ParamTransformModule

class dnois.torch.ParamTransformModule(*args, **kwargs)

A subclass of torch.nn.Module that can create and manage transformed parameters.

Transformed parameter can be registered given either its nominal value (by calling overridden register_parameter()) or latent value (by calling register_latent_parameter()). Note that in first case an inverse transformation must be given to compute the latent value. If a parameter with same name has been registered already, it will be automatically converted to a transformed parameter. Specifying a transformation for existent parameter by calling set_transform() has same effect. Transformations are specified in the form of a Transform instance.

Note

Implementation Detail The latent value of each transformed parameter is a torch.nn.Parameter attribute with name like _latent_<param>.

register_latent_parameter(name: str, param: Parameter | None, transform: Transform)

Similar to register_parameter(), but takes as input the latent value rather than nominal value. In this way, the inverse transformation need not be provided since the initial latent value is known.

If name corresponds to a vanilla parameter (i.e. not transformed parameter) it will be converted to a transformed one.

Parameters:

• name (str) – Name of the parameter.

• param (Parameter) – Latent torch.nn.Parameter instance to be registered.

• transform (Transform) – Transformation object.

register_parameter(name: str, param: Parameter | None, transform: Transform = None) → None

Similar to torch.nn.Module.register_parameter(), but allows you to register a transformed parameter as long as transform is given.

If name corresponds to a vanilla parameter (i.e. not transformed parameter) but transform is given, it will be converted to a transformed one.

Parameters:

• name (str) – Name of the parameter.

• param (Parameter) – Nominal torch.nn.Parameter instance to be registered.

• transform (Transform) – Transformation object.

remove_transform(name: str)

Remove transformation for parameter name.

Parameters:

name (str) – Name of the parameter.

set_transform(name: str, transform: Transform)

Set transformation for parameter name.

If name corresponds to a vanilla parameter (i.e. not transformed parameter) it will be converted to a transformed one.

Parameters:

• name (str) – Name of the parameter.

• transform (Transform) – Transformation object.

property nominal_values: dict[str, Tensor]

A dict whose keys are names of all parameters of this module and values are their values. The values are nominal ones for transformed parameters.

Type:

dict[str, Tensor]

property transformed_parameters: dict[str, tuple[Parameter, Transform]]

A dict whose keys are names of all transformed parameters of this module. The value corresponding to each key is a tuple containing:

• The latent value, a torch.nn.Parameter instance;

• Corresponding transformation object.

Type:

dict[str, tuple[Parameter, Transform]]