from __future__ import annotations

import torch
from schema import ACTIONS
from torch import nn


class NeuralModelMachine(nn.Module):
    def __init__(self) -> None:
        super().__init__()
        self.action_embedding = nn.Embedding(len(ACTIONS), 16)
        self.network = nn.Sequential(
            nn.Linear(8 + 4 + 16, 64),
            nn.LayerNorm(64),
            nn.SiLU(),
            nn.Linear(64, 64),
            nn.SiLU(),
            nn.Linear(64, 9),
        )

    def forward(
        self,
        state: torch.Tensor,
        capabilities: torch.Tensor,
        action: torch.Tensor,
    ) -> torch.Tensor:
        action_features = self.action_embedding(action)
        return self.network(torch.cat([state, capabilities, action_features], dim=1))

    @torch.inference_mode()
    def transition(
        self,
        state: torch.Tensor,
        capabilities: torch.Tensor,
        action: torch.Tensor,
    ) -> tuple[torch.Tensor, torch.Tensor]:
        probabilities = torch.sigmoid(self(state, capabilities, action))
        return (probabilities[:, :8] >= 0.5).float(), probabilities[:, 8]


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