deepinv.distributed

This module provides a simplified API for distributing DeepInverse objects across multiple devices and processes. The core function distribute() automatically wraps your objects (stacked physics, denoisers, data fidelity) into their distributed counterparts, handling all the boilerplate for you.

See the user guide on distributed reconstruction for more information.

Main API

These are the main components most users need:

deepinv.distributed.DistributedContext

Context manager for distributed computing.

deepinv.distributed.distribute

Distribute a DeepInverse object across multiple devices.

Core Classes

These classes are created automatically by deepinv.distributed.distribute(). You typically don’t need to instantiate them directly.

deepinv.distributed.framework.DistributedStackedPhysics	This class distributes a collection of physics operators across multiple processes, where each process owns a subset of the operators.
deepinv.distributed.framework.DistributedStackedLinearPhysics	Distributed linear physics operators.
deepinv.distributed.framework.DistributedProcessing	Distributed signal processing using pluggable tiling and reduction strategies.
deepinv.distributed.framework.DistributedDataFidelity	Distributed data fidelity term for use with distributed physics operators.
deepinv.distributed.framework.DistributedReplicatedParameters	Synchronize gradients for replicated trainable parameters.

Distribution Strategies

Advanced: Custom tiling strategies for spatial distribution of denoisers.

deepinv.distributed.strategies.DistributedSignalStrategy	Abstract base class for distributed signal processing strategies.
deepinv.distributed.strategies.distributed_strategies.TilingStrategy	Smart tiling strategy with padding for N-dimensional data.

deepinv.distributed.strategies.create_strategy

Create a distributed signal strategy.