Composing Normalizing Flows for Inverse Problems
Authors: Jay Whang, Erik Lindgren, Alex Dimakis
ICML 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our method is evaluated on a variety of inverse problems and is shown to produce highquality samples with uncertainty quantification. We further demonstrate that our approach can be amortized for zero-shot inference. We validate the efficacy of our proposed method in terms of both sample and reconstruction quality against three baselines: Langevin Monte Carlo (LMC), Ambient VI, and PLMCMC (Cannella et al., 2020). Both LMC and Pl-MCMC are MCMC techniques that can (asymptotically) sample from the true conditional distribution our method tries to approximate. For the comparisons to be fair, we implemented both methods to run MCMC chains in the latent space of the base model, analogous what our method does for VI. Ambient VI is identical to our method, except it performs VI in the image space and is included for completeness. In addition, we also conduct our experiments on three different datasets (MNIST, CIFAR-10, and Celeb A-HQ) to ensure that our method works across a range of settings. We report four different sample quality metrics: Frechet Inception Distance (FID), Learned Perceptual Image Patch Similarity (LPIPS), and Inception Score (IS) for CIFAR-10 (Heusel et al., 2017; Zhang et al., 2018; Salimans et al., 2016). While not strictly a measure of perceptual similarity, the average mean squared error (MSE) is reported for completeness. Additionally, we also report pairwise LPIPS metric used by Zheng et al. (2019) to measure the diversity of generated samples. |
| Researcher Affiliation | Collaboration | Jay Whang 1 Erik M. Lindgren 2 Alexandros G. Dimakis 3 1Dept. of Computer Science, UT Austin, TX, USA 2Google Research, NY, USA 3Dept. of Electrical and Computer Engineering, UT Austin, TX, USA. |
| Pseudocode | No | The paper does not contain any structured pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not provide a statement or a link indicating the availability of its source code. |
| Open Datasets | Yes | In addition, we also conduct our experiments on three different datasets (MNIST, CIFAR-10, and Celeb A-HQ) to ensure that our method works across a range of settings. |
| Dataset Splits | No | The paper specifies the use of a "test set" and mentions training base models, but it does not explicitly detail the training/validation/test splits (e.g., percentages or sample counts) or how data was partitioned for validation. |
| Hardware Specification | Yes | With our flow-based approximate posterior, this only takes 55 seconds 2. Measured on a single NVidia GTX 2080 GPU. |
| Software Dependencies | No | The paper mentions using "Adam optimizer (Kingma & Ba, 2014)" and "multiscale Real NVP architecture (Dinh et al., 2016)" but does not provide specific version numbers for these or other key software components (e.g., Python, PyTorch, TensorFlow). |
| Experiment Setup | Yes | We refer the reader to Appendix C for model hyperparameters and other details of our experiment setup. |