API References#
TorchSim is organized around one idea: a signal model is the only thing a sequence has to supply, and everything else – differentiation, execution across devices, estimation, reconstruction, design – is written once against that interface.
The pages below follow the subpackages.
- Signal models
What a sequence of your own is written from: the physics, saying what each kind of event does to the spins, and a simulator saying what order they are played in.
- Simulators
The sequences that ship with TorchSim, as classes and as one-call functional wrappers.
- Sequences
The device-agnostic description an acquisition is assembled from: events, the operators that emit them, and builders for whole sequences.
- Parameter estimation
Estimating tissue properties from a measured volume: dictionary matching, lookup tables, nonlinear least squares, kernel regression.
- Model-based reconstruction
Solving for parameter maps straight from k-space, with the signal model inside the forward operator, and the temporal basis that makes it cheap.
- Sequence design
Choosing a sequence’s parameters by minimizing a cost you write.
- Running a simulation
Running a description: the differentiable engine, transmit calibration, and where a per-voxel workload is placed.