformers_available
@@ -632,30 +633,6 @@ def main(args):
unet.set_attn_processor(unet_lora_attn_procs)
def compute_snr(timesteps):
"""
Computes SNR as per https://github.com/TiankaiHang/Min-SNR-Diffusion-Training/blob/521b624bd70c67cee4bdf49225915f5945a872e3/guided_diffusion/gaussian_diffusion.py#L847-L849
"""
alphas_cumprod = noise_scheduler.alphas_cumprod
sqrt_alphas_cumprod = alphas_cumprod**0.5
sqrt_one_minus_alphas_cumprod = (1.0 - alphas_cumprod) ** 0.5
# Expand the tensors.
# Adapted from https://github.com/TiankaiHang/Min-SNR-Diffusion-Training/blob/521b624bd70c67cee4bdf49225915f5945a872e3/guided_diffusion/gaussian_diffusion.py#L1026
sqrt_alphas_cumprod = sqrt_alphas_cumprod.to(device=timesteps.device)[timesteps].float()
while len(sqrt_alphas_cumprod.shape) < len(timesteps.shape):
sqrt_alphas_cumprod = sqrt_alphas_cumprod[..., None]
alpha = sqrt_alphas_cumprod.expand(timesteps.shape)
sqrt_one_minus_alphas_cumprod = sqrt_one_minus_alphas_cumprod.to(device=timesteps.device)[timesteps].float()
while len(sqrt_one_minus_alphas_cumprod.shape) < len(timesteps.shape):
sqrt_one_minus_alphas_cumprod = sqrt_one_minus_alphas_cumprod[..., None]
sigma = sqrt_one_minus_alphas_cumprod.expand(timesteps.shape)
# Compute SNR.
snr = (alpha / sigma) ** 2
return snr
# The text encoder comes from 🤗 transformers, so we cannot directly modify it.
# So, instead, we monkey-patch the forward calls of its attention-blocks.
if args.train_text_encoder:
@@ -1071,7 +1048,7 @@ def main(args):
# 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.
# This is discussed in Section 4.2 of the same paper.
snr = compute_snr(timesteps)
snr = compute_snr(noise_scheduler, timesteps)
base_weight = (
torch.stack([snr, args.snr_gamma * torch.ones_like(timesteps)], dim=1).min(dim=1)[0] / snr
)

28
examples/text_to_image/train_text_to_image_sdxl.py
View File

@@ -51,7 +51,7 @@ from diffusers import (
UNet2DConditionModel,
)
from diffusers.optimization import get_scheduler
from diffusers.training_utils import EMAModel
from diffusers.training_utils import EMAModel, compute_snr
from diffusers.utils import check_min_version, is_wandb_available
from diffusers.utils.import_utils import is_xformers_available
@@ -692,30 +692,6 @@ def main(args):
else:
raise ValueError("xformers is not available. Make sure it is installed correctly")
def compute_snr(timesteps):
"""
Computes SNR as per https://github.com/TiankaiHang/Min-SNR-Diffusion-Training/blob/521b624bd70c67cee4bdf49225915f5945a872e3/guided_diffusion/gaussian_diffusion.py#L847-L849
"""
alphas_cumprod = noise_scheduler.alphas_cumprod
sqrt_alphas_cumprod = alphas_cumprod**0.5
sqrt_one_minus_alphas_cumprod = (1.0 - alphas_cumprod) ** 0.5
# Expand the tensors.
# Adapted from https://github.com/TiankaiHang/Min-SNR-Diffusion-Training/blob/521b624bd70c67cee4bdf49225915f5945a872e3/guided_diffusion/gaussian_diffusion.py#L1026
sqrt_alphas_cumprod = sqrt_alphas_cumprod.to(device=timesteps.device)[timesteps].float()
while len(sqrt_alphas_cumprod.shape) < len(timesteps.shape):
sqrt_alphas_cumprod = sqrt_alphas_cumprod[..., None]
alpha = sqrt_alphas_cumprod.expand(timesteps.shape)
sqrt_one_minus_alphas_cumprod = sqrt_one_minus_alphas_cumprod.to(device=timesteps.device)[timesteps].float()
while len(sqrt_one_minus_alphas_cumprod.shape) < len(timesteps.shape):
sqrt_one_minus_alphas_cumprod = sqrt_one_minus_alphas_cumprod[..., None]
sigma = sqrt_one_minus_alphas_cumprod.expand(timesteps.shape)
# Compute SNR.
snr = (alpha / sigma) ** 2
return snr
# `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
@@ -1088,7 +1064,7 @@ def main(args):
# 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.
# This is discussed in Section 4.2 of the same paper.
snr = compute_snr(timesteps)
snr = compute_snr(noise_scheduler, timesteps)
base_weight = (
torch.stack([snr, args.snr_gamma * torch.ones_like(timesteps)], dim=1).min(dim=1)[0] / snr
)

26
src/diffusers/training_utils.py
View File

@@ -26,6 +26,32 @@ def set_seed(seed: int):
# ^^ safe to call this function even if cuda is not available
def compute_snr(noise_scheduler, timesteps):
"""
Computes SNR as per
https://github.com/TiankaiHang/Min-SNR-Diffusion-Training/blob/521b624bd70c67cee4bdf49225915f5945a872e3/guided_diffusion/gaussian_diffusion.py#L847-L849
"""
alphas_cumprod = noise_scheduler.alphas_cumprod
sqrt_alphas_cumprod = alphas_cumprod**0.5
sqrt_one_minus_alphas_cumprod = (1.0 - alphas_cumprod) ** 0.5
# Expand the tensors.
# Adapted from https://github.com/TiankaiHang/Min-SNR-Diffusion-Training/blob/521b624bd70c67cee4bdf49225915f5945a872e3/guided_diffusion/gaussian_diffusion.py#L1026
sqrt_alphas_cumprod = sqrt_alphas_cumprod.to(device=timesteps.device)[timesteps].float()
while len(sqrt_alphas_cumprod.shape) < len(timesteps.shape):
sqrt_alphas_cumprod = sqrt_alphas_cumprod[..., None]
alpha = sqrt_alphas_cumprod.expand(timesteps.shape)
sqrt_one_minus_alphas_cumprod = sqrt_one_minus_alphas_cumprod.to(device=timesteps.device)[timesteps].float()
while len(sqrt_one_minus_alphas_cumprod.shape) < len(timesteps.shape):
sqrt_one_minus_alphas_cumprod = sqrt_one_minus_alphas_cumprod[..., None]
sigma = sqrt_one_minus_alphas_cumprod.expand(timesteps.shape)
# Compute SNR.
snr = (alpha / sigma) ** 2
return snr
# Adapted from torch-ema https://github.com/fadel/pytorch_ema/blob/master/torch_ema/ema.py#L14
class EMAModel:
"""