DDRM

class deepinv.sampling.DDRM(self, denoiser, sigmas=np.linspace(1, 0, 100), eta=0.85, etab=1.0, verbose=False)[source]

Bases: Module

Denoising Diffusion Restoration Models (DDRM).

This class implements the denoising diffusion restoration model (DDRM) described in https://arxiv.org/abs/2201.11793.

The DDRM is a sampling method that uses a denoiser to sample from the posterior distribution of the inverse problem.

It requires that the physics operator has a singular value decomposition, i.e., it is deepinv.physics.DecomposablePhysics() class.

Parameters:

  • denoiser (torch.nn.Module) – a denoiser model that can handle different noise levels.

  • sigmas (list[int], numpy.array) – a list of noise levels to use in the diffusion, they should be in decreasing order from 1 to 0.

  • eta (float) – hyperparameter

  • etab (float) – hyperparameter

  • verbose (bool) – if True, print progress

Examples:

Denoising diffusion restoration model using a pretrained DRUNet denoiser:

>>> import deepinv as dinv
>>> device = dinv.utils.get_freer_gpu(verbose=False) if torch.cuda.is_available() else 'cpu'
>>> seed = torch.manual_seed(0) # Random seed for reproducibility
>>> seed = torch.cuda.manual_seed(0) # Random seed for reproducibility on GPU
>>> x = 0.5 * torch.ones(1, 3, 32, 32, device=device) # Define plain gray 32x32 image
>>> physics = dinv.physics.Inpainting(
...   mask=0.5, tensor_size=(3, 32, 32),
...   noise_model=dinv.physics.GaussianNoise(0.1),
...   device=device,
... )
>>> y = physics(x) # measurements
>>> denoiser = dinv.models.DRUNet(pretrained="download").to(device)
>>> model = dinv.sampling.DDRM(denoiser=denoiser, sigmas=np.linspace(1, 0, 10), verbose=True) # define the DDRM model
>>> xhat = model(y, physics) # sample from the posterior distribution
>>> dinv.metric.PSNR()(xhat, x) > dinv.metric.PSNR()(y, x) # Should be closer to the original
tensor([True])
forward(y, physics: DecomposablePhysics, seed=None)[source]

Runs the diffusion to obtain a random sample of the posterior distribution.

Parameters:

Examples using DDRM:

Creating a forward operator.

Creating a forward operator.

Image reconstruction with a diffusion model

Image reconstruction with a diffusion model