cdopt.nn.module.utils

cdopt.nn.module.utils#

cdopt.nn.module.utils contains some useful functions for cdopt.nn module, and all its contained functions can be accessed in cdopt.nn.

get_quad_penalty#

sum_quad_pen = cdopt.nn.get_quad_penalty(module)

The function that returns the sum of all the quadratic penalty terms in the module.

Input:

  • module: (torch.nn.Module) The torch.nn.Module class that composed of the layers from torch.nn and cdopt.nn.

Output:

  • The sum of all the quadratic penalty terms in the children of the module, where the penalty parameter equals to penalty_param attributes of each children of the module.

get_constraint_violation#

sum_feas_violation = cdopt.nn.get_constraint_violation(module, ord = None, **kwargs)

The function that returns the feasibility of the module.

Input:

  • module: (torch.nn.Module) The torch.nn.Module class that composed of the layers from torch.nn and cdopt.nn.

Output:

  • The feasibility violation of the variables in module. In this function, we first generate the con_vec whose entries represents the feasibility of each layers. Then the function returns torch.linalg.norm(con_vec, ord = ord, **kwargs).

wvt_flatten2d#

weight_to_var, var_to_weight, var_shape = cdopt.nn.wvt_flatten2d(tensor_shape)

The function that determines the tensor of the shape \((n_1, n_2,...n_m)\) be converted to an 2D tensor that of the shape \((n_1 * n_3 * n_4 *...* n_m, n_2)\).

Inputs:

  • tensor_shape: (Tuple of ints) The shape of the weight matrix of the layer.

Outputs:

  • weight_to_var: (Callable) The function that transform the weight matrix of the layer to the 2D tensor.

  • var_to_weight: (Callable) The function that transform the 2D tensor to the weight matrix of the layer.

  • **var_shape **: (Tuple of ints) The shape of the 2D tensor.

wvt_flatten2d_transp#

weight_to_var, var_to_weight, var_shape = cdopt.nn.wvt_flatten2d_transp(tensor_shape)

The function that determines the tensor of the shape \((n_1, n_2,...n_m)\) be converted to an 2D tensor that of the shape \((n_2 * n_3 * n_4 *...* n_m, n_1)\).

Inputs:

  • tensor_shape: (Tuple of ints) The shape of the weight matrix of the layer.

Outputs:

  • weight_to_var: (Callable) The function that transform the weight matrix of the layer to the 2D tensor.

  • var_to_weight: (Callable) The function that transform the 2D tensor to the weight matrix of the layer.

  • **var_shape **: (Tuple of ints) The shape of the 2D tensor.

wvt_identical#

weight_to_var, var_to_weight, var_shape = cdopt.nn.wvt_identical(tensor_shape)

The function that determines the tensor of the shape \((n_1, n_2,...n_m)\) be converted to an tensor that of the same shape.

Inputs:

  • tensor_shape: (Tuple of ints) The shape of the weight matrix of the layer.

Outputs:

  • weight_to_var: (Callable) The function that transform the weight matrix of the layer to the 2D tensor.

  • var_to_weight: (Callable) The function that transform the 2D tensor to the weight matrix of the layer.

  • **var_shape **: (Tuple of ints) The shape of the tensor.

Contents