Intermediate Layer Optimization for Inverse Problems using Deep Generative Models

Authors: Giannis Daras, Joseph Dean, Ajil Jalal, Alex Dimakis

ICML 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We empirically show that our approach outperforms state-of-the-art methods introduced in Style GAN-2 and PULSE for a wide range of inverse problems including inpainting, denoising, super-resolution and compressed sensing.
Researcher Affiliation Academia 1The University of Texas at Austin.
Pseudocode Yes Algorithm 1 ILO for one layer of the generator
Open Source Code No The paper does not provide an explicit statement about releasing their code or a direct link to a code repository for the ILO method.
Open Datasets Yes To quantify the performance of the different methods we randomly select a few images from Celeba HQ (Liu et al., 2018; Lee et al., 2020) and reconstruct at different levels of sparsity. To demonstrate this, we run the following experiment; we remove entirely the loss functions that relate the generated images with a reference image (i.e. MSE and LPIPS) and we add a new classification loss term using an external classifier trained on a different domain. Essentially, we search for latent codes that lie in an l1 ball around the range of intermediate layers and maximize the probability that the generated image belongs to a certain category. We consider a classifier trained on Image Net (Deng et al., 2009).
Dataset Splits No The paper describes using Celeba-HQ and ImageNet datasets and discusses reconstruction errors, but it does not specify explicit training, validation, or test splits (e.g., percentages or exact counts) for these datasets.
Hardware Specification Yes To obtain the plots, we use a single V100 GPU.
Software Dependencies No The paper mentions software like Style GAN-2, PULSE, DCGAN, Big GAN, BM3D, and various loss functions (LPIPS, MSE), but it does not provide specific version numbers for any software dependencies.
Experiment Setup Yes To ensure that we stay in an l1 ball around the manifold at each layer, we use Projected Gradient Descent (PGD) (Nesterov, 2003). To implement the projection to an l1 ball around the current best solution (see line 4 of Algorithm (1)), we use the method of Duchi et al. (2008). Guided by our theory, we increase the maximum allowed deviation as we move to higher dimensional latent spaces. The radii of the balls are tuned separately as hyperparameters, for a full description see the Appendix. Style GAN-2 typically requires 300 1000 optimization steps (on the first layer) for a good reconstruction (Karras et al., 2019; 2020). However, we observe that running 50 steps in each one of the first four layers outperforms CSGM. We run 300 optimization steps per layer. The leftmost point corresponds to CSGM, i.e. we optimize over only the first layer.