# User Guide Imaging inverse problems are described by the equation $y = \noise{\forw{x}}$ where $x$ is an unknown signal (image, volume, etc.) we want to recover, $y$ are the observed measurements, $A$ is a deterministic (linear or non-linear) operator capturing the physics of the acquisition and $N$ characterizes the noise affecting the measurements. ## Operators The library provides a large variety of imaging operators `physics` modelling $\noise{\forw{\cdot}}$, which can simulate the observation process: ```default x = load_image() y = physics(x) # simulate observation ``` **Introduction** Introduction to the physics module. **Operators** Forward operators and noise distributions. **Functional** Various utilities for forward operators. ## Reconstruction In order to recover an image from its measurements, the library provides many reconstruction methods $\hat{x}=R(y, A)$, which often leverage knowledge of the acquisition `physics`. Given a restoration model `model`, the reconstruction is therefore provided as ```default x_hat = model(y, physics) # reconstruct signal ``` **Introduction** Introduction and summary of reconstruction algorithms. **Pseudoinverse** Simplest reconstruction without priors or training. **Pretrained Models** Ready-to-use pretrained models and denoisers. **Denoisers** Classical and deep denoisers with pretrained weights. **Deep Reconstruction Methods** Deep model architectures for reconstruction. **Optimization** Priors and data-fidelity functions, and optimization algorithms. **Unfolded Algorithms** Unfolded architectures. **Iterative Reconstruction** Plug-and-play, RED, variational methods. **Sampling** Diffusion and MCMC algorithms. **Adversarial Reconstruction** Conditional, unconditional GANs and deep image prior. **Blind Inverse Problems** Reconstruction with misspecified physics and/or noise level. ## Training, Testing and Utilities All the tools from the library, from measurement operator to restoration methods, are implemented as [`torch.nn.Module`](https://docs.pytorch.org/docs/stable/generated/torch.nn.Module.html#torch.nn.Module) and therefore natively support backpropagation. Reconstruction networks `model` can be trained on datasets to improve their performance: ```default trainer = Trainer(model, loss, optimizer, metric, train_dataset, ...) trainer.train() trainer.test(test_dataset) ``` **Training** Training and testing reconstruction models. **Datasets** Utilities to generate and load datasets for training and testing. **Loss** Supervised and self-supervised losses to train the models. **Metrics** Distortion and perceptual metrics to evaluate reconstructions. **Transforms** Transforms for data augmentation and self-supervised learning. **Utils** Plotting and other utilities. ## Distributed Computing In large-scale image reconstruction problems, one GPU might not be enough to handle large images/volumes or large forward operators. We provide multi-GPU tools specific to image reconstruction with `deepinv.distributed`: ```default physics = deepinv.distributed.distribute(physics) # distrubute physics denoiser = deepinv.distributed.distribute(denoiser) # distribute denoiser ``` **Distributed Reconstruction** Distributed reconstruction for large-scale images and volumes. **Distributed Training** Distributed training for large-scale images and volumes.