Samuel Hurault

DeepInverse has joined the PyTorch Ecosystem, making imaging with deep learning easier and more reproducible across research and industry. DeepInverse is an open source framework for solving imaging inverse problems in medical imaging, computational photography, remote sensing, astronomical imaging, microscopy and more. DeepInverse makes imaging with deep learning easy and is developed by a passionate community of researchers, practitioners and engineers. 🔗 Learn more & explore how to get involved: hubs.la/Q03RKyHJ0 #PyTorch #OpenSourceAI #DeepInverse #Imaging #DeepLearning

Summer School's Hands-On Tutorials 💻DeepInverse: A PyTorch Library for Imaging Inverse Problems 🎓By @TachellaJulian (ENS Lyon) @MatthieuTerris (Inria) @HuraultSamuel (CNRS) 📡Build Your Own Radar With MIT’s Next-Generation Kit 🎓By Brad Perry & Ken Kolodziej, MIT Lincoln Lab

Not coming back from Marseille empty handed #DeepInverse @TachellaJulian @MatthieuTerris @ddongchen @minhhai712

Merci pour l’opportunité d’avoir échangé sur mes recherches et mes expériences ! Merci à mes directeurs de thèse @gabrielpeyre et @RemiGribonval pour votre supervision 😊

🚀 deepinverse hackathon🚀 We just finished a couple of days working on deepinverse in the beautiful CIRM venue. deepinv.github.io/deepinv/ Immensely grateful to the dream team of contributors that participated in the hackathon! A thread 🧵

@MatthieuTerris, @HuraultSamuel and I will be giving a tutorial on deepinverse in Luxembourg next year! + the symposium has a great lineup of speakers!

@YTraonmilin @TachellaJulian @MatthieuTerris @ddongchen Working on it, thanks Yann!

@TachellaJulian @HuraultSamuel @MatthieuTerris @ddongchen Great tool ! Suggestion : add a bibtex ref on the site to properly cite 😉

📢📢 Release of DeepInverse library 📢📢 After months of intense work, we are releasing the first stable version of DeepInverse deepinv.github.io/deepinv/, a PyTorch library for solving inverse problems with deep learning. with @HuraultSamuel, @MatthieuTerris and @ddongchen A 🧵

@OwkinScience perfectly captured the atmosphere of our benchopt sprint ! tinyurl.com/bench-sprint-v… A lot of exchanges, coding, and fun! 😄 So proud of what we achieved during these three days! Here is a recap of this sprint's output: notes.inria.fr/s/XDBWzL8ID#

Newbies but goldies 😌: “Keep the Momentum: Conservation Laws beyond Euclidean Gradient Flows” with @gabrielpeyre and @RemiGribonval. We study conservation laws during the (euclidean or not) gradient or momentum flow of neural networks. arxiv.org/abs/2405.12888

📢 DeepInverse v0.2 is out! New features: 📏 Physics for blind inverse + calibration problems 🕸️ Advanced blurs 💣 Random physics generators 🤖New modular Trainer 🪂 Phase retrieval 🎉 Patch and 3D priors by new contributors! A🧵 deepinv.github.io

Very honored to have received the PhD prize « Signal, Image, Vision » from the French research organizations EEA, GdR IASIS and GRETSI. Thanks a lot for this distinction.

@ukmlv @iclr_conf @2ptmvd @docmilanfar @YuyangHu_666 Very interesting read and super clear derivation

@ukmlv @iclr_conf @2ptmvd @docmilanfar @YuyangHu_666 I think that algorithm 1 could be also seen as a special case of our Bregman RED algorithm (arxiv.org/pdf/2306.03466……) with h = ||.||_H^t H, isn't it ?

📢📢 DeepInverse v0.1.0 Release 📢📢 github.com/deepinv/deepinv 💣New SOTA denoisers 🕸️More diffusion-based algorithms (DPS, DiffPIR) 🤗HuggingFace integration 🎁More training utils (on-the-fly data generation, etc) 🤖Explicit priors (L1Prior, etc)

People DeepIng into DeepInv at the DIPopt workshop 🤗

Happy to share that I defended my PhD thesis yesterday. It was a wonderful moment that I got to share with friends and family. Thanks to everybody who listened to me in Bordeaux or from distance. If you are interested, you can find my PhD manuscript here : plmbox.math.cnrs.fr/f/64a93457647f…

🎁Diffusion Posterior Sampling (DPS) is now part of the deepinverse library. Big thanks to @hyungjin_chung for the contribution! The list of deepinv contributors keeps growing 🎉

@tonysilveti @gabrielpeyre Very cool! And backward Euler gives iterations of a Bregman proximity operator ?

@HuraultSamuel @gabrielpeyre Euler discretization gives the so-called natural gradient descent, taking a partial Euler discretization gives mirror descent (see here proceedings.mlr.press/v130/gunasekar…)

For small step sizes, mirror descent is equivalent to a gradient descent on a Hessian manifold. blogs.princeton.edu/imabandit/2013…

2 papers accepted at Neurips 2023 ! - Convergent Bregman Plug-and-Play Image Restoration for Poisson Inverse Probems. With @ukmlv, A. Leclaire and N. Papadakis. - Self-Consistent Velocity Matching of Probability Flows. With Lingxiao Li and @JustinMSolomon.