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 4605.25 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 0x7fc5b53fca30>

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%|          | 30/5000 [00:00<00:16, 295.80it/s]
  1%|          | 61/5000 [00:00<00:16, 301.87it/s]
  2%|▏         | 92/5000 [00:00<00:16, 303.75it/s]
  2%|▏         | 123/5000 [00:00<00:15, 304.83it/s]
  3%|▎         | 154/5000 [00:00<00:15, 305.29it/s]
  4%|▎         | 185/5000 [00:00<00:15, 305.39it/s]
  4%|▍         | 216/5000 [00:00<00:15, 305.44it/s]
  5%|▍         | 247/5000 [00:00<00:15, 305.17it/s]
  6%|▌         | 278/5000 [00:00<00:15, 305.35it/s]
  6%|▌         | 309/5000 [00:01<00:15, 305.42it/s]
  7%|▋         | 340/5000 [00:01<00:15, 305.29it/s]
  7%|▋         | 371/5000 [00:01<00:15, 305.64it/s]
  8%|▊         | 402/5000 [00:01<00:15, 305.65it/s]
  9%|▊         | 433/5000 [00:01<00:14, 305.79it/s]
  9%|▉         | 464/5000 [00:01<00:14, 306.09it/s]
 10%|▉         | 495/5000 [00:01<00:14, 305.76it/s]
 11%|█         | 526/5000 [00:01<00:14, 305.88it/s]
 11%|█         | 557/5000 [00:01<00:14, 306.04it/s]
 12%|█▏        | 588/5000 [00:01<00:14, 306.20it/s]
 12%|█▏        | 619/5000 [00:02<00:14, 306.06it/s]
 13%|█▎        | 650/5000 [00:02<00:14, 306.18it/s]
 14%|█▎        | 681/5000 [00:02<00:14, 305.85it/s]
 14%|█▍        | 712/5000 [00:02<00:14, 305.80it/s]
 15%|█▍        | 743/5000 [00:02<00:13, 306.05it/s]
 15%|█▌        | 774/5000 [00:02<00:13, 306.01it/s]
 16%|█▌        | 805/5000 [00:02<00:13, 306.15it/s]
 17%|█▋        | 836/5000 [00:02<00:13, 306.39it/s]
 17%|█▋        | 867/5000 [00:02<00:13, 306.60it/s]
 18%|█▊        | 898/5000 [00:02<00:13, 306.69it/s]
 19%|█▊        | 929/5000 [00:03<00:13, 306.31it/s]
 19%|█▉        | 960/5000 [00:03<00:13, 306.06it/s]
 20%|█▉        | 991/5000 [00:03<00:13, 305.99it/s]
 20%|██        | 1022/5000 [00:03<00:13, 302.00it/s]
 21%|██        | 1053/5000 [00:03<00:13, 298.23it/s]
 22%|██▏       | 1083/5000 [00:03<00:13, 295.53it/s]
 22%|██▏       | 1113/5000 [00:03<00:13, 293.23it/s]
 23%|██▎       | 1143/5000 [00:03<00:13, 291.83it/s]
 23%|██▎       | 1173/5000 [00:03<00:13, 290.87it/s]
 24%|██▍       | 1203/5000 [00:03<00:13, 290.15it/s]
 25%|██▍       | 1233/5000 [00:04<00:12, 289.80it/s]
 25%|██▌       | 1262/5000 [00:04<00:12, 289.55it/s]
 26%|██▌       | 1291/5000 [00:04<00:12, 288.94it/s]
 26%|██▋       | 1320/5000 [00:04<00:12, 288.82it/s]
 27%|██▋       | 1349/5000 [00:04<00:12, 288.88it/s]
 28%|██▊       | 1378/5000 [00:04<00:12, 289.06it/s]
 28%|██▊       | 1407/5000 [00:04<00:12, 288.51it/s]
 29%|██▊       | 1436/5000 [00:04<00:12, 288.62it/s]
 29%|██▉       | 1465/5000 [00:04<00:12, 288.94it/s]
 30%|██▉       | 1494/5000 [00:04<00:12, 288.99it/s]
 30%|███       | 1523/5000 [00:05<00:12, 288.52it/s]
 31%|███       | 1552/5000 [00:05<00:11, 288.37it/s]
 32%|███▏      | 1581/5000 [00:05<00:11, 288.05it/s]
 32%|███▏      | 1610/5000 [00:05<00:11, 288.19it/s]
 33%|███▎      | 1639/5000 [00:05<00:11, 288.28it/s]
 33%|███▎      | 1668/5000 [00:05<00:11, 288.22it/s]
 34%|███▍      | 1697/5000 [00:05<00:11, 287.83it/s]
 35%|███▍      | 1726/5000 [00:05<00:11, 288.19it/s]
 35%|███▌      | 1755/5000 [00:05<00:11, 288.47it/s]
 36%|███▌      | 1784/5000 [00:05<00:11, 288.70it/s]
 36%|███▋      | 1813/5000 [00:06<00:11, 286.45it/s]
 37%|███▋      | 1842/5000 [00:06<00:11, 287.03it/s]
 37%|███▋      | 1871/5000 [00:06<00:10, 287.31it/s]
 38%|███▊      | 1900/5000 [00:06<00:10, 287.24it/s]
 39%|███▊      | 1929/5000 [00:06<00:10, 287.32it/s]
 39%|███▉      | 1958/5000 [00:06<00:10, 287.15it/s]
 40%|███▉      | 1987/5000 [00:06<00:10, 287.44it/s]
 40%|████      | 2016/5000 [00:06<00:10, 287.72it/s]
 41%|████      | 2045/5000 [00:06<00:10, 287.85it/s]
 41%|████▏     | 2074/5000 [00:06<00:10, 287.77it/s]
 42%|████▏     | 2103/5000 [00:07<00:10, 288.24it/s]
 43%|████▎     | 2132/5000 [00:07<00:09, 288.22it/s]
 43%|████▎     | 2161/5000 [00:07<00:09, 287.91it/s]
 44%|████▍     | 2190/5000 [00:07<00:09, 287.91it/s]
 44%|████▍     | 2219/5000 [00:07<00:09, 287.86it/s]
 45%|████▍     | 2248/5000 [00:07<00:09, 287.54it/s]
 46%|████▌     | 2277/5000 [00:07<00:09, 287.74it/s]
 46%|████▌     | 2306/5000 [00:07<00:09, 287.99it/s]
 47%|████▋     | 2335/5000 [00:07<00:09, 288.15it/s]
 47%|████▋     | 2364/5000 [00:08<00:09, 288.05it/s]
 48%|████▊     | 2393/5000 [00:08<00:09, 288.23it/s]
 48%|████▊     | 2422/5000 [00:08<00:08, 288.08it/s]
 49%|████▉     | 2451/5000 [00:08<00:08, 287.81it/s]
 50%|████▉     | 2480/5000 [00:08<00:08, 288.14it/s]
 50%|█████     | 2509/5000 [00:08<00:08, 288.23it/s]
 51%|█████     | 2538/5000 [00:08<00:08, 287.90it/s]
 51%|█████▏    | 2567/5000 [00:08<00:08, 287.97it/s]
 52%|█████▏    | 2596/5000 [00:08<00:08, 288.17it/s]
 52%|█████▎    | 2625/5000 [00:08<00:08, 288.30it/s]
 53%|█████▎    | 2654/5000 [00:09<00:08, 288.23it/s]
 54%|█████▎    | 2683/5000 [00:09<00:08, 288.08it/s]
 54%|█████▍    | 2712/5000 [00:09<00:07, 287.89it/s]
 55%|█████▍    | 2741/5000 [00:09<00:07, 288.16it/s]
 55%|█████▌    | 2770/5000 [00:09<00:07, 288.25it/s]
 56%|█████▌    | 2799/5000 [00:09<00:07, 288.14it/s]
 57%|█████▋    | 2828/5000 [00:09<00:07, 287.86it/s]
 57%|█████▋    | 2857/5000 [00:09<00:07, 287.84it/s]
 58%|█████▊    | 2886/5000 [00:09<00:07, 288.09it/s]
 58%|█████▊    | 2915/5000 [00:09<00:07, 288.26it/s]
 59%|█████▉    | 2944/5000 [00:10<00:07, 288.04it/s]
 59%|█████▉    | 2973/5000 [00:10<00:07, 287.97it/s]
 60%|██████    | 3002/5000 [00:10<00:06, 287.99it/s]
 61%|██████    | 3031/5000 [00:10<00:06, 288.26it/s]
 61%|██████    | 3060/5000 [00:10<00:06, 288.38it/s]
 62%|██████▏   | 3089/5000 [00:10<00:06, 288.34it/s]
 62%|██████▏   | 3118/5000 [00:10<00:06, 287.87it/s]
 63%|██████▎   | 3147/5000 [00:10<00:06, 288.21it/s]
 64%|██████▎   | 3176/5000 [00:10<00:06, 288.32it/s]
 64%|██████▍   | 3205/5000 [00:10<00:06, 288.56it/s]
 65%|██████▍   | 3234/5000 [00:11<00:06, 288.06it/s]
 65%|██████▌   | 3263/5000 [00:11<00:06, 288.16it/s]
 66%|██████▌   | 3292/5000 [00:11<00:05, 288.29it/s]
 66%|██████▋   | 3321/5000 [00:11<00:05, 288.58it/s]
 67%|██████▋   | 3350/5000 [00:11<00:05, 288.49it/s]
 68%|██████▊   | 3379/5000 [00:11<00:05, 288.47it/s]
 68%|██████▊   | 3408/5000 [00:11<00:05, 288.22it/s]
 69%|██████▊   | 3437/5000 [00:11<00:05, 288.34it/s]
 69%|██████▉   | 3466/5000 [00:11<00:05, 288.57it/s]
 70%|██████▉   | 3495/5000 [00:11<00:05, 288.55it/s]
 70%|███████   | 3524/5000 [00:12<00:05, 288.32it/s]
 71%|███████   | 3553/5000 [00:12<00:05, 288.45it/s]
 72%|███████▏  | 3582/5000 [00:12<00:04, 287.16it/s]
 72%|███████▏  | 3611/5000 [00:12<00:04, 287.07it/s]
 73%|███████▎  | 3640/5000 [00:12<00:04, 287.51it/s]
 73%|███████▎  | 3669/5000 [00:12<00:04, 287.96it/s]
 74%|███████▍  | 3698/5000 [00:12<00:04, 287.96it/s]
 75%|███████▍  | 3727/5000 [00:12<00:04, 288.21it/s]
 75%|███████▌  | 3756/5000 [00:12<00:04, 288.39it/s]
 76%|███████▌  | 3785/5000 [00:12<00:04, 288.47it/s]
 76%|███████▋  | 3814/5000 [00:13<00:04, 288.03it/s]
 77%|███████▋  | 3843/5000 [00:13<00:04, 288.05it/s]
 77%|███████▋  | 3872/5000 [00:13<00:03, 287.89it/s]
 78%|███████▊  | 3901/5000 [00:13<00:03, 288.06it/s]
 79%|███████▊  | 3930/5000 [00:13<00:03, 288.16it/s]
 79%|███████▉  | 3959/5000 [00:13<00:03, 288.19it/s]
 80%|███████▉  | 3988/5000 [00:13<00:03, 287.51it/s]
 80%|████████  | 4017/5000 [00:13<00:03, 287.69it/s]
 81%|████████  | 4046/5000 [00:13<00:03, 287.91it/s]
 82%|████████▏ | 4075/5000 [00:13<00:03, 288.06it/s]
 82%|████████▏ | 4104/5000 [00:14<00:03, 288.17it/s]
 83%|████████▎ | 4133/5000 [00:14<00:03, 288.18it/s]
 83%|████████▎ | 4162/5000 [00:14<00:02, 288.04it/s]
 84%|████████▍ | 4191/5000 [00:14<00:02, 287.83it/s]
 84%|████████▍ | 4220/5000 [00:14<00:02, 287.74it/s]
 85%|████████▍ | 4249/5000 [00:14<00:02, 287.73it/s]
 86%|████████▌ | 4278/5000 [00:14<00:02, 287.51it/s]
 86%|████████▌ | 4307/5000 [00:14<00:02, 287.64it/s]
 87%|████████▋ | 4336/5000 [00:14<00:02, 287.88it/s]
 87%|████████▋ | 4365/5000 [00:14<00:02, 288.02it/s]
 88%|████████▊ | 4394/5000 [00:15<00:02, 287.68it/s]
 88%|████████▊ | 4423/5000 [00:15<00:02, 287.63it/s]
 89%|████████▉ | 4452/5000 [00:15<00:01, 287.68it/s]
 90%|████████▉ | 4481/5000 [00:15<00:01, 287.60it/s]
 90%|█████████ | 4510/5000 [00:15<00:01, 287.98it/s]
 91%|█████████ | 4539/5000 [00:15<00:01, 287.91it/s]
 91%|█████████▏| 4568/5000 [00:15<00:01, 287.57it/s]
 92%|█████████▏| 4597/5000 [00:15<00:01, 287.84it/s]
 93%|█████████▎| 4626/5000 [00:15<00:01, 288.16it/s]
 93%|█████████▎| 4655/5000 [00:15<00:01, 288.08it/s]
 94%|█████████▎| 4684/5000 [00:16<00:01, 288.14it/s]
 94%|█████████▍| 4713/5000 [00:16<00:00, 288.30it/s]
 95%|█████████▍| 4742/5000 [00:16<00:00, 287.41it/s]
 95%|█████████▌| 4771/5000 [00:16<00:00, 287.32it/s]
 96%|█████████▌| 4800/5000 [00:16<00:00, 287.61it/s]
 97%|█████████▋| 4829/5000 [00:16<00:00, 287.65it/s]
 97%|█████████▋| 4858/5000 [00:16<00:00, 287.36it/s]
 98%|█████████▊| 4887/5000 [00:16<00:00, 287.57it/s]
 98%|█████████▊| 4916/5000 [00:16<00:00, 287.67it/s]
 99%|█████████▉| 4945/5000 [00:16<00:00, 287.87it/s]
 99%|█████████▉| 4974/5000 [00:17<00:00, 288.14it/s]
100%|██████████| 5000/5000 [00:17<00:00, 291.40it/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.