AmbientFlow: Invertible generative models from incomplete, noisy measurements
Authors: Varun A. Kelkar, Rucha Deshpande, Arindam Banerjee, Mark Anastasio
TMLR 2024 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive numerical studies demonstrate the effectiveness of Ambient Flow in learning the object distribution. The utility of Ambient Flow in a downstream inference task of image reconstruction is demonstrated. Numerical Studies This section describes the numerical studies used to demonstrate the utility of Ambient Flow for learning object distributions from noisy and incomplete imaging measurements. The studies include toy problems in two dimensions, low-dimensional problems involving a distribution of handwritten digits from the MNIST dataset, problems involving face images from the Celeb A-HQ dataset as well as the problem of recovering the object distribution from stylized magnetic resonance imaging measurements. A case study that demonstrates the utility of Ambient Flow in the downstream tasks of image reconstruction and posterior sampling is also described. |
| Researcher Affiliation | Collaboration | Varun A. Kelkar EMAIL University of Illinois at Urbana-Champaign, Urbana, IL 61801 Currently at Analog Devices, Inc., Boston, MA 02110 Rucha M. Deshpande EMAIL Washington University in St. Louis, St. Louis, MO 63130 Arindam Banerjee EMAIL University of Illinois at Urbana-Champaign, Urbana, IL 61801 Mark A. Anastasio EMAIL University of Illinois at Urbana-Champaign, Urbana, IL 61801 |
| Pseudocode | No | The paper describes the approach using mathematical formulations and theorems (Theorem 3.1, Lemma 3.1, Theorem 3.2) and then proceeds to numerical studies. There are no explicit sections or figures labeled 'Pseudocode' or 'Algorithm'. |
| Open Source Code | Yes | 1Our Py Torch implementation of Ambient Flow can be found at https://github.com/comp-imaging-sci/ambientflow |
| Open Datasets | Yes | Next, a problem of recovering the distribution of MNIST digits from noisy and/or blurred images of MNIST digits was considered (Le Cun et al., 1998). For the face image study, images of size n = 64 64 3 from the Celeb A-HQ dataset were considered (Karras et al., 2017). T2-weighted brain images of size n = 128 128 from the Fast MRI initiative database were considered (Zbontar et al., 2018) as samples from the object distribution. |
| Dataset Splits | Yes | A training dataset of size D = 5 107 was used. For the first task, the compressed sensing using generative models (CSGM) formalism was used to obtained approximate MAP estimates from measurements, for a held-out for a test dataset of size 45 (Asim et al., 2020): For a held-out dataset of size 20, the first task was performed using the compressed sensing using generative models (CSGM) formalism (Asim et al., 2020). |
| Hardware Specification | Yes | Ambient Flow was trained using Py Torch using an NVIDIA Quadro RTX 8000 GPU. |
| Software Dependencies | No | Ambient Flow was trained using Py Torch using an NVIDIA Quadro RTX 8000 GPU. All hyperparameters for the main INN were fixed based on a Py Torch implementation of the Glow architecture (Seonghyeon), except the number of blocks, which was set to scale logarithmically by the image dimension. The paper mentions 'Py Torch' but does not specify a version number or other specific software dependencies with versions. |
| Experiment Setup | Yes | All hyperparameters for the main INN were fixed based on a Py Torch implementation of the Glow architecture (Seonghyeon), except the number of blocks, which was set to scale logarithmically by the image dimension. The regularization parameter for the image reconstruction method was tuned to give the lowest RMS error (RMSE) for every individual reconstructed image. The regularization parameters for each image reconstruction method were tuned to achieve the best RMSE on a single validation image, and then kept constant for the entire test dataset. For this problem, three different forward models were considered, namely the identity operator, and two Gaussian blurring operators Hblur1 and Hblur2 with root-mean-squared (RMS) width values σb = 1.5 and 3.0 pixels. The measurement noise was distributed as n N(0, σ2 n Im), with σn = 0.3. Stylized MRI measurements with undersampling ratio n/m = 1 (fully sampled) and n/m = 4 were simulated using the fast Fourier transform (FFT). |