Mobius Linear example

mobius_linear.py

import torch.nn
import geoopt
# package.nn.modules.py
def create_ball(ball=None, c=None):
    if ball is None:
        assert c is not None, "curvature of the ball should be explicitly specified"
        ball = geoopt.PoincareBall(c)
    elif not isinstance(ball, geoopt.PoincareBall):
        raise ValueError("ball should be an instance of PoncareMall")
    return ball

class MobiusLinear(torch.nn.Linear):
    def __init__(self, *args, nonlin=None, ball=None, c=1.0, **kwargs):
        super().__init__(*args, **kwargs)
        self.ball = create_ball(ball, c)
        if self.bias is not None:
            self.bias = geoopt.ManifoldParameter(self.bias, manifold=self.ball)
        self.nonlin = nonlin
        self.reset_parameters()

    def forward(self, input):
        return mobius_linear(
            input,
            weight=self.weight,
            bias=self.bias,
            nonlin=self.nonlin,
            ball=self.ball,
        )

    @torch.no_grad()
    def reset_parameters(self):
        torch.nn.init.eye_(self.weight)
        self.weight.add_(torch.rand_like(self.weight).mul_(1e-3))
        if self.bias is not None:
            self.bias.zero_()

# package.nn.functional.py
def mobius_linear(input, weight, bias=None, nonlin=None, *, ball: geoopt.PoincareBall):
    output = ball.mobius_matvec(weight, input)
    if bias is not None:
        output = ball.mobius_add(output, bias)
    if nonlin is not None:
        output = ball.mobius_fn_apply(nonlin, output)
    return output