Code accompanying paper G. Jagatap and C. Hegde, "Algorithmic Guarantees for Inverse Imaging with Untrained Network Priors", Advances in Neural Information Processing Systems (NeurIPS), 2019.
For details please refer to the preprint available at: https://arxiv.org/abs/1906.08763
Run the python notebook compressive_imaging.ipynb. The code works in 3 modes:
- Mode 1: model fitting. Find the best network weights w that approximate image x as G(w,z), with fixed code z. Solves the obective function min_w || x - G(w,z) ||2^2. Can also be used for denoising type applications such as super-resolution, inpainting, denoising.
- Mode 2: inverting linear compressive measurements. Reconstruct an image x from linear compressive measurements y=Ax. Solves the objective function min_x ||y - Ax||2^2 such that x = G(w,z).
- Mode 3: inverting magnitude-only compressive measurements. Reconstruct an image x from magnitude-only compressive measurements y=|Ax|. Solves the objective function min_x ||y - |Ax|||2^2 such that x = G(w,z).
Datasets used: MNIST and CelebA
The untrained generative network G(w,z) can assume a decoder architecture (set decodetype='decoder')described in "Deep Decoder: Concise Image Representations from Untrained Non-convolutional Networks", ICLR 2019, https://arxiv.org/abs/1810.03982; or can also assume a DCGAN architecture (set decodetype='transposeconv').