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 4759.25 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 0x7f05ba810710>

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]
  0%|          | 25/5000 [00:00<00:20, 244.90it/s]
  1%|          | 56/5000 [00:00<00:17, 281.22it/s]
  2%|▏         | 86/5000 [00:00<00:17, 288.78it/s]
  2%|▏         | 117/5000 [00:00<00:16, 296.23it/s]
  3%|▎         | 148/5000 [00:00<00:16, 300.54it/s]
  4%|▎         | 179/5000 [00:00<00:15, 302.73it/s]
  4%|▍         | 210/5000 [00:00<00:15, 304.34it/s]
  5%|▍         | 241/5000 [00:00<00:15, 305.26it/s]
  5%|▌         | 272/5000 [00:00<00:15, 305.81it/s]
  6%|▌         | 303/5000 [00:01<00:15, 304.82it/s]
  7%|▋         | 334/5000 [00:01<00:15, 305.30it/s]
  7%|▋         | 365/5000 [00:01<00:15, 305.39it/s]
  8%|▊         | 396/5000 [00:01<00:15, 305.23it/s]
  9%|▊         | 427/5000 [00:01<00:14, 304.96it/s]
  9%|▉         | 458/5000 [00:01<00:14, 305.70it/s]
 10%|▉         | 489/5000 [00:01<00:14, 306.14it/s]
 10%|█         | 520/5000 [00:01<00:14, 306.27it/s]
 11%|█         | 551/5000 [00:01<00:14, 306.43it/s]
 12%|█▏        | 582/5000 [00:01<00:14, 306.46it/s]
 12%|█▏        | 613/5000 [00:02<00:14, 306.90it/s]
 13%|█▎        | 644/5000 [00:02<00:14, 307.27it/s]
 14%|█▎        | 675/5000 [00:02<00:14, 307.54it/s]
 14%|█▍        | 706/5000 [00:02<00:13, 306.73it/s]
 15%|█▍        | 737/5000 [00:02<00:13, 306.90it/s]
 15%|█▌        | 768/5000 [00:02<00:13, 306.85it/s]
 16%|█▌        | 799/5000 [00:02<00:13, 306.76it/s]
 17%|█▋        | 830/5000 [00:02<00:13, 306.54it/s]
 17%|█▋        | 861/5000 [00:02<00:13, 304.53it/s]
 18%|█▊        | 892/5000 [00:02<00:13, 305.25it/s]
 18%|█▊        | 923/5000 [00:03<00:13, 305.84it/s]
 19%|█▉        | 954/5000 [00:03<00:13, 306.21it/s]
 20%|█▉        | 985/5000 [00:03<00:13, 306.30it/s]
 20%|██        | 1016/5000 [00:03<00:13, 303.06it/s]
 21%|██        | 1047/5000 [00:03<00:13, 299.44it/s]
 22%|██▏       | 1077/5000 [00:03<00:13, 297.11it/s]
 22%|██▏       | 1107/5000 [00:03<00:13, 295.46it/s]
 23%|██▎       | 1137/5000 [00:03<00:13, 293.87it/s]
 23%|██▎       | 1167/5000 [00:03<00:13, 293.02it/s]
 24%|██▍       | 1197/5000 [00:03<00:13, 292.53it/s]
 25%|██▍       | 1227/5000 [00:04<00:12, 292.29it/s]
 25%|██▌       | 1257/5000 [00:04<00:12, 291.49it/s]
 26%|██▌       | 1287/5000 [00:04<00:12, 291.44it/s]
 26%|██▋       | 1317/5000 [00:04<00:12, 291.24it/s]
 27%|██▋       | 1347/5000 [00:04<00:12, 291.46it/s]
 28%|██▊       | 1377/5000 [00:04<00:12, 291.63it/s]
 28%|██▊       | 1407/5000 [00:04<00:12, 291.34it/s]
 29%|██▊       | 1437/5000 [00:04<00:12, 291.43it/s]
 29%|██▉       | 1467/5000 [00:04<00:12, 291.65it/s]
 30%|██▉       | 1497/5000 [00:04<00:12, 291.84it/s]
 31%|███       | 1527/5000 [00:05<00:11, 291.79it/s]
 31%|███       | 1557/5000 [00:05<00:11, 292.03it/s]
 32%|███▏      | 1587/5000 [00:05<00:11, 291.51it/s]
 32%|███▏      | 1617/5000 [00:05<00:11, 291.48it/s]
 33%|███▎      | 1647/5000 [00:05<00:11, 291.76it/s]
 34%|███▎      | 1677/5000 [00:05<00:11, 291.92it/s]
 34%|███▍      | 1707/5000 [00:05<00:11, 292.14it/s]
 35%|███▍      | 1737/5000 [00:05<00:11, 291.95it/s]
 35%|███▌      | 1767/5000 [00:05<00:11, 287.84it/s]
 36%|███▌      | 1796/5000 [00:06<00:11, 283.68it/s]
 36%|███▋      | 1825/5000 [00:06<00:11, 280.77it/s]
 37%|███▋      | 1854/5000 [00:06<00:11, 278.75it/s]
 38%|███▊      | 1882/5000 [00:06<00:11, 277.19it/s]
 38%|███▊      | 1910/5000 [00:06<00:11, 275.89it/s]
 39%|███▉      | 1938/5000 [00:06<00:11, 276.44it/s]
 39%|███▉      | 1968/5000 [00:06<00:10, 280.76it/s]
 40%|███▉      | 1998/5000 [00:06<00:10, 284.19it/s]
 41%|████      | 2028/5000 [00:06<00:10, 286.33it/s]
 41%|████      | 2058/5000 [00:06<00:10, 287.99it/s]
 42%|████▏     | 2088/5000 [00:07<00:10, 288.73it/s]
 42%|████▏     | 2118/5000 [00:07<00:09, 289.67it/s]
 43%|████▎     | 2148/5000 [00:07<00:09, 290.44it/s]
 44%|████▎     | 2178/5000 [00:07<00:09, 290.92it/s]
 44%|████▍     | 2208/5000 [00:07<00:09, 291.00it/s]
 45%|████▍     | 2238/5000 [00:07<00:09, 291.16it/s]
 45%|████▌     | 2268/5000 [00:07<00:09, 291.33it/s]
 46%|████▌     | 2298/5000 [00:07<00:09, 291.40it/s]
 47%|████▋     | 2328/5000 [00:07<00:09, 291.46it/s]
 47%|████▋     | 2358/5000 [00:07<00:09, 291.37it/s]
 48%|████▊     | 2388/5000 [00:08<00:08, 291.15it/s]
 48%|████▊     | 2418/5000 [00:08<00:08, 291.32it/s]
 49%|████▉     | 2448/5000 [00:08<00:08, 291.45it/s]
 50%|████▉     | 2478/5000 [00:08<00:08, 291.52it/s]
 50%|█████     | 2508/5000 [00:08<00:08, 291.80it/s]
 51%|█████     | 2538/5000 [00:08<00:08, 291.92it/s]
 51%|█████▏    | 2568/5000 [00:08<00:08, 292.18it/s]
 52%|█████▏    | 2598/5000 [00:08<00:08, 291.98it/s]
 53%|█████▎    | 2628/5000 [00:08<00:08, 291.74it/s]
 53%|█████▎    | 2658/5000 [00:09<00:08, 291.52it/s]
 54%|█████▍    | 2688/5000 [00:09<00:07, 291.40it/s]
 54%|█████▍    | 2718/5000 [00:09<00:07, 291.56it/s]
 55%|█████▍    | 2748/5000 [00:09<00:07, 291.15it/s]
 56%|█████▌    | 2778/5000 [00:09<00:07, 290.45it/s]
 56%|█████▌    | 2808/5000 [00:09<00:07, 290.02it/s]
 57%|█████▋    | 2838/5000 [00:09<00:07, 290.41it/s]
 57%|█████▋    | 2868/5000 [00:09<00:07, 290.80it/s]
 58%|█████▊    | 2898/5000 [00:09<00:07, 290.96it/s]
 59%|█████▊    | 2928/5000 [00:09<00:07, 291.02it/s]
 59%|█████▉    | 2958/5000 [00:10<00:07, 290.75it/s]
 60%|█████▉    | 2988/5000 [00:10<00:06, 291.05it/s]
 60%|██████    | 3018/5000 [00:10<00:06, 291.17it/s]
 61%|██████    | 3048/5000 [00:10<00:06, 290.93it/s]
 62%|██████▏   | 3078/5000 [00:10<00:06, 291.13it/s]
 62%|██████▏   | 3108/5000 [00:10<00:06, 291.11it/s]
 63%|██████▎   | 3138/5000 [00:10<00:06, 291.28it/s]
 63%|██████▎   | 3168/5000 [00:10<00:06, 291.42it/s]
 64%|██████▍   | 3198/5000 [00:10<00:06, 291.69it/s]
 65%|██████▍   | 3228/5000 [00:10<00:06, 291.57it/s]
 65%|██████▌   | 3258/5000 [00:11<00:05, 291.67it/s]
 66%|██████▌   | 3288/5000 [00:11<00:05, 291.87it/s]
 66%|██████▋   | 3318/5000 [00:11<00:05, 291.72it/s]
 67%|██████▋   | 3348/5000 [00:11<00:05, 291.70it/s]
 68%|██████▊   | 3378/5000 [00:11<00:05, 291.10it/s]
 68%|██████▊   | 3408/5000 [00:11<00:05, 290.93it/s]
 69%|██████▉   | 3438/5000 [00:11<00:05, 291.11it/s]
 69%|██████▉   | 3468/5000 [00:11<00:05, 291.44it/s]
 70%|██████▉   | 3498/5000 [00:11<00:05, 291.61it/s]
 71%|███████   | 3528/5000 [00:12<00:05, 291.56it/s]
 71%|███████   | 3558/5000 [00:12<00:04, 291.61it/s]
 72%|███████▏  | 3588/5000 [00:12<00:04, 291.84it/s]
 72%|███████▏  | 3618/5000 [00:12<00:04, 291.79it/s]
 73%|███████▎  | 3648/5000 [00:12<00:04, 291.88it/s]
 74%|███████▎  | 3678/5000 [00:12<00:04, 291.87it/s]
 74%|███████▍  | 3708/5000 [00:12<00:04, 292.02it/s]
 75%|███████▍  | 3738/5000 [00:12<00:04, 292.12it/s]
 75%|███████▌  | 3768/5000 [00:12<00:04, 292.01it/s]
 76%|███████▌  | 3798/5000 [00:12<00:04, 291.82it/s]
 77%|███████▋  | 3828/5000 [00:13<00:04, 290.21it/s]
 77%|███████▋  | 3858/5000 [00:13<00:03, 290.85it/s]
 78%|███████▊  | 3888/5000 [00:13<00:03, 291.52it/s]
 78%|███████▊  | 3918/5000 [00:13<00:03, 291.26it/s]
 79%|███████▉  | 3948/5000 [00:13<00:03, 291.56it/s]
 80%|███████▉  | 3978/5000 [00:13<00:03, 287.54it/s]
 80%|████████  | 4008/5000 [00:13<00:03, 288.90it/s]
 81%|████████  | 4038/5000 [00:13<00:03, 289.72it/s]
 81%|████████▏ | 4068/5000 [00:13<00:03, 290.47it/s]
 82%|████████▏ | 4098/5000 [00:13<00:03, 290.91it/s]
 83%|████████▎ | 4128/5000 [00:14<00:02, 291.10it/s]
 83%|████████▎ | 4158/5000 [00:14<00:02, 291.07it/s]
 84%|████████▍ | 4188/5000 [00:14<00:02, 290.98it/s]
 84%|████████▍ | 4218/5000 [00:14<00:02, 291.19it/s]
 85%|████████▍ | 4248/5000 [00:14<00:02, 291.17it/s]
 86%|████████▌ | 4278/5000 [00:14<00:02, 291.17it/s]
 86%|████████▌ | 4308/5000 [00:14<00:02, 291.33it/s]
 87%|████████▋ | 4338/5000 [00:14<00:02, 291.64it/s]
 87%|████████▋ | 4368/5000 [00:14<00:02, 291.89it/s]
 88%|████████▊ | 4398/5000 [00:14<00:02, 292.16it/s]
 89%|████████▊ | 4428/5000 [00:15<00:01, 292.23it/s]
 89%|████████▉ | 4458/5000 [00:15<00:01, 292.41it/s]
 90%|████████▉ | 4488/5000 [00:15<00:01, 292.12it/s]
 90%|█████████ | 4518/5000 [00:15<00:01, 292.01it/s]
 91%|█████████ | 4548/5000 [00:15<00:01, 291.94it/s]
 92%|█████████▏| 4578/5000 [00:15<00:01, 292.02it/s]
 92%|█████████▏| 4608/5000 [00:15<00:01, 292.32it/s]
 93%|█████████▎| 4638/5000 [00:15<00:01, 292.30it/s]
 93%|█████████▎| 4668/5000 [00:15<00:01, 292.49it/s]
 94%|█████████▍| 4698/5000 [00:16<00:01, 292.50it/s]
 95%|█████████▍| 4728/5000 [00:16<00:00, 292.46it/s]
 95%|█████████▌| 4758/5000 [00:16<00:00, 292.36it/s]
 96%|█████████▌| 4788/5000 [00:16<00:00, 292.07it/s]
 96%|█████████▋| 4818/5000 [00:16<00:00, 292.27it/s]
 97%|█████████▋| 4848/5000 [00:16<00:00, 292.33it/s]
 98%|█████████▊| 4878/5000 [00:16<00:00, 292.23it/s]
 98%|█████████▊| 4908/5000 [00:16<00:00, 292.14it/s]
 99%|█████████▉| 4938/5000 [00:16<00:00, 292.39it/s]
 99%|█████████▉| 4968/5000 [00:16<00:00, 292.46it/s]
100%|█████████▉| 4998/5000 [00:17<00:00, 292.35it/s]
100%|██████████| 5000/5000 [00:17<00:00, 293.23it/s]
Iteration 4999, current converge crit. = 1.43E-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.