cdopt.nn.utils.set_constraints

set_constraints.set_constraint_dissolving

set_constraint_dissolving(module, attr_name, manifold_class = euclidean_torch, weight_var_transfer = None, manifold_args = {}, penalty_param = 0)

Set the manifold constraints to the attribute attr_name to the Module module.

Parameters:

• module (torch.nn.Module) – the module to be set the manifold constraints

• attr_name (str) – the name of the attribute to be set the manifold constraints

• manifold_class (cdopt.manifold) – the manifold classes from cdopt.manifold_torch.

• weight_var_transfer (Callable) – the function that determines how the variables from manifold_class are transformed to the attribute attr_name of the module.

• manifold_args (dict) – arguments to be passed when instantiating the manifold class from manifold_class

• penalty_param (float) – the value of the penalty parameter.

Returns:

• The module with manifold constraints on its attribute attr_name.

Note

The set_constraint_dissolving() function is developed based on the torch.nn.utils.parametrize functions. Therefore, the module returned belongs to the ParametrizationList module in PyTorch, which is different from the predefined neural layers from cdopt.nn.

The list of parametrizations on the tensor weight will be accessible under module.parametrizations. And the variables of the manifold will be accessible under module.parametrizations[attr_name].original. Moreover, the constraint dissolving mapping \(\mathcal{A}\) can be accessed at module.parametrizations[attr_name].A, the constriants can be accessed at module.parametrizations[attr_name].C, the quadratic penalty term can be accessed at module.parametrizations[attr_name].quad_penalty(), and the penalty parameter can be accessed at module.parametrizations[attr_name].penalty_param.