source/model.py

from __future__ import annotations

import math

import torch
from torch import nn


class CouplingLayer(nn.Module):
    def __init__(self, mask: tuple[float, float]) -> None:
        super().__init__()
        self.register_buffer("mask", torch.tensor(mask))
        self.network = nn.Sequential(
            nn.Linear(2, 48),
            nn.SiLU(),
            nn.Linear(48, 48),
            nn.SiLU(),
            nn.Linear(48, 4),
        )

    def parameters_for(self, masked: torch.Tensor) -> tuple[torch.Tensor, ...]:
        scale, translation = self.network(masked).chunk(2, dim=1)
        scale = 1.4 * torch.tanh(scale) * (1 - self.mask)
        translation = translation * (1 - self.mask)
        return scale, translation

    def forward(self, values: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
        masked = values * self.mask
        scale, translation = self.parameters_for(masked)
        transformed = masked + (1 - self.mask) * (
            values * torch.exp(scale) + translation
        )
        return transformed, scale.sum(1)

    def inverse(self, values: torch.Tensor) -> torch.Tensor:
        masked = values * self.mask
        scale, translation = self.parameters_for(masked)
        return masked + (1 - self.mask) * (
            (values - translation) * torch.exp(-scale)
        )


class RealNVP(nn.Module):
    def __init__(self, layers: int = 8) -> None:
        super().__init__()
        self.layers = nn.ModuleList(
            [
                CouplingLayer((1.0, 0.0) if index % 2 == 0 else (0.0, 1.0))
                for index in range(layers)
            ]
        )

    def forward(self, values: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
        log_determinant = values.new_zeros(len(values))
        latent = values
        for layer in self.layers:
            latent, change = layer(latent)
            log_determinant += change
        return latent, log_determinant

    def inverse(self, latent: torch.Tensor) -> torch.Tensor:
        values = latent
        for layer in reversed(self.layers):
            values = layer.inverse(values)
        return values

    def log_probability(self, values: torch.Tensor) -> torch.Tensor:
        latent, log_determinant = self(values)
        base = -0.5 * (latent**2).sum(1) - math.log(2 * math.pi)
        return base + log_determinant

    @torch.inference_mode()
    def sample(
        self,
        samples: int,
        *,
        seed: int,
        temperature: float = 1.0,
    ) -> torch.Tensor:
        generator = torch.Generator().manual_seed(seed)
        latent = temperature * torch.randn(samples, 2, generator=generator)
        return self.inverse(latent)


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