source/model.py

from __future__ import annotations

import torch
from torch import nn
from torch.nn import functional as F


class LinearSplineLayer(nn.Module):
    def __init__(
        self,
        input_dimensions: int,
        output_dimensions: int,
        grid_points: int,
    ) -> None:
        super().__init__()
        centers = torch.linspace(-1, 1, grid_points)
        self.register_buffer("centers", centers)
        self.spacing = float(centers[1] - centers[0])
        self.coefficients = nn.Parameter(
            torch.randn(output_dimensions, input_dimensions, grid_points) * 0.05
        )
        self.base_weight = nn.Parameter(
            torch.randn(output_dimensions, input_dimensions) * 0.1
        )
        self.bias = nn.Parameter(torch.zeros(output_dimensions))

    def forward(self, inputs: torch.Tensor) -> torch.Tensor:
        distance = torch.abs(inputs[:, :, None] - self.centers)
        basis = F.relu(1 - distance / self.spacing)
        spline = torch.einsum("big,oig->bo", basis, self.coefficients)
        return spline + F.linear(inputs, self.base_weight, self.bias)


class SplineKAN(nn.Module):
    def __init__(self) -> None:
        super().__init__()
        self.first = LinearSplineLayer(2, 16, 14)
        self.second = LinearSplineLayer(16, 1, 8)

    def forward(self, inputs: torch.Tensor) -> torch.Tensor:
        hidden = torch.tanh(self.first(inputs))
        return self.second(hidden)


class MatchedMLP(nn.Module):
    def __init__(self) -> None:
        super().__init__()
        self.network = nn.Sequential(
            nn.Linear(2, 32),
            nn.Tanh(),
            nn.Linear(32, 16),
            nn.Tanh(),
            nn.Linear(16, 1),
        )

    def forward(self, inputs: torch.Tensor) -> torch.Tensor:
        return self.network(inputs)


def parameter_count(model: nn.Module) -> int:
    return sum(parameter.numel() for parameter in model.parameters())