LION.models.post_processing.cFBPConvNet

Classes

ConvODEFunc(channels[, relu_type, ...])	The right-hand side of a first-order neural ODE: convolutions through the channels listed, with batch normalisation and relu_type.
ConvSODEFunc(channels[, relu_type, ...])	The right-hand side of a second-order neural ODE: the state is (position, velocity) along the channels.
Down()	Downscaling with maxpool
InitialVelocity(channels)	The initial velocity of a second-order neural ODE, from the initial position.
ODEBlock(odefunc[, tol, adjoint, solver])	Solves the neural ODE odefunc from time 0 to 1 (torchdiffeq).
Up(channels[, stride, relu_type])	Upscaling with transpose conv
cFBPConvNet([operator_params, model_parameters])	Continuous FBPConvNet: FBPConvNet with every convolution block replaced by a neural ODE block.

---

class LION.models.post_processing.cFBPConvNet.ConvODEFunc(channels, relu_type='ReLU', relu_last=True, kernel_size=3)

Bases: Module

The right-hand side of a first-order neural ODE: convolutions through the channels listed, with batch normalisation and relu_type.

Initialize internal Module state, shared by both nn.Module and ScriptModule.

forward(t, x)

dx/dt at x (t unused: autonomous).

class LION.models.post_processing.cFBPConvNet.ConvSODEFunc(channels, relu_type='ReLU', relu_last=True, kernel_size=3)

Bases: Module

The right-hand side of a second-order neural ODE: the state is (position, velocity) along the channels.

Initialize internal Module state, shared by both nn.Module and ScriptModule.

forward(t, x)

The state’s derivative at x: (velocity, acceleration).

class LION.models.post_processing.cFBPConvNet.Down

Bases: Module

Downscaling with maxpool

Initialize internal Module state, shared by both nn.Module and ScriptModule.

forward(x)

x halved in size (max pooling).

class LION.models.post_processing.cFBPConvNet.InitialVelocity(channels)

Bases: Module

The initial velocity of a second-order neural ODE, from the initial position.

Initialize internal Module state, shared by both nn.Module and ScriptModule.

forward(x0)

The ODE’s state at time 0: the position x0 and its velocity, along the channels.

class LION.models.post_processing.cFBPConvNet.ODEBlock(odefunc, tol=0.001, adjoint=False, solver='rk4')

Bases: Module

Solves the neural ODE odefunc from time 0 to 1 (torchdiffeq).

odefunc: the ODE’s right-hand side; tol: the solver’s relative and absolute tolerance; adjoint: gradients by the adjoint method (less memory); solver: a torchdiffeq method.

forward(x, eval_times=None, solver=None)

The ODE’s solution at time 1 from x (or at each of eval_times), with solver (the block’s if None).

class LION.models.post_processing.cFBPConvNet.Up(channels, stride=2, relu_type='ReLU')

Bases: Module

Upscaling with transpose conv

Initialize internal Module state, shared by both nn.Module and ScriptModule.

forward(x)

x doubled in size (transposed convolution).

class LION.models.post_processing.cFBPConvNet.cFBPConvNet(operator_params=None, model_parameters=None)

Bases: LIONmodel

Continuous FBPConvNet: FBPConvNet with every convolution block replaced by a neural ODE block. Takes the FDK reconstruction as input (IMAGE) and returns input + residual, like FBPConvNetImage.

Initialize internal Module state, shared by both nn.Module and ScriptModule.

static default_parameters()

FBPConvNet’s channels, each block a neural ODE (its input lifted by a 1x1 convolution), and the ODEs’ settings:

• tol: the solver’s tolerance (1e-3);

• ode_adjoint: gradients by the adjoint method (False);

• do_second_order: second-order ODEs (False; not implemented yet);

• solver: a torchdiffeq method (“rk4”).

static cite(cite_format='bib')

Where cFBPConvNet comes from: its citation, BibTeX (“bib”) or “MLA”.

forward(x)

x: FDK reconstructions, (B, 1, H, W).