model.py

"""
relgnn/model.py
Core RelGNN — Atenção sobre Rotas Atômicas (sem grafo estático).

Arquitetura:
  1. TableEncoder:      embeddings por tabela via MLP sobre features numéricas
  2. RouteAggregator:   attention ao longo de cada rota (sequência de tabelas)
  3. HierarchicalAgg:   agrega múltiplas rotas com pesos aprendidos
  4. FraudHead:         classificador binário final
"""

from dataclasses import dataclass
from typing import List, Dict, Tuple, Optional
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F

from data.routes import AtomicRoute


# ─── CONFIG ───────────────────────────────────────────────────────────────────

@dataclass
class RelGNNConfig:
    hidden_dim: int   = 64
    num_epochs: int   = 50
    learning_rate: float = 1e-3
    dropout: float    = 0.2
    num_heads: int    = 4
    seed: int         = 42


# ─── TABLE ENCODER ────────────────────────────────────────────────────────────

class TableEncoder(nn.Module):
    """
    Codifica as features de uma tabela em um embedding de tamanho `hidden_dim`.
    Opera direto nas colunas numéricas — sem conversão para grafo.
    """
    def __init__(self, input_dim: int, hidden_dim: int, dropout: float = 0.2):
        super().__init__()
        self.net = nn.Sequential(
            nn.Linear(input_dim, hidden_dim * 2),
            nn.LayerNorm(hidden_dim * 2),
            nn.ReLU(),
            nn.Dropout(dropout),
            nn.Linear(hidden_dim * 2, hidden_dim),
            nn.LayerNorm(hidden_dim),
            nn.ReLU(),
        )

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        return self.net(x)


# ─── ROUTE ATTENTION ──────────────────────────────────────────────────────────

class RouteAttention(nn.Module):
    """
    Mecanismo de atenção sobre uma Rota Atômica.
    Recebe sequência de embeddings [h1, h2, ..., hK] (K = n_hops + 1)
    e retorna um embedding agregado representando a rota.

    Implementa atenção scaled-dot-product entre os hops.
    """
    def __init__(self, hidden_dim: int, num_heads: int = 4, dropout: float = 0.2):
        super().__init__()
        self.attn = nn.MultiheadAttention(
            embed_dim=hidden_dim,
            num_heads=num_heads,
            dropout=dropout,
            batch_first=True,
        )
        self.norm = nn.LayerNorm(hidden_dim)
        self.mlp  = nn.Sequential(
            nn.Linear(hidden_dim, hidden_dim * 2),
            nn.ReLU(),
            nn.Dropout(dropout),
            nn.Linear(hidden_dim * 2, hidden_dim),
        )

    def forward(self, hop_embeddings: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
        """
        Args:
            hop_embeddings: [batch, n_hops, hidden_dim]
        Returns:
            route_emb: [batch, hidden_dim]  — representação da rota
            alpha:     [batch, n_hops]      — pesos de atenção por hop
        """
        # Self-attention entre os hops da rota
        attn_out, alpha = self.attn(hop_embeddings, hop_embeddings, hop_embeddings)

        # Residual + norm
        attn_out = self.norm(attn_out + hop_embeddings)

        # Agrega via mean-pooling ponderado (último hop = entidade alvo)
        # O primeiro token (tabela alvo) agrega informações dos vizinhos
        route_emb = attn_out[:, 0, :]   # [batch, hidden_dim]
        route_emb = route_emb + self.mlp(route_emb)

        alpha_weights = alpha.mean(dim=1)[:, 0, :]   # [batch, n_hops]
        return route_emb, alpha_weights


# ─── HIERARCHICAL ROUTE AGGREGATOR ───────────────────────────────────────────

class HierarchicalRouteAgg(nn.Module):
    """
    Agrega embeddings de múltiplas rotas com pesos aprendidos.
    Cada rota contribui de forma diferente para a predição final.
    """
    def __init__(self, hidden_dim: int, num_routes: int):
        super().__init__()
        self.route_weights = nn.Parameter(torch.ones(num_routes))
        self.output_proj   = nn.Linear(hidden_dim, hidden_dim)

    def forward(self, route_embeddings: List[torch.Tensor]) -> torch.Tensor:
        """
        Args:
            route_embeddings: lista de [batch, hidden_dim], uma por rota
        Returns:
            agg: [batch, hidden_dim]
        """
        stacked = torch.stack(route_embeddings, dim=1)   # [batch, R, hidden]
        weights = F.softmax(self.route_weights, dim=0)   # [R]
        weighted = (stacked * weights.unsqueeze(0).unsqueeze(-1)).sum(dim=1)
        return self.output_proj(weighted)


# ─── FRAUD HEAD ───────────────────────────────────────────────────────────────

class FraudHead(nn.Module):
    def __init__(self, hidden_dim: int, dropout: float = 0.2):
        super().__init__()
        self.net = nn.Sequential(
            nn.Linear(hidden_dim, hidden_dim // 2),
            nn.ReLU(),
            nn.Dropout(dropout),
            nn.Linear(hidden_dim // 2, 1),
        )

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        return self.net(x).squeeze(-1)   # [batch]


# ─── RELGNN ───────────────────────────────────────────────────────────────────

class RelGNN(nn.Module):
    """
    RelGNN completo.

    Fluxo:
      tabelas SQL
        → TableEncoder (por tabela)
        → RouteAttention (por rota atômica)
        → HierarchicalRouteAgg
        → FraudHead
        → sigmoid(logit) = P(fraude)
    """

    def __init__(self, config: RelGNNConfig):
        super().__init__()
        self.config = config
        torch.manual_seed(config.seed)

    def build(self, feature_dims: Dict[str, int], routes: List[AtomicRoute]):
        """Instancia os módulos após conhecer as dimensões das features."""
        H = self.config.hidden_dim
        D = self.config.dropout

        self.table_encoders = nn.ModuleDict({
            table: TableEncoder(dim, H, D)
            for table, dim in feature_dims.items()
        })

        self.route_attns = nn.ModuleList([
            RouteAttention(H, self.config.num_heads, D)
            for _ in routes
        ])

        self.hierarchical = HierarchicalRouteAgg(H, len(routes))
        self.fraud_head    = FraudHead(H, D)
        self.routes        = routes

    def forward(
        self,
        table_features: Dict[str, torch.Tensor],
    ) -> Tuple[torch.Tensor, Dict]:
        """
        Args:
            table_features: {table_name: [batch, feature_dim]}
        Returns:
            logits: [batch]
            attention_info: dict com pesos de atenção por rota
        """
        # 1. Encoder por tabela
        table_embs = {
            table: encoder(table_features[table])
            for table, encoder in self.table_encoders.items()
            if table in table_features
        }

        # 2. Attention por rota atômica
        route_embs = []
        attention_info = {}

        for i, (route, attn_module) in enumerate(zip(self.routes, self.route_attns)):
            # Coleta embeddings das tabelas na rota
            available = [t for t in route.path if t in table_embs]
            if len(available) < 2:
                # Usa embedding da tabela alvo repetido se rota incompleta
                e = table_embs.get(route.path[0], list(table_embs.values())[0])
                route_embs.append(e)
                continue

            hop_list = [table_embs[t] for t in available]
            hop_tensor = torch.stack(hop_list, dim=1)   # [batch, K, H]

            route_emb, alpha = attn_module(hop_tensor)
            route_embs.append(route_emb)
            attention_info[f"route_{i}"] = alpha.detach().cpu().numpy()

        # 3. Agrega rotas hierarquicamente
        agg = self.hierarchical(route_embs)

        # 4. Classificador de fraude
        logits = self.fraud_head(agg)

        return logits, attention_info

    def fit(self, tables, routes, log_fn=print, progress_fn=None):
        """Wrapper de treinamento completo."""
        from relgnn.trainer import Trainer
        trainer = Trainer(self, self.config)
        return trainer.fit(tables, routes, log_fn=log_fn, progress_fn=progress_fn)