pmlayer.torch.layers.PiecewiseLinear

Monotonically increasing piesewise linear layer for PyTorch.

class PiecewiseLinear(
    input_keypoints,
    num_dims=1,
    output_min=0.0,
    output_max=1.0,
    indices_increasing=[],
    indices_decreasing=[],
    indices_transform=[])

This layer transforms each input feature by using a piecewise linear function. Note that the input features are handled individually. The output of this layer is monotonically increasing with respect to the input features specified by indices_increasing, and the output of this layer is monotonically decreasing with respect to the input features specified by indices_decreasing. The parameter input_keypoints specifies the knots of the piecewise linear function.

Parameters

Args	Type	Description
input_keypoints	Tensor	The boundaries of input features.
num_dims	int	The number of input features.
output_min	float	The minimum value of the output.
output_max	float	The maximum value of the output.
indices_increasing	list of int	The list of indices of monotonically increasing features. Default: [].
indices_decreasing	list of int	The list of indices of monotonically decreasing features. Default: [].
indices_transform	list of int	The list of indices of features that are transformed by using piece-wise linear function without monotonicity constraints. Default: [].

Tensor Shape

I/O	Shape
Input	(N, num_dims)
Output	(N, num_dims)

N usually corresponds the batch size.

Example

The following code transforms input tensor x into output tensor y, and the output (the shape of y) is (128,10).

b = torch.linspace(0.0, 1.0, steps=5)
l = PiecewiseLinear(b, 10, indices_increasing=[2,3])
x = torch.randn(128, 10)
y = l(x)
print(y.shape)