Learning

bartorch.learning holds neural networks for complex, high-dimensional images and what places them in this package’s iterations: a network for images carrying frames or subspace coefficients, patchwise execution on a device of an image held on the host, unrolled iterations, the partition of the acquired samples for training without a reference, and a voxel-wise uncertainty. A network used as a regularizer is a ImplicitPrior, whose sigma may be a schedule over the iterations and which with step=True hands the network the iteration index. Differentiation through reconstruction describes how gradients pass through an unrolled iteration and the memory each backward strategy uses.

Networks

Object	Description
UNet	Residual UNet on 2D or 3D real channels, optionally over a frame axis, conditioned on the iteration index, the noise level and a class
ComplexNet	A real network applied to complex images, their leading axes and real and imaginary parts as its channels, normalized by the peak or whitened
Patchwise	A network applied patch by patch on a device, with a randomly shifted grid and mixed precision, to an image held elsewhere

Precision on the device	Patchwise(dtype="auto")
Card with bfloat16 (Ampere and later: A40, RTX 4000 Ada)	bfloat16
Card without it (Turing: T4)	float16
Host	single precision, no mixed precision

Iterations as networks

Object	Description
Unrolled	A fixed number of applications of an iteration block from bartorch.optim or bartorch.nlop, with shared or per-iteration parameters

Unrolled setting	Graph recorded	Memory	Gradient
default	The whole stack	Iterations × one step	End to end
detach=True	One iteration at a time	One step	Of each step alone (greedy training with Unrolled.steps())
checkpoint=True	The states between iterations	States + one step	End to end; each step is computed twice

Training without a reference

Object	Description
split	Partition of the acquired samples into a set reconstructed from and a set held out, for self-supervised training

Uncertainty

Object	Description
moments	Voxel-wise mean and variance of a randomized reconstruction repeated
calibrate	The factor turning a spread into an interval with a stated coverage, by split conformal calibration

Complex values as channels

Object	Description
as_real	Real and imaginary parts on a new leading axis
as_complex	Inverse of as_real

Training stages

These two need lightning and torchio, which pip install bartorch[learning] installs; they are imported the first time either name is asked for, so import bartorch.learning imports neither.

Object	Description
Reconstruction	A LightningModule training a denoiser, an unrolled network greedily, or an unrolled network end to end, supervised or self-supervised, with the batch left on the host
RandomGain	A torchio transform multiplying every image of a subject by one random complex gain

The examples of Learned regularization train networks built from these objects.