deepinv - Release v0.3.3

New Features

• Automatic Aadjoint, Uadjoint and V computation for user-defined physics
• Add RAM model
• FastMRI better raw data loading: load targets from different folder for test sets, load mask from test set, prewhitening, normalisation
• SKM-TEA raw MRI dataset
• New downsampling physics that matches MATLAB bicubic imresize

Changed

• Rename the normalizing function deepinv.utils.rescaleimg to normalizesignal
• Changed default linear solver from CG to lsqr
• Added positive clipping by default and gain minimum in PoissonGaussianNoise

Fixed

• Fix downsampling generator batching
• Fix memory leak in deepinv.physics.tomography when using autograd
• Fix the circular padded UNet
• Clamp constant signals in deepinv.utils.rescale_img to ensure they are normalized
• Fix ZeroNoise default missing in Physics

- Python
Published by tachella 10 months ago

deepinv - Release v0.3.2

New Features

• Add support for astra-toolbox CT operators (parallel, fan, cone) with deepinv.physics.TomographyWithAstra (#474 by Romain Vo)
• Add Physics.clone (#534 by Jérémy Scanvic)

Changed

• Make autograd use the base linear operator for deepinv.physics.adjoint_function (#519 by Jérémy Scanvic)
• Parallelize the test suite making it 15% faster (#522 by Jérémy Scanvic)
• Adjust backward paths for tomography (#535 by Johannes Hertrich)

Fixed

• Fix the total loss reported by the trainer (#515 by Jérémy Scanvic)
• Fix the gradient norm reported by the trainer (#520 by Jérémy Scanvic)
• Fix that the maxpixel option in PSNR and SSIM and add analgous minpixel option (#535 by Johannes Hertrich)
• Fix some issues related to denoisers: ICNN grad not working inside torch.no_grad(), batch of image and batch of sigma for some denoisers (DiffUNet, BM3D, TV, Wavemet), EPLL error when batch size > 1 (#530 by Minh Hai Nguyen)
  
• Batching WaveletPrior and fix iwt(#530 by Minh Hai Nguyen)

- Python
Published by github-actions[bot] 11 months ago

deepinv - Release v0.3.1

• Added deepinv.physics.SaltPepperNoise for impulse noise
• Add measurement augmentation VORTEX loss
• Add non-geometric data augmentations (noise, phase errors)
• Make PhysicsGenerator.average use batches
• MRI losses subclass, weighted-SSDU, Robust-SSDU loss functions + more mask generators
• Multi-coil MRI estimates sens maps with sigpy ESPIRiT, MRISliceTransform better loads raw data by estimating coil maps and generating masks
• Add HaarPSI metric + metric standardization

- Python
Published by tachella 12 months ago

deepinv - Release v0.3.0

• Diffusion models: SDE class and solvers, noisy data fidelity terms, added EDM models
• Physics: Added CASSI, HyperSpectral Unmixing, Ptychography and StackedPhysics (for sensor fusion)
• Validation: added validation dataset to Trainer, with early_stopping and best model saving options.
• Datasets: SimpleFastMRISliceDataset, CMRxRecon, NBU satellite image dataset
• Linear Solvers: added BiCGStab, LSQR and MINRES
• Custom models: MoDL, VarNet/E2E-VarNet, PanNet

- Python
Published by tachella about 1 year ago

deepinv - Release v0.2.2

• new documentation
• Metrics classes
• Advanced MRI: 3D and multicoil support
• Bregman potentials
• Advanced transforms: diffeomorphisms, time transforms
• new datasets: Kohler image deblurring
• L12 prior
• support temporal/video physics and models

- Python
Published by tachella over 1 year ago

deepinv - Release v0.2.1

New features:

• New self-supervised learning losses: advanced Splitting Losses (Noise2Void, Noise2Self, Artifact2Artifact, etc.), Noise2Score, new EquivariantImaging transforms
• New operators: Dynamic MRI, RadioInterferometry, Gamma noise
• New PhysicsGenerators (advanced MRI masking)
• Easy-to-load standard datasets: Fluorescent Microscopy, FastMRI, LIDC, Flickr2k, LSDIR
• Improved Trainer
• Adversarial training losses and examples
• Move to python 3.9

- Python
Published by tachella over 1 year ago

deepinv - Release v0.2.0

New features!

• Physics parameterization for blind inverse + calibration problems
• Random physics generators
• New Trainer
• Metrics
• 3D denoisers
• Patch priors

- Python
Published by tachella about 2 years ago

deepinv - Release v0.1.1

New Features

• Added r2r loss (:gh:148 by Brayan Monroy_) - 30/01/2024
• Added scale transform (:gh:135 by Jérémy Scanvic_) - 19/12/2023
• Added priors for total variation and l12 mixed norm (:gh:156 by Nils Laurent_) - 09/02/2023

Fixed

• Fixed issue in noise forward of Decomposable class (:gh:154 by Matthieu Terris_) - 08/02/2024
• Fixed new black version 24.1.1 style changes (:gh:151 by Julian Tachella_) - 31/01/2024
• Fixed test for sigma as torch tensor with gpu enable (:gh:145 by Brayan Monroy_) - 23/12/2023
• Fixed :gh:139 BM3D tensor format grayscale (:gh:140 by Matthieu Terris_) - 23/12/2023
• Fixed :gh:136 noise additive model for DecomposablePhysics (:gh:138 by Matthieu Terris_) - 22/12/2023
• Importing deepinv does not modify matplotlib config anymore (:gh1501 by Thomas Moreau_) - 30/01/2024

Changed

• Rephrased the README (:gh:142 by Jérémy Scanvic_) - 09/01/2024

- Python
Published by tachella over 2 years ago

deepinv - Release v0.1.0

• better network training utils (on-the-fly data generation, etc)
• new SOTA denoisers (diffUNet, SwinIR, etc)
• more diffusion-based algorithms (DPS, DiffPIR)
• HuggingFace integration
• explicit priors (L1Prior, Tikhonov, etc)

- Python
Published by tachella over 2 years ago

deepinv - Release v0.0.1

First stable release.

- Python
Published by tachella almost 3 years ago