Change Log

This change log is for the main branch. It contains changes for each release, with the date and author of each change.

Current

New Features

• Add dataset for patch sampling from (large) nD images without loading entire images into memory (#806 by Vicky De Ridder)

• Add support for 3D CNNs (#869 by Vicky De Ridder)

• Add support for complex dtypes in WaveletDenoiser, WaveletDictDenoiser and WaveletPrior (#738 by Chaithya G R)

• dinv.io functions for loading DICOM, NIFTI, COS, GEOTIFF etc. (#768 by Andrew Wang)

Changed

• load_np_url now returns tensors, load_url helper function moved to io (#768 by Andrew Wang)

• utils/signal.py renamed to signals.py to avoid stdlib conflict (#768 by Andrew Wang)

• utils.get_data_home now creates folder if not exist (#768 by Andrew Wang)

Fixed

• Blur physics objects now put new filters to physics device regardless of input filter device (#844 by Vicky De Ridder)

• Set14HR dataset now downloads from a different source (and has slightly different folderstructure), since old link broke. (#845 by Vicky De Ridder)

• LsdirHR dataset now downloads from a different source (since old link broke) and correctly contains the specific folder images, instead of everything in root (#866 by Vicky De Ridder)

• Fix typo in docstring of ItohFidelity and SpatialUnwrapping demo (#860 by Brayan Monroy)

• Fix unhandled import error in CBSD68 if datasets is not installed (#868 by Johannes Hertrich)

• Add support for complex signals in PSNR (#738 by Jérémy Scanvic)

v0.3.5

New Features

• Add statistics for SAR imaging + fix variance of deepinv.physics.GammaNoise in doc (#740 by Louise Friot Giroux)

• Add imshow kwargs to plot (#791 by Andrew Wang)

• Add deepinv.physics.RicianNoise model (#805 by Vicky De Ridder)

• Add manual physics to reduced resolution loss (#808 by Andrew Wang)

• Multi-coil MRI coil-map estimation acceleration via CuPy (#781 by Andrew Wang)

• Add SpatialUnwrapping forward model and ItohFidelity data fidelity (#723 by Brayan Monroy)

Changed

• (Breaking) Make deepinv.datasets.HDF5Dataset similar to deepinv.Trainer in the unsupervised setting by using NaNs for ground truths instead of a copy of the measurements (#761 by Jérémy Scanvic)

• Add squared parameter to LinearPhysics.compute_norm() and compute_sqnorm() method (#832 by Jérémy Scanvic)

• Allow self-supervised eval by removing the model.eval() from Trainer.train() (#777 by Julian Tachella)

• Make tqdm progress bar auto-resize (#835 by Andrew Wang)

• (Breaking) Normalize the Tomography operator with proper spectral norm computation. Set the default normalization behavior to True for both CT operators (#715 by Romain Vo)

Fixed

• Use the learning-free model for learning-free metrics in deepinv.Trainer (#788 by Jérémy Scanvic)

• Fix device deepinv.utils.dirac_like() and deepinv.physics.blur.bilinear_filter(), deepinv.physics.blur.bicubic_filter() and deepinv.physics.blur.gaussian_blur() filters (#785 by Julian Tachella)

• Fix positivity + batching gamma least squares solvers (#785 by Julian Tachella and Minh Hai Nguyen)

• Fix and test deepinv.models.RAM scaling issues (#785 by Julian Tachella)

• Reduced CI python version tests (#746 by Matthieu Terris)

• Fix scaling issue in deepinv.sampling.DiffusionSDE (#772 by Minh Hai Nguyen)

• All splitting losses fixed to work with changing image sizes and with multicoil MRI (#778 by Andrew Wang)

• deepinv.Trainer treats batch of nans as no ground truth (#793 by Andrew Wang)

• Save training loss history (#777 by Julian Tachella)

• Fix docstring formatting in BDSDS500 dataset (#816 by Brayan Monroy)

• Remove unnecessary tensor cloning from DDRM and DPS (#834 by Vicky De Ridder)

• Change deprecated torch.norm calls to torch.linalg.vector_norm (#840 by Minh Hai Nguyen)

v0.3.4

New Features

• Quickstart tutorials + clean examples (#622 by Andrew Wang)

• Dataset base class + deepinv.datasets.ImageFolder and deepinv.datasets.TensorDataset classes (#622 by Andrew Wang)

• Added GitHub action checking import time (#680 by Julian Tachella)

• Client model for server-side inference for using models in the cloud (#691 by Andrew Wang)

• Reduced resolution self-supervised loss (#735 by Andrew Wang)

• Add deepinv.utils.disable_tex() to disable LaTeX (#726 by Andrew Wang)

• Add BSDS500 dataset (#749 by Johannes Hertrich and Sebastian Neumayer)

• O(1) memory backprop for linear solver and example (#739 by Minh Hai Nguyen)

Changed

• Move mixins to utils and reduce number of cross-submodule top-level imports (#680 by Andrew Wang)

• deepinv.datasets.PatchDataset returns tensors and not tuples (#622 by Andrew Wang)

Fixed

• Fixed natsorted issue (#680 by Julian Tachella)

• Fix full-reference metrics used with measurement-only dataset (#622 by Andrew Wang)

• Batching deepinv.physics.generator.DownsamplingGenerator in the case of multiple filters (#690 by Matthieu Terris)

• NaN motion blur generator (#685 by Matthieu Terris)

• Fix the condition for break in compute_norm (#699 by Quentin Barthélemy)

• Python 3.9 backward compatibility and zip_strict (#707 by Andrew Wang)

• Fix numerical instability of deepinv.optim.utils.bicgstab() solver (#739 by Minh Hai Nguyen)

v0.3.3

New Features

• Automatic A_adjoint, U_adjoint and V computation for user-defined physics (#658 by Julian Tachella)

• Add deepinv.models.RAM model (#524 by Matthieu Terris)

• FastMRI better raw data loading: load targets from different folder for test sets, load mask from test set, prewhitening, normalisation (#608 by Andrew Wang)

• SKM-TEA raw MRI dataset (#608 by Andrew Wang)

• New downsampling physics that matches MATLAB bicubic imresize (#608 by Andrew Wang)

Changed

• Rename the normalizing function deepinv.utils.rescale_img to deepinv.utils.normalize_signal() (#641 by Jérémy Scanvic)

• Changed default linear solver from CG to deepinv.optim.utils.lsqr() (#658 by Julian Tachella)

• Added positive clipping by default and gain minimum in deepinv.physics.PoissonGaussianNoise (#658 by Julian Tachella).

Fixed

• Fix downsampling generator batching (#608 by Andrew Wang)

• Fix memory leak in deepinv.physics.Tomography when using autograd (#651 by Minh Hai Nguyen)

• Fix the circular padded deepinv.models.UNet (#653 by Victor Sechaud)

• Clamp constant signals in deepinv.utils.normalize_signal() to ensure they are normalized (#641 by Jérémy Scanvic)

• Fix ZeroNoise default missing in deepinv.physics.ZeroNoise (#658 by Julian Tachella)

v0.3.2

New Features

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

• Add deepinv.physics.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)

• Update python version to 3.10+ (#605 by Minh Hai Nguyen)

• Update the library dependencies, issue template, codecov report on linux only (#654 by Minh Hai Nguyen)

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 max_pixel option in PSNR and SSIM and add analgous min_pixel 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)

• Fix on unreliable/inconsistent automatic choosing GPU with most free VRAM (#570 by Fedor Goncharov)

v0.3.1

New Features

• Added deepinv.physics.SaltPepperNoise for impulse noise (#472 by Thomas Moreau).

• Add measurement augmentation VORTEX loss (#410 by Andrew Wang)

• Add non-geometric data augmentations (noise, phase errors) (#410 by Andrew Wang)

• Make deepinv.physics.generator.PhysicsGenerator.average use batches (#488 by Jérémy Scanvic)

• MRI losses subclass, weighted-SSDU, Robust-SSDU loss functions + more mask generators (#416 by Keying Guo and Andrew Wang)

• Multi-coil MRI estimates sens maps with sigpy ESPIRiT, MRISliceTransform better loads raw data by estimating coil maps and generating masks (#416 by Andrew Wang)

• Add HaarPSI metric + metric standardization (#416 by Andrew Wang)

• Add ENSURE loss (#454 by Andrew Wang)

Changed

• Added cake_cutting, zig_zag and xy orderings in deepinv.physics.SinglePixelCamera physics (#475 by Brayan Monroy).

Fixed

• Fix images not showing in sphinx examples (#478 by Matthieu Terris)

• Fix plot_inset not showing (#455 by Andrew Wang)

• Fix latex rendering in deepinv.utils.plotting.config_matplotlib() (#452 by Romain Vo)

• Get rid of unnecessary file system writes in deepinv.utils.get_freer_gpu() (#468 by Jérémy Scanvic)

• Fixed sequency ordering in deepinv.physics.SinglePixelCamera (#475 by Brayan Monroy)

• Change array operations from numpy to PyTorch in deepinv.physics.SinglePixelCamera (#483 by Jérémy Scanvic)

• Get rid of commented out code (#485 by Jérémy Scanvic)

• Changed deepinv.physics.SinglePixelCamera parameters in demos (#493 by Brayan Monroy)

• Improved code coverage by mocking datasets (#490 by Jérémy Scanvic)

• Fix MRI mask generator update img_size on-the-fly not updating n_lines (#416 by Andrew Wang)

• Upgrade deprecated typing.T types in the code base (#501 by Jérémy Scanvic)

v0.3

New Features

• Added early-stopping callback for deepinv.Trainer and best model saving (#437 by Julian Tachella and Andrew Wang)

• Add various generators for the physics module (downsampling, variable masks for inpainting, PoissonGaussian generators etc) (#384 by Matthieu Terris)

• Add minres least squared solver (#425 by Sebastian Neumayer and Johannes Hertrich)

• New least squared solvers (BiCGStab & LSQR) (#393 by Julian Tachella)

• Typehints are used automatically in the documentation (#379 by Julian Tachella)

• Add Ptychography operator in physics.phase_retrieval (#351 by Victor Sechaud)

• Multispectral: NBU satellite image dataset, ERGAS+SAM metrics, PanNet, generalised pansharpening and decolorize (#371 by Julian Tachella and Andrew Wang)

• StackedPhysics: class definition, loss and data-fidelity (#371 by Julian Tachella and Andrew Wang)

• Added HyperSpectral Unmixing operator (#353 by Dongdong Chen and Andrew Wang)

• Add CASSI operator (#377 by Andrew Wang)

• Add validation dataset to data generator (#363 by Andrew Wang)

• Add Rescale and ToComplex torchvision-style transforms (#363 by Andrew Wang)

• Add SimpleFastMRISliceDataset, simplify FastMRISliceDataset, add FastMRI tests (#363 by Andrew Wang)

• FastMRI now compatible with MRI and MultiCoilMRI physics (#363 by Andrew Wang)

• Add VarNet/E2E-VarNet model and generalise ArtifactRemoval (#363 by Andrew Wang)

• deepinv.Trainer now can log train progress per batch or per epoch (#388 by Andrew Wang)

• CMRxRecon dataset and generalised dataset metadata caching (#385 by Andrew Wang)

• Online training with noisy physics now can repeat the same noise each epoch (#414 by Andrew Wang)

• deepinv.Trainer test can return unaggregated metrics (#420 by Andrew Wang)

• MoDL model (#435 by Andrew Wang)

• Add conversion to Hounsfield Units (HUs) for LIDC IDRI (#459 by Jérémy Scanvic)

• Add ComposedLinearPhysics (via __mul__ method) (#462 by Minh Hai Nguyen and Julian Tachella )

• Register physics-dependent parameters to module buffers (#462 by Minh Hai Nguyen)

• Add example on optimizing physics parameters (#462 by Minh Hai Nguyen)

• Add device property to deepinv.utils.TensorList (#462 by Minh Hai Nguyen)

• Add test physics device transfer and differentiablity (#462 by Minh Hai Nguyen)

Fixed

• Fixed MRI noise bug in kernel of mask (#384 by Matthieu Terris)

• Support for multi-physics / multi-dataset during training fixed (#384 by Matthieu Terris)

• Fixed device bug (#415 by Dongdong Chen)

• Fixed hyperlinks throughout docs (#379 by Julian Tachella)

• Missing sigma normalization in L2Denoiser (#371 by Julian Tachella and Andrew Wang)

• deepinv.Trainer discards checkpoint after loading (#385 by Andrew Wang)

• Fix offline training with noise generator not updating noise params (#414 by Andrew Wang)

• Fix wrong reference link in auto examples (#432 by Minh Hai Nguyen)

• Fix paths in LidcIdriSliceDataset (#446 by Jérémy Scanvic)

• Fix device inconsistency in test_physics, physics classes and noise models (#462 by Minh Hai Nguyen)

• Fix Ptychography can not handle multi-channels input (#494 by Minh Hai Nguyen)

• Fix argument name (img_size, in_shape, …) inconsistency (#494 by Minh Hai Nguyen)

Changed

• Add bibtex references (#575 by Samuel Hurault)

• Set sphinx warnings as errors (#379 by Julian Tachella)

• Added single backquotes default to code mode in docs (#379 by Julian Tachella)

• Changed the __add__ method for stack method for stacking physics (#371 by Julian Tachella and Andrew Wang)

• Changed the R2R loss to handle multiple noise distributions (#380 by Brayan Monroy)

• deepinv.Trainer.get_samples_online() using physics generator now updates physics params via both deepinv.physics.Physics.update_parameters() and forward pass (#386 by Andrew Wang)

• Deprecate deepinv.Trainer freq_plot in favour of plot_interval (#388 by Andrew Wang)

v0.2.2

New Features

• Added NCNSpp, ADMUNet model and pretrained weights (by Minh Hai Nguyen)

• Added SDE class (DiffusionSDE (OU Process), VESDE) for image generation (by Minh Hai Nguyen and Samuel Hurault)

• Added SDE solvers (Euler, Heun) (by Minh Hai Nguyen and Samuel Hurault)

• Added example on image generation, working for deepinv.models.NCSNpp, deepinv.models.ADMUNet, deepinv.models.DRUNet and deepinv.models.DiffUNet (by Minh Hai Nguyen and Matthieu Terris)

• Added VP-SDE for image generation and posterior sampling (#434 by Minh Hai Nguyen)

• global path for datasets deepinv.utils.get_data_home() (#347 by Julian Tachella and Thomas Moreau)

• New docs user guide (#347 by Julian Tachella and Thomas Moreau)

• Added UNSURE loss (#313 by Julian Tachella)

• Add transform symmetrisation, further transform arithmetic, and new equivariant denoiser (#259 by Andrew Wang)

• New transforms: multi-axis reflect, time-shift and diffeomorphism (#259 by Andrew Wang)

• Add wrapper classes for adapting models to take time-sequence 2D+t input (#296 by Andrew Wang)

• Add sequential MRI operator (#296 by Andrew Wang)

• Add multi-operator equivariant imaging loss (#296 by Andrew Wang)

• Add loss schedulers (#296 by Andrew Wang)

• Add transform symmetrisation, further transform arithmetic, and new equivariant denoiser (#259 by Andrew Wang)

• New transforms: multi-axis reflect, time-shift and diffeomorphism (#259 by Andrew Wang)

• Multi-coil MRI, 3D MRI, MRI Mixin (#287 by Andrew Wang, Brett Levac)

• Add Metric baseclass, unified params (for complex, norm, reduce), typing, tests, L1L2 metric, QNR metric, metrics docs section, Metric functional wrapper (#309, #343 by Andrew Wang)

• generate_dataset features: complex numbers, save/load physics_generator params, overwrite bool (#324, #352 by Andrew Wang)

• Add the Köhler dataset (#271 by Jérémy Scanvic)

Fixed

• Fixed sphinx warnings (#347 by Julian Tachella and Thomas Moreau)

• Fix cache file initialization in FastMRI Dataloader (#300 by Pierre-Antoine Comby)

• Fixed prox_l2 no learning option in deepinv.Trainer (#304 by Julian Tachella)

• Fixed SSIM to use lightweight torchmetrics function + add MSE and NMSE as metrics + allow PSNR & SSIM to set max pixel on the fly (#296 by Andrew Wang)

• Fix generate_dataset error with physics_generator and batch_size != 1. (#315 by apolychronou)

• Fix generate_dataset error not using random physics generator (#324 by Andrew Wang)

• Fix Scale transform rng device error (#324 by Andrew Wang)

• Fix bug when using cuda device in dinv.datasets.generate_dataset (#334 by Tobias Liaudat)

• Update outdated links in the readme (#366 by Jérémy Scanvic)

Changed

• Added direct option to ArtifactRemoval (#347 by Julian Tachella and Thomas Moreau)

• Sphinx template to pydata (#347 by Julian Tachella and Thomas Moreau)

• Remove metrics from utils and consolidate complex and normalisation options (#309 by Andrew Wang)

• get_freer_gpu falls back to torch.cuda when nvidia-smi fails (#352 by Andrew Wang)

• libcpab now is a PyPi package for diffeomorphisms, add rngs and devices to transforms (#370 by Andrew Wang)

v0.2.1

New Features

• Mirror Descent algorithm with Bregman potentials (#282 by Samuel Hurault)

• Added Gaussian-weighted splitting mask (from Yaman et al.), Artifact2Artifact (Liu et al.) and Phase2Phase (Eldeniz et al.) (#279 by Andrew Wang)

• Added time-agnostic network wrapper (#279 by Andrew Wang)

• Add sinc filter (#280 by Julian Tachella)

• Add Noise2Score method (#280 by Julian Tachella)

• Add Gamma Noise (#280 by Julian Tachella)

• Add 3D Blur physics operator, with 3D diffraction microscope blur generators (#277 by Florian Sarron, Pierre Weiss, Paul Escande, Minh Hai Nguyen) - 12/07/2024

• Add ICNN model (#281 by Samuel Hurault)

• Dynamic MRI physics operator (#242 by Andrew Wang)

• Add support for adversarial losses and models (GANs) (#183 by Andrew Wang)

• Base transform class for transform arithmetic (#240 by Andrew Wang) - 26/06/2024.

• Plot video/animation functionality (#245 by Andrew Wang)

• Added update_parameters for parameter-dependent physics (#241 by Julian Tachella) - 11/06/2024

• Added evaluation functions for R2R and Splitting losses (#241 by Julian Tachella) - 11/06/2024

• Added a new Physics class for the Radio Interferometry problem (#230 by Chao Tang, Tobias Liaudat) - 07/06/2024

• Add projective and affine transformations for EI or data augmentation (#173 by Andrew Wang)

• Add k-t MRI mask generators using Gaussian, random uniform and equispaced sampling stratgies (#206 by Andrew Wang)

• Added Lidc-Idri buit-in datasets (#270 by Maxime SONG) - 12/07/2024

• Added Flickr2k / LSDIR / Fluorescent Microscopy Denoising buit-in datasets (#276 by Maxime SONG) - 15/07/2024

• Added rng a random number generator to each deepinv.physics.generator.PhysicsGenerator and a seed number argument to deepinv.physics.generator.PhysicsGenerator.step() (by Minh Hai Nguyen)

• Added an equivalent of deepinv.physics.functional.random_choice() in torch, available as deepinv.physics.functional.random_choice() (by Minh Hai Nguyen)

Fixed

• Disable unecessary gradient computation to prevent memory explosion (#301 by Dylan Sechet, Samuel Hurault)

• Wandb logging (#280 by Julian Tachella)

• SURE improvements (#280 by Julian Tachella)

• Fixed padding in conv_transpose2d and made conv_2d a true convolution (by Florian Sarron, Pierre Weiss, Paul Escande, Minh Hai Nguyen) - 12/07/2024

• Fixed the gradient stopping in EILoss (#263 by Jérémy Scanvic) - 27/06/2024

• Fixed averaging loss over epochs deepinv.Trainer (#241 by Julian Tachella) - 11/06/2024

• Fixed deepinv.Trainer save_path timestamp problem on Windows (#245 by Andrew Wang)

• Fixed inpainting/SplittingLoss mask generation + more flexible tensor size handling + pixelwise masking (#267 by Andrew Wang)

• Fixed the deepinv.physics.generator.ProductConvolutionBlurGenerator, allowing for batch generation (previously does not work) by (Minh Hai Nguyen)

Changed

• Redefine Prior, DataFidelity and Bregman with a common parent class Potential (#282 by Samuel Hurault)

• Changed to Python 3.9+ (#280 by Julian Tachella)

• Improved support for parameter-dependent operators (#227 by Jérémy Scanvic) - 28/05/2024

• Added a divergence check in the conjugate gradient implementation (#225 by Jérémy Scanvic) - 22/05/2024

v0.2.0

Many of the features in this version were developed by Minh Hai Nguyen, Pierre Weiss, Florian Sarron, Julian Tachella and Matthieu Terris during the IDRIS hackathon.

New Features

• Added a parameterization of the operators and noiselevels for the physics class

• Added a physics.functional submodule

• Modified the Blur class to handle color, grayscale, single and multi-batch images

• Added a PhysicsGenerator class to synthetize parameters for the forward operators

• Added the possibility to sum generators

• Added a MotionBlur generator

• Added a DiffractionBlur generator

• Added a MaskGenerator for MRI

• Added a SigmaGenerator for the Gaussian noise

• Added a tour of blur operators

• Added ProductConvolution expansions

• Added a ThinPlateSpline interpolation function

• Added d-dimensional histograms

• Added GeneratorMixture to mix physics generators

• Added the SpaceVarying blur class

• Added the SpaceVarying blur generators

• Added pytests and examples for all the new features

• A few speed ups by carefully profiling the training codes

• made sigma in drunet trainable

• Added deepinv.Trainer, deepinv.loss.Loss class and eval metric (LPIPS, NIQE, SSIM) (#181 by Julian Tachella) - 02/04/2024

• PhaseRetrieval class (#176 by Zhiyuan Hu) - 20/03/2024

• Added 3D wavelets (#164 by Matthieu Terris) - 07/03/2024

• Added patch priors loss (#164 by Johannes Hertrich) - 07/03/2024

• Added Restormer model (#185 by Antoine Regnier and Maxime SONG) - 18/04/2024

• Added DIV2K built-in dataset (#203 by Maxime SONG) - 03/05/2024

• Added Urban100 built-in dataset (#237 by Maxime SONG) - 07/06/2024

• Added Set14 / CBSD68 / fastMRI buit-in datasets (#248 #249 #229 by Maxime SONG) - 25/06/2024

Fixed

• Fixed the None prior (#233 by Samuel Hurault) - 04/06/2024

• Fixed the conjugate gradient torch.nograd for teh demos, accelerated)

• Fixed torch.nograd in demos for faster generation of the doc

• Corrected the padding for the convolution

• Solved pan-sharpening issues

• Many docstring fixes

• Fixed slow drunet sigma and batched conjugate gradient (#181 by Minh Hai Nguyen) - 02/04/2024

• Fixed g dependence on sigma in optim docs (#165 by Julian Tachella) - 28/02/2024

Changed

• Refactored the documentation completely for the physics

• Refactor unfolded docs (#181 by Julian Tachella) - 02/04/2024

• Refactor model docs (#172 by Julian Tachella) - 12/03/2024

• Changed WaveletPrior to WaveletDenoiser (#165 by Julian Tachella) - 28/02/2024

• Move from torchwavelets to ptwt (#162 by Matthieu Terris) - 22/02/2024

v0.1.1

New Features

• Added r2r loss (#148 by Brayan Monroy) - 30/01/2024

• Added scale transform (#135 by Jérémy Scanvic) - 19/12/2023

• Added priors for total variation and l12 mixed norm (#156 by Nils Laurent) - 09/02/2023

Fixed

• Fixed issue in noise forward of Decomposable class (#154 by Matthieu Terris) - 08/02/2024

• Fixed new black version 24.1.1 style changes (#151 by Julian Tachella) - 31/01/2024

• Fixed test for sigma as torch tensor with gpu enable (#145 by Brayan Monroy) - 23/12/2023

• Fixed #139 BM3D tensor format grayscale (#140 by Matthieu Terris) - 23/12/2023

• Fixed #136 noise additive model for DecomposablePhysics (#138 by Matthieu Terris) - 22/12/2023

• Importing deepinv does not modify matplotlib config anymore (:gh`1501` by Thomas Moreau) - 30/01/2024

Changed

• Rephrased the README (#142 by Jérémy Scanvic) - 09/01/2024

v0.1.0

New Features

• Added autoadjoint capabilities (#151 by Julian Tachella) - 31/01/2024

• Added equivariant transforms (#125 by Matthieu Terris) - 07/12/2023

• Moved datasets and weights to HuggingFace (#121 by Samuel Hurault) - 01/12/2023

• Added L1 prior, change distance in DataFidelity (#108 by Samuel Hurault) - 03/11/2023

• Added Kaiming init (#102 by Matthieu Terris) - 29/10/2023

• Added Anderson Acceleration (#86 by Samuel Hurault) - 23/10/2023

• Added deepinv.sampling.DPS() diffusion method (#92 by Julian Tachella and Hyungjin Chung) - 20/10/2023

• Added on-the-fly physics computations in training (#88 by Matthieu Terris) - 10/10/2023

• Added no_grad parameter (#80 by Jérémy Scanvic) - 20/08/2023

• Added prox of TV (#79 by Matthieu Terris) - 16/08/2023

• Added diffpir demo + model (#77 by Matthieu Terris) - 08/08/2023

• Added SwinIR model (#76 by Jérémy Scanvic) - 02/08/2023

• Added hard-threshold (#71 by Matthieu Terris) - 18/07/2023

• Added discord server (#64 by Julian Tachella) - 10/07/2023

• Added changelog file (#64 by Julian Tachella) - 10/07/2023

Fixed

• doc fixes + training fixes (#124 by Julian Tachella) - 06/12/2023

• Add doc weights (#97 by Matthieu Terris) - 24/10/2023

• Fix BlurFFT adjoint (#89 by Matthieu Terris) - 15/10/2023

• Doc typos (#88 by Matthieu Terris) - 10/10/2023

• Minor fixes DiffPIR + other typos (#81 by Matthieu Terris) - 10/09/2023

• Call wandb.init only when needed (#78 by Jérémy Scanvic) - 09/08/2023

• Log epoch loss instead of batch loss (#73 by Jérémy Scanvic) - 21/07/2023

• Automatically disable backtracking is no explicit cost (#68 by Samuel Hurault) - 12/07/2023

• Added missing indent (#63 by Jérémy Scanvic) - 12/07/2023

• Fixed get_freer_gpu grep statement to work for different versions of nvidia-smi (#82 by Alexander Mehta) - 20/09/2023

• Fixed get_freer_gpu to work on different operating systems (#87 by Andrea Sebastiani) - 10/10/2023

• Fixed Discord server and contributiong links (#87 by Andrea Sebastiani) - 10/10/2023

Changed

• Update CI (#95 #99 by Thomas Moreau) - 24/10/2023

• Changed normalization CS and SPC to 1/m (#72 by Julian Tachella) - 21/07/2023

• Update docstring (#68 by Samuel Hurault) - 12/07/2023