Uncertainty quantification with PnP-ULA.

This code shows you how to use sampling algorithms to quantify uncertainty of a reconstruction from incomplete and noisy measurements.

ULA obtains samples by running the following iteration:

\[x_{k+1} = x_k + \alpha \eta \nabla \log p_{\sigma}(x_k) + \eta \nabla \log p(y|x_k) + \sqrt{2 \eta} z_k\]

where \(z_k \sim \mathcal{N}(0, I)\) is a Gaussian random variable, \(\eta\) is the step size and \(\alpha\) is a parameter controlling the regularization.

The PnP-ULA method is described in the paper Laumont et al.[1].

import deepinv as dinv
from deepinv.utils.plotting import plot
import torch
from deepinv.utils import load_example

Load image from the internet

This example uses an image of Messi.

device = dinv.utils.get_device()

x = load_example("messi.jpg", img_size=32).to(device)

Selected GPU 0 with 6547.625 MiB free memory

Define forward operator and noise model

This example uses inpainting as the forward operator and Gaussian noise as the noise model.

sigma = 0.1  # noise level
physics = dinv.physics.Inpainting(mask=0.5, img_size=x.shape[1:], device=device)
physics.noise_model = dinv.physics.GaussianNoise(sigma=sigma)

# Set the global random seed from pytorch to ensure reproducibility of the example.
torch.manual_seed(0)

<torch._C.Generator object at 0x7f82f894f3d0>

Define the likelihood

Since the noise model is Gaussian, the negative log-likelihood is the L2 loss.

\[-\log p(y|x) \propto \frac{1}{2\sigma^2} \|y-Ax\|^2\]

# load Gaussian Likelihood
likelihood = dinv.optim.data_fidelity.L2(sigma=sigma)

Define the prior

The score a distribution can be approximated using Tweedie’s formula via the deepinv.optim.ScorePrior class.

\[\nabla \log p_{\sigma}(x) \approx \frac{1}{\sigma^2} \left(D(x,\sigma)-x\right)\]

This example uses a pretrained DnCNN model. From a Bayesian point of view, the score plays the role of the gradient of the negative log prior The hyperparameter sigma_denoiser (\(sigma\)) controls the strength of the prior.

In this example, we use a pretrained DnCNN model using the deepinv.loss.FNEJacobianSpectralNorm loss, which makes sure that the denoiser is firmly non-expansive (see Terris et al.[2]), and helps to stabilize the sampling algorithm.

sigma_denoiser = 2 / 255
prior = dinv.optim.ScorePrior(
    denoiser=dinv.models.DnCNN(pretrained="download_lipschitz")
).to(device)

Create the MCMC sampler

Here we use the Unadjusted Langevin Algorithm (ULA) to sample from the posterior defined in deepinv.sampling.ULAIterator. The hyperparameter step_size controls the step size of the MCMC sampler, regularization controls the strength of the prior and iterations controls the number of iterations of the sampler.

regularization = 0.9
step_size = 0.01 * (sigma**2)
iterations = int(5e3) if torch.cuda.is_available() else 10
params = {
    "step_size": step_size,
    "alpha": regularization,
    "sigma": sigma_denoiser,
}
f = dinv.sampling.sampling_builder(
    "ULA",
    prior=prior,
    data_fidelity=likelihood,
    max_iter=iterations,
    params_algo=params,
    thinning=1,
    verbose=True,
)

Generate the measurement

We apply the forward model to generate the noisy measurement.

y = physics(x)

Run sampling algorithm and plot results

The sampling algorithm returns the posterior mean and variance. We compare the posterior mean with a simple linear reconstruction.

mean, var = f.sample(y, physics)

# compute linear inverse
x_lin = physics.A_adjoint(y)

# compute PSNR
print(f"Linear reconstruction PSNR: {dinv.metric.PSNR()(x, x_lin).item():.2f} dB")
print(f"Posterior mean PSNR: {dinv.metric.PSNR()(x, mean).item():.2f} dB")

# plot results
error = (mean - x).abs().sum(dim=1).unsqueeze(1)  # per pixel average abs. error
std = var.sum(dim=1).unsqueeze(1).sqrt()  # per pixel average standard dev.
imgs = [x_lin, x, mean, std / std.flatten().max(), error / error.flatten().max()]
plot(
    imgs,
    titles=["measurement", "ground truth", "post. mean", "post. std", "abs. error"],
)

  0%|          | 0/5000 [00:00<?, ?it/s]
  1%|▏         | 66/5000 [00:00<00:07, 651.00it/s]
  3%|▎         | 137/5000 [00:00<00:07, 681.85it/s]
  4%|▍         | 221/5000 [00:00<00:06, 752.16it/s]
  6%|▌         | 305/5000 [00:00<00:05, 786.57it/s]
  8%|▊         | 389/5000 [00:00<00:05, 805.35it/s]
  9%|▉         | 473/5000 [00:00<00:05, 816.73it/s]
 11%|█         | 557/5000 [00:00<00:05, 821.39it/s]
 13%|█▎        | 641/5000 [00:00<00:05, 826.42it/s]
 14%|█▍        | 725/5000 [00:00<00:05, 830.38it/s]
 16%|█▌        | 809/5000 [00:01<00:05, 832.59it/s]
 18%|█▊        | 893/5000 [00:01<00:04, 827.77it/s]
 20%|█▉        | 976/5000 [00:01<00:04, 810.69it/s]
 21%|██        | 1058/5000 [00:01<00:04, 795.58it/s]
 23%|██▎       | 1138/5000 [00:01<00:04, 789.93it/s]
 24%|██▍       | 1218/5000 [00:01<00:04, 791.93it/s]
 26%|██▌       | 1298/5000 [00:01<00:04, 793.62it/s]
 28%|██▊       | 1378/5000 [00:01<00:04, 795.11it/s]
 29%|██▉       | 1458/5000 [00:01<00:04, 795.44it/s]
 31%|███       | 1538/5000 [00:01<00:04, 796.36it/s]
 32%|███▏      | 1618/5000 [00:02<00:04, 795.94it/s]
 34%|███▍      | 1698/5000 [00:02<00:04, 795.59it/s]
 36%|███▌      | 1778/5000 [00:02<00:04, 795.83it/s]
 37%|███▋      | 1858/5000 [00:02<00:03, 794.85it/s]
 39%|███▉      | 1938/5000 [00:02<00:03, 795.00it/s]
 40%|████      | 2018/5000 [00:02<00:03, 796.11it/s]
 42%|████▏     | 2098/5000 [00:02<00:03, 795.89it/s]
 44%|████▎     | 2178/5000 [00:02<00:03, 793.40it/s]
 45%|████▌     | 2258/5000 [00:02<00:03, 794.27it/s]
 47%|████▋     | 2338/5000 [00:02<00:03, 794.55it/s]
 48%|████▊     | 2418/5000 [00:03<00:03, 793.93it/s]
 50%|████▉     | 2498/5000 [00:03<00:03, 794.90it/s]
 52%|█████▏    | 2578/5000 [00:03<00:03, 795.67it/s]
 53%|█████▎    | 2658/5000 [00:03<00:02, 792.84it/s]
 55%|█████▍    | 2738/5000 [00:03<00:02, 779.31it/s]
 56%|█████▋    | 2816/5000 [00:03<00:02, 769.67it/s]
 58%|█████▊    | 2894/5000 [00:03<00:02, 768.38it/s]
 59%|█████▉    | 2974/5000 [00:03<00:02, 776.95it/s]
 61%|██████    | 3054/5000 [00:03<00:02, 781.40it/s]
 63%|██████▎   | 3134/5000 [00:03<00:02, 786.64it/s]
 64%|██████▍   | 3214/5000 [00:04<00:02, 790.25it/s]
 66%|██████▌   | 3294/5000 [00:04<00:02, 791.24it/s]
 67%|██████▋   | 3374/5000 [00:04<00:02, 793.48it/s]
 69%|██████▉   | 3454/5000 [00:04<00:01, 790.85it/s]
 71%|███████   | 3534/5000 [00:04<00:01, 785.50it/s]
 72%|███████▏  | 3614/5000 [00:04<00:01, 787.58it/s]
 74%|███████▍  | 3694/5000 [00:04<00:01, 788.37it/s]
 75%|███████▌  | 3774/5000 [00:04<00:01, 791.08it/s]
 77%|███████▋  | 3855/5000 [00:04<00:01, 793.77it/s]
 79%|███████▊  | 3935/5000 [00:04<00:01, 795.16it/s]
 80%|████████  | 4015/5000 [00:05<00:01, 795.88it/s]
 82%|████████▏ | 4095/5000 [00:05<00:01, 796.15it/s]
 84%|████████▎ | 4175/5000 [00:05<00:01, 796.92it/s]
 85%|████████▌ | 4255/5000 [00:05<00:00, 795.19it/s]
 87%|████████▋ | 4335/5000 [00:05<00:00, 796.45it/s]
 88%|████████▊ | 4415/5000 [00:05<00:00, 796.85it/s]
 90%|████████▉ | 4495/5000 [00:05<00:00, 785.79it/s]
 91%|█████████▏| 4574/5000 [00:05<00:00, 773.60it/s]
 93%|█████████▎| 4652/5000 [00:05<00:00, 769.05it/s]
 95%|█████████▍| 4731/5000 [00:05<00:00, 773.50it/s]
 96%|█████████▌| 4811/5000 [00:06<00:00, 781.10it/s]
 98%|█████████▊| 4892/5000 [00:06<00:00, 786.77it/s]
 99%|█████████▉| 4972/5000 [00:06<00:00, 789.76it/s]
100%|██████████| 5000/5000 [00:06<00:00, 791.62it/s]
Iteration 4999, current converge crit. = 1.42E-05, objective = 1.00E-03
Iteration 4999, current converge crit. = 3.42E-04, objective = 1.00E-03
Linear reconstruction PSNR: 8.55 dB
Posterior mean PSNR: 22.31 dB

References:

[1]

Rémi Laumont, Valentin De Bortoli, Andrés Almansa, Julie Delon, Alain Durmus, and Marcelo Pereyra. Bayesian imaging using plug & play priors: when langevin meets tweedie. SIAM Journal on Imaging Sciences, 15(2):701–737, 2022.

[2]

Matthieu Terris, Audrey Repetti, Jean-Christophe Pesquet, and Yves Wiaux. Building firmly nonexpansive convolutional neural networks. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 8658–8662. IEEE, 2020.