Abstract
DDPM model trained on huggan/anime-faces dataset.
Training Arguments
Argument | Value |
---|---|
image_size | 64 |
train_batch_size | 16 |
eval_batch_size | 16 |
num_epochs | 50 |
gradient_accumulation_steps | 1 |
learning_rate | 1e-4 |
lr_warmup_steps | 500 |
mixed_precision | "fp16" |
For training code, please refer to this link.
Inference
This project aims to implement DDPM from scratch, so DDPMScheduler
is not used. Instead, I use only UNet2DModel
and implement a simple scheduler myself. The inference code is:
import torch
from tqdm import tqdm
from diffusers import UNet2DModel
class DDPM:
def __init__(
self,
num_train_timesteps:int = 1000,
beta_start: float = 0.0001,
beta_end: float = 0.02,
):
self.num_train_timesteps = num_train_timesteps
self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32)
self.alphas = 1.0 - self.betas
self.alphas_cumprod = torch.cumprod(self.alphas, dim=0)
self.timesteps = torch.arange(num_train_timesteps - 1, -1, -1)
def add_noise(
self,
original_samples: torch.Tensor,
noise: torch.Tensor,
timesteps: torch.Tensor,
):
alphas_cumprod = self.alphas_cumprod.to(device=original_samples.device ,dtype=original_samples.dtype)
noise = noise.to(original_samples.device)
timesteps = timesteps.to(original_samples.device)
# \sqrt{\bar\alpha_t}
sqrt_alpha_prod = alphas_cumprod[timesteps].flatten() ** 0.5
while len(sqrt_alpha_prod.shape) < len(original_samples.shape):
sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1)
# \sqrt{1 - \bar\alpha_t}
sqrt_one_minus_alpha_prod = (1.0 - alphas_cumprod[timesteps]).flatten() ** 0.5
while len(sqrt_one_minus_alpha_prod.shape) < len(original_samples.shape):
sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1)
return sqrt_alpha_prod * original_samples + sqrt_one_minus_alpha_prod * noise
@torch.no_grad()
def sample(
self,
unet: UNet2DModel,
batch_size: int,
in_channels: int,
sample_size: int,
):
betas = self.betas.to(unet.device)
alphas = self.alphas.to(unet.device)
alphas_cumprod = self.alphas_cumprod.to(unet.device)
timesteps = self.timesteps.to(unet.device)
images = torch.randn((batch_size, in_channels, sample_size, sample_size), device=unet.device)
for timestep in tqdm(timesteps, desc='Sampling'):
pred_noise: torch.Tensor = unet(images, timestep).sample
# mean of q(x_{t-1}|x_t)
alpha_t = alphas[timestep]
alpha_cumprod_t = alphas_cumprod[timestep]
sqrt_alpha_t = alpha_t ** 0.5
one_minus_alpha_t = 1.0 - alpha_t
sqrt_one_minus_alpha_cumprod_t = (1 - alpha_cumprod_t) ** 0.5
mean = (images - one_minus_alpha_t / sqrt_one_minus_alpha_cumprod_t * pred_noise) / sqrt_alpha_t
# variance of q(x_{t-1}|x_t)
if timestep > 1:
beta_t = betas[timestep]
one_minus_alpha_cumprod_t_minus_one = 1.0 - alphas_cumprod[timestep - 1]
one_divided_by_sigma_square = alpha_t / beta_t + 1.0 / one_minus_alpha_cumprod_t_minus_one
variance = (1.0 / one_divided_by_sigma_square) ** 0.5
else:
variance = torch.zeros_like(timestep)
epsilon = torch.randn_like(images)
images = mean + variance * epsilon
images = (images / 2.0 + 0.5).clamp(0, 1).cpu().permute(0, 2, 3, 1).numpy()
return images
model = UNet2DModel.from_pretrained('ddpm-animefaces-64').cuda()
ddpm = DDPM()
images = ddpm.sample(model, 32, 3, 64)
from diffusers.utils import make_image_grid, numpy_to_pil
image_grid = make_image_grid(numpy_to_pil(images), rows=4, cols=8)
image_grid.save('ddpm-sample-results.png')
This can also be found in this link.
- Downloads last month
- 54
Inference API (serverless) does not yet support diffusers models for this pipeline type.