PyTorch guide#

mononet.torch provides monotonic layers as torch.nn.Module subclasses. They drop into any existing training loop (plain PyTorch, PyTorch Lightning, etc.) and compose with the native torch.nn.Sequential.

Install#

pip install "mononet[torch]"

Public API#

mononet.torch.layers.MonoLinear — monotonic analogue of torch.nn.Linear (non-decreasing in all inputs).
mononet.torch.layers.MonoResidual — dual-gated monotone residual block, warm-started near identity.
mononet.torch.layers.MonoInput — sign-flip layer encoding per-feature monotonicity directions.

mononet ships layers only — stack them yourself; there is no composed MonoMLP model.

Example#

import torch
from mononet.torch import MonoInput, MonoLinear

# A monotonic MLP: non-decreasing in every input feature.
net = torch.nn.Sequential(
    MonoInput(1),                     # +1 => non-decreasing; -1 => non-increasing
    MonoLinear(4, 32, mode="switch"),
    MonoLinear(32, 1, mode="switch"),
)
y = net(torch.randn(8, 4))            # (8, 1), guaranteed monotone in all inputs

For per-feature monotonicity directions, pass a MonotonicityMask (a 1-D array of {-1, +1}) to MonoInput.

See also#