torch_openreml.covariance.simple_param_specs¶

torch_openreml.covariance.simple_param_specs(n, default=None, trans=None)[source]¶

Create a parameter specification dictionary with n parameters.

Each parameter is named "theta_0", "theta_1", …, is not fixed, and uses a common transform for all parameters. All parameters share the same default value.

Parameters:

n (int) – Number of parameters to create.
default (torch.Tensor, optional) – Default value for each parameter, given as a 1D tensor of shape (1,). Defaults to torch.tensor([0.0]).
trans (Transform, optional) – Transform to apply to all parameters. Defaults to TransformIdentity (unconstrained).

Returns:

A dictionary mapping parameter names to specification dicts of the form {"fixed": False, "default": tensor, "trans": trans}.

Return type:

dict

Raises:

TypeError – If default is not a Torch tensor.
ValueError – If default does not have shape (1,).

Example:

from torch_openreml.covariance.param import simple_param_specs

simple_param_specs(3)

{'theta_0': {'fixed': False,
  'default': tensor([0.]),
  'trans': TransformIdentity()},
 'theta_1': {'fixed': False,
  'default': tensor([0.]),
  'trans': TransformIdentity()},
 'theta_2': {'fixed': False,
  'default': tensor([0.]),
  'trans': TransformIdentity()}}

import torch

simple_param_specs(2, default=torch.tensor([1.5]))

{'theta_0': {'fixed': False,
  'default': tensor([1.5000]),
  'trans': TransformIdentity()},
 'theta_1': {'fixed': False,
  'default': tensor([1.5000]),
  'trans': TransformIdentity()}}