vendor/scripts/experiment_interaction_attack_pv_gnn.py

from __future__ import annotations

from dataclasses import dataclass
from io import BytesIO
from pathlib import Path
import base64
import html
import json
import random
import sys

import matplotlib.pyplot as plt
from matplotlib import colors
from matplotlib.patches import Rectangle
from mplsoccer import Pitch
import numpy as np
import pandas as pd
import torch
from torch import nn
from torch.utils.data import DataLoader, TensorDataset


SCRIPT_DIR = Path(__file__).resolve().parent
PROJECT_ROOT = SCRIPT_DIR.parent
MODEL_DIR = PROJECT_ROOT / "data" / "modeling"
REPORTS_DIR = PROJECT_ROOT / "reports"
REPORT_PATH = REPORTS_DIR / "attack_pv_interaction_gnn_report.html"
JSON_PATH = MODEL_DIR / "attack_pv_interaction_gnn_metrics.json"
MODEL_PATH = MODEL_DIR / "attack_pv_interaction_gnn_bundle.pt"
ATTACK_PRED_PATH = MODEL_DIR / "attack_interaction_gnn_test_predictions.parquet"
PV_PRED_PATH = MODEL_DIR / "pv_interaction_gnn_test_predictions.parquet"

if str(SCRIPT_DIR) not in sys.path:
    sys.path.insert(0, str(SCRIPT_DIR))

import train_attack_prediction_ffn as base  # noqa: E402


RANDOM_SEED = 42
BATCH_SIZE = 256
MAX_EPOCHS = 260
PATIENCE = 32
LEARNING_RATE = 4e-4
WEIGHT_DECAY = 1e-5
SMOOTH_LAMBDA = 0.005
GATE_ENTROPY_LAMBDA = 0.001


ZONE_ORDER = [
    "Deep_Cross__Der_",
    "Half_Space__Der_",
    "Creativity_Zone",
    "Half_Space__Izq_",
    "Deep_Cross__Izq_",
    "Cross__Der_",
    "Cut_Back__Der_",
    "Scoring_Zone",
    "Cut_Back__Izq_",
    "Cross__Izq_",
]

PRETTY_ZONE = {
    "Scoring_Zone": "Scoring Zone",
    "Creativity_Zone": "Creativity Zone",
    "Half_Space__Izq_": "Half-Space Izq",
    "Half_Space__Der_": "Half-Space Der",
    "Cut_Back__Izq_": "Cut-Back Izq",
    "Cut_Back__Der_": "Cut-Back Der",
    "Cross__Izq_": "Cross Izq",
    "Cross__Der_": "Cross Der",
    "Deep_Cross__Izq_": "Deep Cross Izq",
    "Deep_Cross__Der_": "Deep Cross Der",
}

ATTACK_PREV_JSON = MODEL_DIR / "attack_distribution_gnn_metrics.json"
PV_PREV_JSON = MODEL_DIR / "pv_distribution_gnn_metrics.json"


@dataclass
class TaskData:
    df: pd.DataFrame
    global_x: np.ndarray
    attack_node_x: np.ndarray
    defense_node_x: np.ndarray
    y_dist: np.ndarray
    baseline_long: np.ndarray
    baseline_short: np.ndarray
    train_idx: np.ndarray
    val_idx: np.ndarray
    test_idx: np.ndarray
    val_start_date: str
    task_name: str


@dataclass
class SplitData:
    global_x: np.ndarray
    attack_node_x: np.ndarray
    defense_node_x: np.ndarray
    y_dist: np.ndarray
    baseline_long: np.ndarray
    baseline_short: np.ndarray
    metadata: pd.DataFrame


def _set_seed(seed: int = RANDOM_SEED) -> None:
    random.seed(seed)
    np.random.seed(seed)
    torch.manual_seed(seed)


def _attack_zone_rectangles() -> dict[str, list[tuple[float, float, float, float]]]:
    zones: dict[str, list[tuple[float, float, float, float]]] = {}

    def add(z: str, x0: float, x1: float, y0: float, y1: float) -> None:
        zones.setdefault(z, []).append((x0, y0, x1 - x0, y1 - y0))

    add("Scoring_Zone", 83, 100, 37, 63)
    add("Cut_Back__Izq_", 83, 100, 63, 79)
    add("Cross__Izq_", 83, 100, 79, 100)
    add("Cut_Back__Der_", 83, 100, 21, 37)
    add("Cross__Der_", 83, 100, 0, 21)
    add("Creativity_Zone", 60, 83, 37, 63)
    add("Half_Space__Izq_", 60, 83, 63, 79)
    add("Deep_Cross__Izq_", 60, 83, 79, 100)
    add("Half_Space__Der_", 60, 83, 21, 37)
    add("Deep_Cross__Der_", 60, 83, 0, 21)
    return zones


ZONES_RECTS = _attack_zone_rectangles()


def _img_to_base64(fig: plt.Figure) -> str:
    buf = BytesIO()
    fig.savefig(buf, format="png", dpi=180, bbox_inches="tight", facecolor=fig.get_facecolor())
    plt.close(fig)
    return base64.b64encode(buf.getvalue()).decode("ascii")


def _normalized_graphs() -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
    neighbors = {
        "Deep_Cross__Der_": ["Half_Space__Der_", "Cross__Der_", "Deep_Cross__Izq_"],
        "Half_Space__Der_": ["Deep_Cross__Der_", "Creativity_Zone", "Cut_Back__Der_", "Half_Space__Izq_"],
        "Creativity_Zone": ["Half_Space__Der_", "Half_Space__Izq_", "Scoring_Zone", "Cut_Back__Der_", "Cut_Back__Izq_"],
        "Half_Space__Izq_": ["Creativity_Zone", "Deep_Cross__Izq_", "Cut_Back__Izq_", "Half_Space__Der_"],
        "Deep_Cross__Izq_": ["Half_Space__Izq_", "Cross__Izq_", "Deep_Cross__Der_"],
        "Cross__Der_": ["Deep_Cross__Der_", "Cut_Back__Der_", "Cross__Izq_"],
        "Cut_Back__Der_": ["Cross__Der_", "Scoring_Zone", "Half_Space__Der_", "Cut_Back__Izq_", "Creativity_Zone"],
        "Scoring_Zone": ["Cut_Back__Der_", "Cut_Back__Izq_", "Creativity_Zone"],
        "Cut_Back__Izq_": ["Cross__Izq_", "Scoring_Zone", "Half_Space__Izq_", "Cut_Back__Der_", "Creativity_Zone"],
        "Cross__Izq_": ["Deep_Cross__Izq_", "Cut_Back__Izq_", "Cross__Der_"],
    }
    n = len(ZONE_ORDER)
    zone_to_idx = {z: i for i, z in enumerate(ZONE_ORDER)}
    adj = np.zeros((n, n), dtype=np.float32)
    cross = np.eye(n, dtype=np.float32)
    edge_pairs: list[tuple[int, int]] = []
    for z, neighs in neighbors.items():
        i = zone_to_idx[z]
        for neigh in neighs:
            j = zone_to_idx[neigh]
            adj[i, j] = 1.0
            cross[i, j] = 1.0
            edge_pairs.append((i, j))
    deg = np.where(adj.sum(axis=1, keepdims=True) > 0, adj.sum(axis=1, keepdims=True), 1.0)
    adj = adj / deg
    cross_deg = np.where(cross.sum(axis=1, keepdims=True) > 0, cross.sum(axis=1, keepdims=True), 1.0)
    cross = cross / cross_deg
    return torch.tensor(adj, dtype=torch.float32), torch.tensor(cross, dtype=torch.float32), torch.tensor(edge_pairs, dtype=torch.long)


ADJ_MATRIX, CROSS_MATRIX, EDGE_INDEX = _normalized_graphs()


def _standardize_global(global_features: pd.DataFrame, train_idx: np.ndarray) -> tuple[np.ndarray, dict]:
    fill_values = global_features.iloc[train_idx].median(numeric_only=False)
    filled = global_features.fillna(fill_values)
    means = filled.iloc[train_idx].mean(axis=0)
    stds = filled.iloc[train_idx].std(axis=0, ddof=0).replace(0, 1.0)
    scaled = ((filled - means) / stds).to_numpy(dtype=np.float32)
    return scaled, {
        "fill_values": fill_values.to_dict(),
        "means": means.to_dict(),
        "stds": stds.to_dict(),
        "global_feature_columns": list(global_features.columns),
    }


def _standardize_nodes(node_tensor: np.ndarray, train_idx: np.ndarray) -> tuple[np.ndarray, dict]:
    train = node_tensor[train_idx]
    fill = np.nanmedian(train, axis=0)
    filled = np.where(np.isnan(node_tensor), fill[None, :, :], node_tensor)
    means = filled[train_idx].mean(axis=0)
    stds = filled[train_idx].std(axis=0, ddof=0)
    stds = np.where(stds > 0, stds, 1.0)
    scaled = (filled - means[None, :, :]) / stds[None, :, :]
    return scaled.astype(np.float32), {
        "fill_values": fill.tolist(),
        "means": means.tolist(),
        "stds": stds.tolist(),
    }


def _make_split(task: TaskData, idx: np.ndarray) -> SplitData:
    return SplitData(
        global_x=task.global_x[idx].astype(np.float32),
        attack_node_x=task.attack_node_x[idx].astype(np.float32),
        defense_node_x=task.defense_node_x[idx].astype(np.float32),
        y_dist=task.y_dist[idx].astype(np.float32),
        baseline_long=task.baseline_long[idx].astype(np.float32),
        baseline_short=task.baseline_short[idx].astype(np.float32),
        metadata=task.df.iloc[idx].copy().reset_index(drop=True),
    )


def _make_loader(split: SplitData, shuffle: bool) -> DataLoader:
    dataset = TensorDataset(
        torch.from_numpy(split.global_x),
        torch.from_numpy(split.attack_node_x),
        torch.from_numpy(split.defense_node_x),
        torch.from_numpy(split.y_dist),
        torch.from_numpy(split.baseline_long),
        torch.from_numpy(split.baseline_short),
    )
    return DataLoader(dataset, batch_size=BATCH_SIZE, shuffle=shuffle)


class IntraGraphBlock(nn.Module):
    def __init__(self, in_dim: int, out_dim: int, dropout: float) -> None:
        super().__init__()
        self.self_lin = nn.Linear(in_dim, out_dim)
        self.neigh_lin = nn.Linear(in_dim, out_dim)
        self.norm = nn.LayerNorm(out_dim)
        self.dropout = nn.Dropout(dropout)

    def forward(self, x: torch.Tensor, adj: torch.Tensor) -> torch.Tensor:
        neigh = torch.einsum("ij,bjf->bif", adj, x)
        h = self.self_lin(x) + self.neigh_lin(neigh)
        h = self.norm(h)
        h = torch.relu(h)
        return self.dropout(h)


class CrossGraphBlock(nn.Module):
    def __init__(self, src_dim: int, dst_dim: int, out_dim: int, dropout: float) -> None:
        super().__init__()
        self.dst_lin = nn.Linear(dst_dim, out_dim)
        self.src_lin = nn.Linear(src_dim, out_dim)
        self.norm = nn.LayerNorm(out_dim)
        self.dropout = nn.Dropout(dropout)

    def forward(self, dst: torch.Tensor, src: torch.Tensor, cross_adj: torch.Tensor) -> torch.Tensor:
        src_msg = torch.einsum("ij,bjf->bif", cross_adj, src)
        h = self.dst_lin(dst) + self.src_lin(src_msg)
        h = self.norm(h)
        h = torch.relu(h)
        return self.dropout(h)


class InteractionDistributionGNN(nn.Module):
    def __init__(self, attack_node_dim: int, defense_node_dim: int, global_dim: int, hidden_dim: int = 96, global_hidden: int = 96, num_layers: int = 3) -> None:
        super().__init__()
        self.global_encoder = nn.Sequential(
            nn.Linear(global_dim, 192),
            nn.ReLU(),
            nn.Dropout(0.10),
            nn.Linear(192, global_hidden),
            nn.ReLU(),
        )
        self.attack_encoder = nn.Sequential(nn.Linear(attack_node_dim + global_hidden, hidden_dim), nn.ReLU())
        self.defense_encoder = nn.Sequential(nn.Linear(defense_node_dim + global_hidden, hidden_dim), nn.ReLU())
        self.attack_blocks = nn.ModuleList([IntraGraphBlock(hidden_dim, hidden_dim, 0.06) for _ in range(num_layers)])
        self.defense_blocks = nn.ModuleList([IntraGraphBlock(hidden_dim, hidden_dim, 0.06) for _ in range(num_layers)])
        self.cross_to_attack = nn.ModuleList([CrossGraphBlock(hidden_dim, hidden_dim, hidden_dim, 0.05) for _ in range(num_layers)])
        self.cross_to_defense = nn.ModuleList([CrossGraphBlock(hidden_dim, hidden_dim, hidden_dim, 0.05) for _ in range(num_layers)])
        self.gate_head = nn.Linear(hidden_dim * 2, 1)
        self.delta_head = nn.Linear(hidden_dim * 2, 1)

    def forward(
        self,
        attack_node_x: torch.Tensor,
        defense_node_x: torch.Tensor,
        global_x: torch.Tensor,
        baseline_long: torch.Tensor,
        baseline_short: torch.Tensor,
        adj: torch.Tensor,
        cross: torch.Tensor,
    ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
        g = self.global_encoder(global_x)
        g_rep = g.unsqueeze(1).expand(-1, attack_node_x.size(1), -1)
        att = self.attack_encoder(torch.cat([attack_node_x, g_rep], dim=-1))
        dfn = self.defense_encoder(torch.cat([defense_node_x, g_rep], dim=-1))

        for att_block, dfn_block, cross_att, cross_dfn in zip(self.attack_blocks, self.defense_blocks, self.cross_to_attack, self.cross_to_defense):
            att = att + att_block(att, adj) + cross_att(att, dfn, cross)
            dfn = dfn + dfn_block(dfn, adj) + cross_dfn(dfn, att, cross)

        pair = torch.cat([att, dfn], dim=-1)
        gate = torch.sigmoid(self.gate_head(pair)).squeeze(-1)
        mixed = gate * baseline_short + (1.0 - gate) * baseline_long
        delta = self.delta_head(pair).squeeze(-1)
        logits = torch.log(torch.clamp(mixed, min=1e-6)) + delta
        pred = torch.softmax(logits, dim=1)
        return pred, delta, gate, mixed


def _distribution_loss(pred: torch.Tensor, target: torch.Tensor, delta: torch.Tensor, gate: torch.Tensor) -> tuple[torch.Tensor, dict[str, float]]:
    eps = 1e-8
    kl = torch.nn.functional.kl_div(torch.log(torch.clamp(pred, min=eps)), target, reduction="batchmean")
    smooth = torch.mean((delta[:, EDGE_INDEX[:, 0]] - delta[:, EDGE_INDEX[:, 1]]) ** 2)
    gate_entropy = -torch.mean(gate * torch.log(torch.clamp(gate, min=eps)) + (1 - gate) * torch.log(torch.clamp(1 - gate, min=eps)))
    loss = kl + (SMOOTH_LAMBDA * smooth) + (GATE_ENTROPY_LAMBDA * gate_entropy)
    return loss, {"kl": float(kl.item()), "smooth": float(smooth.item()), "gate_entropy": float(gate_entropy.item())}


def _train_model(task: TaskData) -> tuple[InteractionDistributionGNN, list[dict[str, float]]]:
    train = _make_split(task, task.train_idx)
    val = _make_split(task, task.val_idx)
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    model = InteractionDistributionGNN(
        attack_node_dim=train.attack_node_x.shape[2],
        defense_node_dim=train.defense_node_x.shape[2],
        global_dim=train.global_x.shape[1],
    ).to(device)
    optimizer = torch.optim.AdamW(model.parameters(), lr=LEARNING_RATE, weight_decay=WEIGHT_DECAY)
    train_loader = _make_loader(train, shuffle=True)
    val_loader = _make_loader(val, shuffle=False)
    best_state = None
    best_val = float("inf")
    patience_left = PATIENCE
    history: list[dict[str, float]] = []
    adj = ADJ_MATRIX.to(device)
    cross = CROSS_MATRIX.to(device)

    for epoch in range(1, MAX_EPOCHS + 1):
        model.train()
        running = 0.0
        n_batches = 0
        for global_x, attack_node_x, defense_node_x, y_dist, baseline_long, baseline_short in train_loader:
            global_x = global_x.to(device)
            attack_node_x = attack_node_x.to(device)
            defense_node_x = defense_node_x.to(device)
            y_dist = y_dist.to(device)
            baseline_long = baseline_long.to(device)
            baseline_short = baseline_short.to(device)
            optimizer.zero_grad()
            pred, delta, gate, mixed = model(attack_node_x, defense_node_x, global_x, baseline_long, baseline_short, adj, cross)
            loss, parts = _distribution_loss(pred, y_dist, delta, gate)
            loss.backward()
            optimizer.step()
            running += float(loss.item())
            n_batches += 1

        val_metrics = _evaluate_loader(model, val_loader, device)
        history.append({"epoch": epoch, "train_loss": running / max(n_batches, 1), "val_loss": val_metrics["loss"], "val_kl": val_metrics["kl"]})
        if val_metrics["loss"] < best_val - 1e-6:
            best_val = val_metrics["loss"]
            best_state = {k: v.detach().cpu().clone() for k, v in model.state_dict().items()}
            patience_left = PATIENCE
        else:
            patience_left -= 1
        if patience_left <= 0:
            break

    if best_state is None:
        best_state = {k: v.detach().cpu().clone() for k, v in model.state_dict().items()}
    model.load_state_dict(best_state)
    return model, history


def _evaluate_loader(model: InteractionDistributionGNN, loader: DataLoader, device: torch.device) -> dict[str, float]:
    model.eval()
    total = 0.0
    total_kl = 0.0
    n_batches = 0
    adj = ADJ_MATRIX.to(device)
    cross = CROSS_MATRIX.to(device)
    with torch.no_grad():
        for global_x, attack_node_x, defense_node_x, y_dist, baseline_long, baseline_short in loader:
            global_x = global_x.to(device)
            attack_node_x = attack_node_x.to(device)
            defense_node_x = defense_node_x.to(device)
            y_dist = y_dist.to(device)
            baseline_long = baseline_long.to(device)
            baseline_short = baseline_short.to(device)
            pred, delta, gate, mixed = model(attack_node_x, defense_node_x, global_x, baseline_long, baseline_short, adj, cross)
            loss, parts = _distribution_loss(pred, y_dist, delta, gate)
            total += float(loss.item())
            total_kl += parts["kl"]
            n_batches += 1
    return {"loss": total / max(n_batches, 1), "kl": total_kl / max(n_batches, 1)}


def _predict(model: InteractionDistributionGNN, split: SplitData) -> tuple[np.ndarray, np.ndarray, np.ndarray]:
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    model.eval()
    preds: list[np.ndarray] = []
    gates: list[np.ndarray] = []
    mixeds: list[np.ndarray] = []
    loader = _make_loader(split, shuffle=False)
    adj = ADJ_MATRIX.to(device)
    cross = CROSS_MATRIX.to(device)
    with torch.no_grad():
        for global_x, attack_node_x, defense_node_x, y_dist, baseline_long, baseline_short in loader:
            global_x = global_x.to(device)
            attack_node_x = attack_node_x.to(device)
            defense_node_x = defense_node_x.to(device)
            baseline_long = baseline_long.to(device)
            baseline_short = baseline_short.to(device)
            pred, delta, gate, mixed = model(attack_node_x, defense_node_x, global_x, baseline_long, baseline_short, adj, cross)
            preds.append(pred.cpu().numpy())
            gates.append(gate.cpu().numpy())
            mixeds.append(mixed.cpu().numpy())
    return np.vstack(preds), np.vstack(gates), np.vstack(mixeds)


def _metrics(y_true: np.ndarray, pred_model: np.ndarray, pred_long: np.ndarray, pred_short: np.ndarray) -> dict[str, float]:
    def kl_proxy(y, p):
        return float(np.mean(np.sum(y * (np.log(np.clip(y, 1e-8, 1.0)) - np.log(np.clip(p, 1e-8, 1.0))), axis=1)))

    return {
        "model_mae": base._mean_abs_error(y_true, pred_model),
        "season_baseline_mae": base._mean_abs_error(y_true, pred_long),
        "short8_baseline_mae": base._mean_abs_error(y_true, pred_short),
        "model_jsd": base._jsd_mean(y_true, pred_model),
        "season_baseline_jsd": base._jsd_mean(y_true, pred_long),
        "short8_baseline_jsd": base._jsd_mean(y_true, pred_short),
        "model_kl_proxy": kl_proxy(y_true, pred_model),
        "season_baseline_kl_proxy": kl_proxy(y_true, pred_long),
        "short8_baseline_kl_proxy": kl_proxy(y_true, pred_short),
    }


def _load_prev_metrics(path: Path) -> dict[str, float]:
    return json.loads(path.read_text(encoding="utf-8"))["test_metrics"]


def _history_plot(history: list[dict[str, float]], title: str) -> str:
    hist = pd.DataFrame(history)
    fig, ax = plt.subplots(figsize=(7.5, 4.2), facecolor="#F6F7F4")
    ax.plot(hist["epoch"], hist["train_loss"], label="Train", color="#2B7A5A", linewidth=2)
    ax.plot(hist["epoch"], hist["val_loss"], label="Validacion", color="#D1495B", linewidth=2)
    ax.set_xlabel("Epoch")
    ax.set_ylabel("Loss")
    ax.set_title(title, fontsize=13, fontweight="bold", color="#14342B")
    ax.grid(alpha=0.2)
    ax.legend(frameon=False)
    return _img_to_base64(fig)


def _draw_pitch_distribution(ax: plt.Axes, values: dict[str, float], title: str) -> None:
    pitch = Pitch(pitch_type="opta", pitch_length=100, pitch_width=100, line_color="#D9E0DA", linewidth=1.2)
    pitch.draw(ax=ax)
    ax.set_facecolor("#F6F7F4")
    vmax = max(values.values()) if values else 1.0
    norm = colors.Normalize(vmin=0.0, vmax=max(vmax, 1e-6))
    cmap = plt.cm.Greens
    for zone, rects in ZONES_RECTS.items():
        value = values.get(zone, 0.0)
        for x, y, w, h in rects:
            ax.add_patch(Rectangle((x, y), w, h, facecolor=cmap(norm(value)), edgecolor="#FFFFFF", linewidth=1.5, alpha=0.84, zorder=1))
            ax.text(x + w / 2, y + h / 2, f"{value * 100:.1f}%", ha="center", va="center", fontsize=8.5, fontweight="bold", color="#16352C", zorder=3)
    ax.set_title(title, fontsize=12, fontweight="bold", color="#14342B", pad=10)


def _task_match_quad(row: pd.Series, prefix_target: str) -> str:
    fig, axes = plt.subplots(1, 4, figsize=(16, 4.6), facecolor="#F6F7F4")
    fig.subplots_adjust(wspace=0.08)
    real = {zone: float(row[f"{prefix_target}__{zone}"]) for zone in ZONE_ORDER}
    model = {zone: float(row[f"pred_model__{prefix_target}__{zone}"]) for zone in ZONE_ORDER}
    season = {zone: float(row[f"pred_season__{prefix_target}__{zone}"]) for zone in ZONE_ORDER}
    short8 = {zone: float(row[f"pred_short8__{prefix_target}__{zone}"]) for zone in ZONE_ORDER}
    _draw_pitch_distribution(axes[0], real, "Real")
    _draw_pitch_distribution(axes[1], model, "Interaccion")
    _draw_pitch_distribution(axes[2], season, "Baseline temporada")
    _draw_pitch_distribution(axes[3], short8, "Baseline ultimos 8")
    fig.suptitle(
        f"{row['fecha'].strftime('%Y-%m-%d')} | {row.get('team_name', 'Equipo')} vs {row.get('opponent_name', 'Rival')}",
        fontsize=15,
        fontweight="bold",
        color="#14342B",
        y=1.02,
    )
    return _img_to_base64(fig)


def _render_compare_table(title: str, current: dict[str, float], previous: dict[str, float], season: dict[str, float], short8: dict[str, float]) -> str:
    return f"""
    <section class="card">
      <h3>{html.escape(title)}</h3>
      <table>
        <tr><th>Modelo</th><th>MAE</th><th>JSD</th><th>KL</th></tr>
        <tr><td>Interaccion GNN</td><td>{current['model_mae']:.4f}</td><td>{current['model_jsd']:.4f}</td><td>{current['model_kl_proxy']:.4f}</td></tr>
        <tr><td>GNN anterior</td><td>{previous['model_mae']:.4f}</td><td>{previous['model_jsd']:.4f}</td><td>{previous['model_kl_proxy']:.4f}</td></tr>
        <tr><td>Baseline temporada</td><td>{season['mae']:.4f}</td><td>{season['jsd']:.4f}</td><td>{season['kl']:.4f}</td></tr>
        <tr><td>Baseline ultimos 8</td><td>{short8['mae']:.4f}</td><td>{short8['jsd']:.4f}</td><td>{short8['kl']:.4f}</td></tr>
      </table>
    </section>
    """


def _build_report(summary: dict, attack_hist_img: str, pv_hist_img: str, attack_racing_sections: list[str], pv_racing_sections: list[str]) -> str:
    attack_test = summary["attack"]["test_metrics"]
    pv_test = summary["pv"]["test_metrics"]
    return f"""<!DOCTYPE html>
<html lang="es">
<head>
  <meta charset="utf-8" />
  <title>Interaccion ataque-defensa GNN</title>
  <style>
    body {{ margin: 0; background: #f1f4ef; color: #14342B; font-family: -apple-system, BlinkMacSystemFont, "Segoe UI", sans-serif; }}
    .wrap {{ max-width: 1400px; margin: 0 auto; padding: 30px 24px 48px; }}
    h1 {{ margin: 0 0 10px; font-size: 40px; }}
    .lead {{ margin: 0 0 22px; font-size: 18px; color: #35574D; }}
    .hero {{ display: grid; grid-template-columns: repeat(4, minmax(0, 1fr)); gap: 14px; margin-bottom: 20px; }}
    .stat, .card, .match {{ background: white; border-radius: 20px; padding: 18px 20px; box-shadow: 0 8px 24px rgba(12, 36, 28, 0.08); }}
    .stat h3 {{ margin: 0 0 8px; font-size: 12px; text-transform: uppercase; letter-spacing: .08em; color: #587468; }}
    .stat p {{ margin: 0; font-size: 28px; font-weight: 800; }}
    .grid2 {{ display: grid; grid-template-columns: repeat(2, minmax(0, 1fr)); gap: 16px; margin-bottom: 18px; }}
    table {{ width: 100%; border-collapse: collapse; font-size: 14px; }}
    th, td {{ padding: 10px 8px; border-bottom: 1px solid #E5ECE6; text-align: left; }}
    th {{ color: #587468; text-transform: uppercase; font-size: 12px; letter-spacing: .06em; }}
    img {{ width: 100%; border-radius: 16px; display: block; }}
    .card {{ margin-bottom: 18px; }}
    .match {{ margin-bottom: 18px; }}
    @media (max-width: 980px) {{ .hero, .grid2 {{ grid-template-columns: 1fr; }} }}
  </style>
</head>
<body>
  <div class="wrap">
    <h1>GNN de interaccion ataque-defensa</h1>
    <p class="lead">Cada zona tiene dos representaciones: una ofensiva propia y una defensiva del rival. Hay message passing dentro de cada grafo y tambien entre ambos grafos, para modelar explicitamente el matchup zona a zona. Se compara contra los dos baselines y contra la GNN anterior.</p>
    <section class="hero">
      <div class="stat"><h3>Attack Test MAE</h3><p>{attack_test['model_mae']:.4f}</p></div>
      <div class="stat"><h3>PV Test MAE</h3><p>{pv_test['model_mae']:.4f}</p></div>
      <div class="stat"><h3>Attack Mejora vs Prev</h3><p>{summary['attack']['previous_test_metrics']['model_mae'] - attack_test['model_mae']:+.4f}</p></div>
      <div class="stat"><h3>PV Mejora vs Prev</h3><p>{summary['pv']['previous_test_metrics']['model_mae'] - pv_test['model_mae']:+.4f}</p></div>
    </section>

    <div class="grid2">
      {_render_compare_table(
        "Ataque - test",
        summary["attack"]["test_metrics"],
        summary["attack"]["previous_test_metrics"],
        {"mae": summary["attack"]["test_metrics"]["season_baseline_mae"], "jsd": summary["attack"]["test_metrics"]["season_baseline_jsd"], "kl": summary["attack"]["test_metrics"]["season_baseline_kl_proxy"]},
        {"mae": summary["attack"]["test_metrics"]["short8_baseline_mae"], "jsd": summary["attack"]["test_metrics"]["short8_baseline_jsd"], "kl": summary["attack"]["test_metrics"]["short8_baseline_kl_proxy"]},
      )}
      {_render_compare_table(
        "PV - test",
        summary["pv"]["test_metrics"],
        summary["pv"]["previous_test_metrics"],
        {"mae": summary["pv"]["test_metrics"]["season_baseline_mae"], "jsd": summary["pv"]["test_metrics"]["season_baseline_jsd"], "kl": summary["pv"]["test_metrics"]["season_baseline_kl_proxy"]},
        {"mae": summary["pv"]["test_metrics"]["short8_baseline_mae"], "jsd": summary["pv"]["test_metrics"]["short8_baseline_jsd"], "kl": summary["pv"]["test_metrics"]["short8_baseline_kl_proxy"]},
      )}
    </div>

    <div class="grid2">
      <section class="card">
        <h3>Ataque - entrenamiento</h3>
        <img src="data:image/png;base64,{attack_hist_img}" alt="Historial ataque" />
      </section>
      <section class="card">
        <h3>PV - entrenamiento</h3>
        <img src="data:image/png;base64,{pv_hist_img}" alt="Historial PV" />
      </section>
    </div>

    <section>
      <h2>Racing - Ataque</h2>
      {''.join(attack_racing_sections)}
    </section>
    <section>
      <h2>Racing - PV</h2>
      {''.join(pv_racing_sections)}
    </section>
  </div>
</body>
</html>"""


def _build_attack_task(df_base: pd.DataFrame) -> TaskData:
    df = df_base[df_base["usable_for_model"]].copy().reset_index(drop=True)
    attack_targets = [f"target_attack_share__{zone}" for zone in ZONE_ORDER]
    y_dist = df[attack_targets].to_numpy(dtype=np.float32)
    baseline_long = base._normalize_rows(df[[f"long_mean__actual_attack_share__{zone}" for zone in ZONE_ORDER]].to_numpy(dtype=float)).astype(np.float32)
    baseline_short = base._normalize_rows(df[[f"short_mean__actual_attack_share__{zone}" for zone in ZONE_ORDER]].to_numpy(dtype=float)).astype(np.float32)

    attack_node_cols = [
        [f"short_mean__actual_attack_share__{zone}", f"long_mean__actual_attack_share__{zone}", f"short_mean__actual_pv_share__{zone}", f"long_mean__actual_pv_share__{zone}"]
        for zone in ZONE_ORDER
    ]
    defense_node_cols = [
        [f"opp__short_mean__actual_conceded_attack_share__{zone}", f"opp__long_mean__actual_conceded_attack_share__{zone}", f"opp__short_mean__actual_conceded_pv_share__{zone}", f"opp__long_mean__actual_conceded_pv_share__{zone}"]
        for zone in ZONE_ORDER
    ]
    attack_node_x = np.stack([df[cols].apply(pd.to_numeric, errors="coerce").to_numpy(dtype=float) for cols in attack_node_cols], axis=1)
    defense_node_x = np.stack([df[cols].apply(pd.to_numeric, errors="coerce").to_numpy(dtype=float) for cols in defense_node_cols], axis=1)

    wide_features, _, _, _, _ = base._feature_matrix(df)
    node_cols = {c for cols in attack_node_cols + defense_node_cols for c in cols} | set(attack_targets)
    global_features = wide_features[[c for c in wide_features.columns if c not in node_cols]].copy()

    train_idx_pd, val_idx_pd, val_start_date = base._build_temporal_validation(df[df["split"] == "train"].copy())
    train_idx = train_idx_pd.to_numpy()
    val_idx = val_idx_pd.to_numpy()
    test_idx = df.index[df["split"] == "test"].to_numpy()
    global_x, _ = _standardize_global(global_features, train_idx)
    attack_node_x, _ = _standardize_nodes(attack_node_x, train_idx)
    defense_node_x, _ = _standardize_nodes(defense_node_x, train_idx)
    return TaskData(df, global_x, attack_node_x, defense_node_x, y_dist, baseline_long, baseline_short, train_idx, val_idx, test_idx, val_start_date, "attack")


def _clip_and_normalize(mat: np.ndarray) -> np.ndarray:
    clipped = np.clip(mat, 0.0, None)
    total = clipped.sum(axis=1, keepdims=True)
    out = np.zeros_like(clipped, dtype=np.float32)
    mask = total.squeeze(-1) > 0
    if mask.any():
        out[mask] = (clipped[mask] / total[mask]).astype(np.float32)
    return out


def _build_pv_task(df_base: pd.DataFrame) -> TaskData:
    df = df_base.copy()
    zone_cols = [f"zone_pvAdded__{zone}" for zone in ZONE_ORDER]
    df["pv_positive_total"] = np.clip(df[zone_cols].to_numpy(dtype=float), 0, None).sum(axis=1)
    df = df[df["usable_for_model"] & (df["pv_positive_total"] > 0)].copy().reset_index(drop=True)

    y_dist = _clip_and_normalize(df[[f"zone_pvAdded__{zone}" for zone in ZONE_ORDER]].to_numpy(dtype=float))
    baseline_long = _clip_and_normalize(df[[f"long_mean__zone_pvAdded__{zone}" for zone in ZONE_ORDER]].to_numpy(dtype=float))
    baseline_short = _clip_and_normalize(df[[f"short_mean__zone_pvAdded__{zone}" for zone in ZONE_ORDER]].to_numpy(dtype=float))

    attack_node_cols = [
        [f"short_mean__actual_pv_share__{zone}", f"long_mean__actual_pv_share__{zone}", f"short_mean__actual_attack_share__{zone}", f"long_mean__actual_attack_share__{zone}"]
        for zone in ZONE_ORDER
    ]
    defense_node_cols = [
        [f"opp__short_mean__actual_conceded_pv_share__{zone}", f"opp__long_mean__actual_conceded_pv_share__{zone}", f"opp__short_mean__actual_conceded_attack_share__{zone}", f"opp__long_mean__actual_conceded_attack_share__{zone}"]
        for zone in ZONE_ORDER
    ]
    attack_node_x = np.stack([df[cols].apply(pd.to_numeric, errors="coerce").to_numpy(dtype=float) for cols in attack_node_cols], axis=1)
    defense_node_x = np.stack([df[cols].apply(pd.to_numeric, errors="coerce").to_numpy(dtype=float) for cols in defense_node_cols], axis=1)

    wide_features, _, _, _, _ = base._feature_matrix(df)
    node_cols = {c for cols in attack_node_cols + defense_node_cols for c in cols}
    global_features = wide_features[[c for c in wide_features.columns if c not in node_cols]].copy()

    train_idx_pd, val_idx_pd, val_start_date = base._build_temporal_validation(df[df["split"] == "train"].copy())
    train_idx = train_idx_pd.to_numpy()
    val_idx = val_idx_pd.to_numpy()
    test_idx = df.index[df["split"] == "test"].to_numpy()
    global_x, _ = _standardize_global(global_features, train_idx)
    attack_node_x, _ = _standardize_nodes(attack_node_x, train_idx)
    defense_node_x, _ = _standardize_nodes(defense_node_x, train_idx)
    return TaskData(df, global_x, attack_node_x, defense_node_x, y_dist, baseline_long, baseline_short, train_idx, val_idx, test_idx, val_start_date, "pv")


def _run_task(task: TaskData) -> dict:
    model, history = _train_model(task)
    test_split = _make_split(task, task.test_idx)
    val_split = _make_split(task, task.val_idx)
    test_pred, test_gates, test_mixed = _predict(model, test_split)
    val_pred, val_gates, val_mixed = _predict(model, val_split)
    test_metrics = _metrics(test_split.y_dist.astype(float), test_pred, test_split.baseline_long.astype(float), test_split.baseline_short.astype(float))
    val_metrics = _metrics(val_split.y_dist.astype(float), val_pred, val_split.baseline_long.astype(float), val_split.baseline_short.astype(float))
    return {
        "model_state_dict": {k: v.detach().cpu() for k, v in model.state_dict().items()},
        "history": history,
        "test_split": test_split,
        "val_split": val_split,
        "test_pred": test_pred,
        "val_pred": val_pred,
        "test_gates": test_gates,
        "test_mixed": test_mixed,
        "test_metrics": test_metrics,
        "val_metrics": val_metrics,
    }


def _attach_predictions(split: SplitData, pred: np.ndarray, gates: np.ndarray, mixed: np.ndarray, target_prefix: str) -> pd.DataFrame:
    out = split.metadata.copy().reset_index(drop=True)
    for i, zone in enumerate(ZONE_ORDER):
        out[f"{target_prefix}__{zone}"] = split.y_dist[:, i]
        out[f"pred_model__{target_prefix}__{zone}"] = pred[:, i]
        out[f"pred_season__{target_prefix}__{zone}"] = split.baseline_long[:, i]
        out[f"pred_short8__{target_prefix}__{zone}"] = split.baseline_short[:, i]
        out[f"gate__{zone}"] = gates[:, i]
        out[f"mixed_base__{zone}"] = mixed[:, i]
    return out


def _racing_sections(df: pd.DataFrame, target_prefix: str) -> list[str]:
    racing = df[df["teamId"] == base.RACING_TEAM_ID].sort_values(["fecha", "matchId"]).tail(3)
    sections = []
    for _, row in racing.iterrows():
        real = row[[f"{target_prefix}__{zone}" for zone in ZONE_ORDER]].to_numpy(dtype=float)
        model = row[[f"pred_model__{target_prefix}__{zone}" for zone in ZONE_ORDER]].to_numpy(dtype=float)
        season = row[[f"pred_season__{target_prefix}__{zone}" for zone in ZONE_ORDER]].to_numpy(dtype=float)
        short8 = row[[f"pred_short8__{target_prefix}__{zone}" for zone in ZONE_ORDER]].to_numpy(dtype=float)
        sections.append(
            f"""
            <section class="match">
              <h3>{html.escape(row['fecha'].strftime('%Y-%m-%d'))} | {html.escape(str(row.get('team_name', 'Equipo')))} vs {html.escape(str(row.get('opponent_name', 'Rival')))}</h3>
              <p>MAE interaccion {base._mean_abs_error(real[None, :], model[None, :]):.4f} | baseline temporada {base._mean_abs_error(real[None, :], season[None, :]):.4f} | baseline ultimos 8 {base._mean_abs_error(real[None, :], short8[None, :]):.4f}</p>
              <img src="data:image/png;base64,{_task_match_quad(row, target_prefix)}" alt="Distribuciones" />
            </section>
            """
        )
    return sections


def main() -> None:
    _set_seed()
    MODEL_DIR.mkdir(parents=True, exist_ok=True)
    REPORTS_DIR.mkdir(parents=True, exist_ok=True)

    df_base = base._load_dataset()
    attack_task = _build_attack_task(df_base)
    pv_task = _build_pv_task(df_base)

    attack_result = _run_task(attack_task)
    pv_result = _run_task(pv_task)

    attack_prev = _load_prev_metrics(ATTACK_PREV_JSON)
    pv_prev = _load_prev_metrics(PV_PREV_JSON)

    attack_pred_df = _attach_predictions(attack_result["test_split"], attack_result["test_pred"], attack_result["test_gates"], attack_result["test_mixed"], "target_attack_share")
    pv_pred_df = _attach_predictions(pv_result["test_split"], pv_result["test_pred"], pv_result["test_gates"], pv_result["test_mixed"], "target_pv_dist")

    attack_keep = ["matchId", "fecha", "league", "season", "teamId", "team_name", "opponent_name", "is_home", "goals_for", "goals_against", "n_prior_matches", "opp_n_prior_matches"]
    attack_keep += [f"target_attack_share__{zone}" for zone in ZONE_ORDER]
    attack_keep += [f"pred_model__target_attack_share__{zone}" for zone in ZONE_ORDER]
    attack_keep += [f"pred_season__target_attack_share__{zone}" for zone in ZONE_ORDER]
    attack_keep += [f"pred_short8__target_attack_share__{zone}" for zone in ZONE_ORDER]
    attack_keep += [f"gate__{zone}" for zone in ZONE_ORDER]
    attack_keep += [f"mixed_base__{zone}" for zone in ZONE_ORDER]
    attack_pred_df[attack_keep].to_parquet(ATTACK_PRED_PATH, index=False)

    pv_keep = ["matchId", "fecha", "league", "season", "teamId", "team_name", "opponent_name", "is_home", "goals_for", "goals_against", "n_prior_matches", "opp_n_prior_matches"]
    pv_keep += [f"target_pv_dist__{zone}" for zone in ZONE_ORDER]
    pv_keep += [f"pred_model__target_pv_dist__{zone}" for zone in ZONE_ORDER]
    pv_keep += [f"pred_season__target_pv_dist__{zone}" for zone in ZONE_ORDER]
    pv_keep += [f"pred_short8__target_pv_dist__{zone}" for zone in ZONE_ORDER]
    pv_keep += [f"gate__{zone}" for zone in ZONE_ORDER]
    pv_keep += [f"mixed_base__{zone}" for zone in ZONE_ORDER]
    pv_pred_df[pv_keep].to_parquet(PV_PRED_PATH, index=False)

    summary = {
        "attack": {
            "train_rows": int(len(attack_task.train_idx)),
            "val_rows": int(len(attack_task.val_idx)),
            "test_rows": int(len(attack_task.test_idx)),
            "val_start_date": attack_task.val_start_date,
            "test_metrics": attack_result["test_metrics"],
            "val_metrics": attack_result["val_metrics"],
            "previous_test_metrics": attack_prev,
            "mean_gate_test": {zone: float(attack_result["test_gates"][:, i].mean()) for i, zone in enumerate(ZONE_ORDER)},
        },
        "pv": {
            "train_rows": int(len(pv_task.train_idx)),
            "val_rows": int(len(pv_task.val_idx)),
            "test_rows": int(len(pv_task.test_idx)),
            "val_start_date": pv_task.val_start_date,
            "test_metrics": pv_result["test_metrics"],
            "val_metrics": pv_result["val_metrics"],
            "previous_test_metrics": pv_prev,
            "mean_gate_test": {zone: float(pv_result["test_gates"][:, i].mean()) for i, zone in enumerate(ZONE_ORDER)},
        },
    }

    torch.save(
        {
            "attack_model_state_dict": attack_result["model_state_dict"],
            "pv_model_state_dict": pv_result["model_state_dict"],
            "zone_order": ZONE_ORDER,
            "summary": summary,
        },
        MODEL_PATH,
    )
    JSON_PATH.write_text(json.dumps(summary, ensure_ascii=False, indent=2), encoding="utf-8")

    report_html = _build_report(
        summary,
        _history_plot(attack_result["history"], "Ataque - interaccion"),
        _history_plot(pv_result["history"], "PV - interaccion"),
        _racing_sections(attack_pred_df, "target_attack_share"),
        _racing_sections(pv_pred_df, "target_pv_dist"),
    )
    REPORT_PATH.write_text(report_html, encoding="utf-8")

    print(f"Modelo guardado en: {MODEL_PATH}")
    print(f"Predicciones ataque guardadas en: {ATTACK_PRED_PATH}")
    print(f"Predicciones PV guardadas en: {PV_PRED_PATH}")
    print(f"Metricas guardadas en: {JSON_PATH}")
    print(f"Reporte guardado en: {REPORT_PATH}")
    print(json.dumps(summary, ensure_ascii=False, indent=2))


if __name__ == "__main__":
    main()