LION.models.learned_regularizer.ACR

Classes

ACR([operator_params, model_parameters])	Adversarial convex regulariser (Mukherjee et al., 2021): as the adversarial regulariser (AR), but convex in the image: an input-convex network plus a small learnable ||x||^2 term.
ICNN_layer(channels[, kernel_size, stride, ...])	A layer of an input-convex network: a non-negative convolution of the previous layer's output plus a convolution of the input, then a convex, non-decreasing activation (leaky ReLU).
L2net([l2_penalty])	softplus(w) * ||x||^2 per sample, with a learnable weight w.
Positive(*args, **kwargs)	A parametrisation keeping a weight non-negative (clipped at 0), for the convex layers.

---

class LION.models.learned_regularizer.ACR.ACR(operator_params=None, model_parameters=None)

Bases: LIONmodel

Adversarial convex regulariser (Mukherjee et al., 2021): as the adversarial regulariser (AR), but convex in the image: an input-convex network plus a small learnable ||x||^2 term. Trained by ARSolver, used by LearnedRegularizerReconstructor. Takes images.

model_parameters (default_parameters()):

• channels: the channels of each layer (16);

• kernel_size, stride: the convolutions’ (5, 1);

• relu_type: the activation, “LeakyReLU” (the only convex, non-decreasing one supported);

• layers: the convex layers (5);

• l2_penalty: the initial w of the softplus(w) ||x||^2 term (-9).

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

initialize_weights(min_val=0.0, max_val=0.001)

Draws the non-negative weights uniformly in [min_val, max_val], as the paper does. Returns the model.

forward(x)

The regulariser’s value for each image of x: (B, 1).

static default_parameters()

The paper’s ACR: 5 convex layers of 16 channels, 5x5 kernels (see the class).

static cite(cite_format='bib')

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

class LION.models.learned_regularizer.ACR.ICNN_layer(channels, kernel_size=5, stride=1, relu_type='LeakyReLU')

Bases: Module

A layer of an input-convex network: a non-negative convolution of the previous layer’s output plus a convolution of the input, then a convex, non-decreasing activation (leaky ReLU).

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

forward(z, x0)

The layer’s output from the previous one, z, and the network’s input x0.

class LION.models.learned_regularizer.ACR.L2net(l2_penalty=-9.0)

Bases: Module

softplus(w) * ||x||^2 per sample, with a learnable weight w.

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

forward(x)

softplus(w) * ||x||^2 for each image of x: (B, 1).

class LION.models.learned_regularizer.ACR.Positive(*args, **kwargs)

Bases: Module

A parametrisation keeping a weight non-negative (clipped at 0), for the convex layers.

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

forward(X)

X clipped at 0.