def test_bias_false(self):
lc = LocallyConnected(1, 1, 1, bias=False)
assert lc.bias is None
def __init__(
self,
n_features: int,
dist_types: List[DistTypeBase],
use_bias: bool = False,
hidden_layer_units: Iterable[int] = (0, ),
bounds: List[Tuple[int, int]] = None,
lasso_beta: float = 0.0,
ridge_beta: float = 0.0,
nonlinear_clamp: float = 1e-2,
):
"""
Constructor for NOTEARS MLP class.
Args:
n_features: number of input features.
dist_types: list of data type objects used to fit the NOTEARS algorithm.
use_bias: True to add the intercept to the model
hidden_layer_units: An iterable where its length determine the number of layers used,
and the numbers determine the number of nodes used for the layer in order.
bounds: bound constraint for each parameter.
lasso_beta: Constant that multiplies the lasso term (l1 regularisation).
It only applies to dag_layer weight.
ridge_beta: Constant that multiplies the ridge term (l2 regularisation).
It applies to both dag_layer and loc_lin_layer weights.
nonlinear_clamp: Value used to soft clamp the nonlinear layer normalisation.
Prevents the weights from being scaled above 1/nonlinear_clamp.
"""
super().__init__()
self.device = torch.device("cpu")
self.lasso_beta = lasso_beta
self.ridge_beta = ridge_beta
self.nonlinear_clamp = nonlinear_clamp
# cast to list for later concat.
self.dims = ([n_features] + list(hidden_layer_units) +
[1] if hidden_layer_units[0] else [n_features, 1])
# dag_layer: initial linear layer
self.dag_layer = nn.Linear(self.dims[0],
self.dims[0] * self.dims[1],
bias=use_bias).float()
nn.init.zeros_(self.dag_layer.weight)
if use_bias:
nn.init.zeros_(self.dag_layer.bias)
# loc_lin_layer: local linear layers
layers = [
LocallyConnected(self.dims[0],
input_features,
output_features,
bias=use_bias).float() for input_features,
output_features in zip(self.dims[1:-1], self.dims[2:])
]
self._loc_lin_layer_weights = nn.ModuleList(layers)
for layer in layers:
layer.reset_parameters()
# set the bounds as an attribute on the weights object
self.dag_layer.weight.bounds = bounds
# set the dist types
self.dist_types = dist_types
# type the adjacency matrix
self.adj = None
self.adj_mean_effect = None
def test_bias_true(self):
lc = LocallyConnected(1, 1, 1, bias=True)
assert isinstance(lc.bias, nn.Parameter)