train_utils.py

import os
import json
import time
from datetime import datetime
from typing import List, Tuple

import torch
import torch.nn as nn
import torch.optim as optim

from config import MODEL_DIR, META_DIR, DATASET_DISPLAY_NAME
from data_utils import make_loaders
from metrics_utils import compute_classification_metrics, save_confusion_matrix_figure
from model import SimpleCNN, ResNet18Classifier


def model_weight_path(model_name: str) -> str:
    return os.path.join(MODEL_DIR, f"{model_name}.pt")


def model_meta_path(model_name: str) -> str:
    return os.path.join(META_DIR, f"{model_name}.json")


def list_saved_models() -> List[str]:
    names = []
    for fn in os.listdir(META_DIR):
        if fn.endswith(".json"):
            names.append(fn[:-5])
    return sorted(names, reverse=True)


def get_runtime_device() -> torch.device:
    return torch.device("cuda" if torch.cuda.is_available() else "cpu")


def save_model(model: nn.Module, model_name: str, config: dict, training_summary: dict):
    cpu_state_dict = {k: v.detach().cpu() for k, v in model.state_dict().items()}
    torch.save(cpu_state_dict, model_weight_path(model_name))

    payload = {
        "model_name": model_name,
        "config": config,
        "training_summary": training_summary,
        "created_at": datetime.now().strftime("%Y-%m-%d %H:%M:%S"),
    }

    with open(model_meta_path(model_name), "w", encoding="utf-8") as f:
        json.dump(payload, f, indent=2, ensure_ascii=False)


def load_model(model_name: str, device: torch.device) -> Tuple[nn.Module, dict]:
    meta_file = model_meta_path(model_name)
    weight_file = model_weight_path(model_name)

    if not os.path.exists(meta_file):
        raise FileNotFoundError(f"Métadonnées introuvables pour le modèle : {model_name}")

    if not os.path.exists(weight_file):
        raise FileNotFoundError(f"Poids introuvables pour le modèle : {model_name}")

    with open(meta_file, "r", encoding="utf-8") as f:
        meta = json.load(f)

    cfg = meta["config"]

    if cfg.get("model_type", "cnn") == "resnet18":
        model = ResNet18Classifier(
            num_classes=cfg["num_classes"],
            dropout=cfg.get("dropout", 0.4),
            fc_dim=cfg.get("fc_dim", 256),
        )
    else:
        model = SimpleCNN(
            num_classes=cfg["num_classes"],
            num_conv_blocks=cfg.get("num_conv_blocks", 3),
            base_filters=cfg.get("base_filters", 32),
            kernel_size=cfg.get("kernel_size", 3),
            use_batchnorm=cfg.get("use_batchnorm", True),
            dropout=cfg.get("dropout", 0.4),
            fc_dim=cfg.get("fc_dim", 256),
        )

    state_dict = torch.load(weight_file, map_location="cpu")
    model.load_state_dict(state_dict)
    model.to(device)
    model.eval()

    return model, meta


def evaluate_loss_acc(model, loader, criterion, device):
    model.eval()

    total_loss = 0.0
    total = 0
    correct = 0

    with torch.no_grad():
        for images, labels in loader:
            images, labels = images.to(device), labels.to(device)

            outputs = model(images)
            loss = criterion(outputs, labels)

            total_loss += loss.item() * images.size(0)
            preds = outputs.argmax(dim=1)

            correct += (preds == labels).sum().item()
            total += labels.size(0)

    avg_loss = total_loss / total if total else 0.0
    acc = correct / total if total else 0.0

    return avg_loss, acc


def collect_predictions(model, loader, device):
    model.eval()

    y_true = []
    y_pred = []

    with torch.no_grad():
        for images, labels in loader:
            images = images.to(device)

            outputs = model(images)
            preds = outputs.argmax(dim=1).detach().cpu().tolist()

            y_pred.extend(preds)
            y_true.extend(labels.tolist())

    return y_true, y_pred


def train_model(
    model_type: str = "cnn",
    num_conv_blocks: int = 3,
    base_filters: int = 32,
    kernel_size: int = 3,
    use_batchnorm: bool = True,
    dropout: float = 0.4,
    fc_dim: int = 256,
    learning_rate: float = 0.001,
    weight_decay: float = 0.0001,
    batch_size: int = 16,
    epochs: int = 30,
    model_tag: str = "",
):
    device = get_runtime_device()

    train_loader, val_loader, test_loader, class_names = make_loaders(batch_size)
    num_classes = len(class_names)

    if model_type == "resnet18":
        model = ResNet18Classifier(
            num_classes=num_classes,
            dropout=dropout,
            fc_dim=fc_dim,
        ).to(device)
    else:
        model = SimpleCNN(
            num_classes=num_classes,
            num_conv_blocks=num_conv_blocks,
            base_filters=base_filters,
            kernel_size=kernel_size,
            use_batchnorm=use_batchnorm,
            dropout=dropout,
            fc_dim=fc_dim,
        ).to(device)

    trainable_params = sum(p.numel() for p in model.parameters() if p.requires_grad)
    total_params = sum(p.numel() for p in model.parameters())

    criterion = nn.CrossEntropyLoss()

    optimizer = optim.AdamW(
        filter(lambda p: p.requires_grad, model.parameters()),
        lr=learning_rate,
        weight_decay=weight_decay,
    )

    # Réduit le LR de moitié si val_loss ne s'améliore pas pendant 8 époques
    # patience élevée car le val set est très petit (bruit important)
    scheduler = optim.lr_scheduler.ReduceLROnPlateau(
        optimizer,
        mode="min",
        factor=0.5,
        patience=8,
        min_lr=learning_rate * 0.2,
    )

    history = []
    logs = []
    start_time = time.time()

    best_val_loss = float("inf")
    best_state_dict = None

    for epoch in range(1, epochs + 1):
        model.train()

        running_loss = 0.0
        total = 0
        correct = 0

        for images, labels in train_loader:
            images, labels = images.to(device), labels.to(device)

            optimizer.zero_grad()
            outputs = model(images)

            loss = criterion(outputs, labels)
            loss.backward()

            # Important: prevents unstable fine-tuning / exploding gradients
            torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=1.0)

            optimizer.step()

            running_loss += loss.item() * images.size(0)

            preds = outputs.argmax(dim=1)
            correct += (preds == labels).sum().item()
            total += labels.size(0)

        train_loss = running_loss / total if total else 0.0
        train_acc = correct / total if total else 0.0

        val_loss, val_acc = evaluate_loss_acc(model, val_loader, criterion, device)
        scheduler.step(val_loss)
        current_lr = optimizer.param_groups[0]["lr"]

        if val_loss < best_val_loss:
            best_val_loss = val_loss
            best_state_dict = {
                k: v.detach().cpu().clone()
                for k, v in model.state_dict().items()
            }

        row = {
            "epoch": epoch,
            "train_loss": round(train_loss, 4),
            "train_acc": round(train_acc, 4),
            "val_loss": round(val_loss, 4),
            "val_acc": round(val_acc, 4),
        }

        history.append(row)

        logs.append(
            f"Époque {epoch}/{epochs} | "
            f"perte entraînement={train_loss:.4f}, précision entraînement={train_acc:.4f}, "
            f"perte validation={val_loss:.4f}, précision validation={val_acc:.4f}, "
            f"lr={current_lr:.6f}"
        )

    if best_state_dict is not None:
        model.load_state_dict(best_state_dict)

    test_loss, test_acc = evaluate_loss_acc(model, test_loader, criterion, device)
    y_true, y_pred = collect_predictions(model, test_loader, device)

    metrics = compute_classification_metrics(y_true, y_pred, class_names)

    elapsed = time.time() - start_time

    timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
    safe_tag = model_tag.strip().replace(" ", "_") if model_tag.strip() else "charcoal_resnet18"
    model_name = f"{safe_tag}_{timestamp}"

    cm_path = save_confusion_matrix_figure(metrics["confusion_matrix"], model_name)

    if model_type == "resnet18":
        architecture = "ResNet18 pré-entraîné (layer4 + classifieur)"
    else:
        architecture = f"CNN simple ({num_conv_blocks} blocs, filtres={base_filters}, noyau={kernel_size}x{kernel_size})"

    config = {
        "dataset_name": DATASET_DISPLAY_NAME,
        "model_type": model_type,
        "architecture": architecture,
        "num_classes": num_classes,
        "class_names": class_names,
        "num_conv_blocks": num_conv_blocks,
        "base_filters": base_filters,
        "kernel_size": kernel_size,
        "use_batchnorm": use_batchnorm,
        "dropout": dropout,
        "fc_dim": fc_dim,
        "learning_rate": learning_rate,
        "weight_decay": weight_decay,
        "batch_size": batch_size,
        "epochs": epochs,
    }

    training_summary = {
        "final_train_loss": history[-1]["train_loss"] if history else None,
        "final_train_acc": history[-1]["train_acc"] if history else None,
        "best_val_loss": round(best_val_loss, 4),
        "final_val_loss": history[-1]["val_loss"] if history else None,
        "final_val_acc": history[-1]["val_acc"] if history else None,
        "test_cross_entropy_loss": round(test_loss, 4),
        "test_accuracy": round(test_acc, 4),
        "test_f1_macro": metrics["f1_macro"],
        "test_f1_weighted": metrics["f1_weighted"],
        "elapsed_seconds": round(elapsed, 2),
        "device": str(device),
        "total_params": total_params,
        "trainable_params": trainable_params,
    }

    save_model(model, model_name, config, training_summary)

    logs.append("")
    logs.append("Entraînement terminé.")
    logs.append(f"Modèle sauvegardé : {model_name}")
    logs.append(f"Appareil utilisé : {device}")
    logs.append(f"Architecture : {architecture}")
    logs.append(f"Nombre total de paramètres : {total_params}")
    logs.append(f"Paramètres entraînables : {trainable_params}")
    logs.append(f"Perte test cross-entropy : {test_loss:.4f}")
    logs.append(f"Accuracy test : {test_acc:.4f}")
    logs.append(f"F1 macro test : {metrics['f1_macro']:.4f}")
    logs.append(f"F1 pondéré test : {metrics['f1_weighted']:.4f}")
    logs.append(f"Temps écoulé : {elapsed:.1f}s")

    return {
        "logs": "\n".join(logs),
        "history": history,
        "summary": training_summary,
        "model_name": model_name,
        "classification_report": metrics["classification_report"],
        "confusion_matrix": metrics["confusion_matrix"],
        "confusion_matrix_path": cm_path,
    }


def evaluate_saved_model(model_name: str):
    if not model_name:
        raise ValueError("Aucun modèle sélectionné.")

    device = get_runtime_device()
    model, meta = load_model(model_name, device)

    batch_size = int(meta["config"].get("batch_size", 16))
    _, _, test_loader, class_names = make_loaders(batch_size)

    criterion = nn.CrossEntropyLoss()

    test_loss, test_acc = evaluate_loss_acc(model, test_loader, criterion, device)
    y_true, y_pred = collect_predictions(model, test_loader, device)

    metrics = compute_classification_metrics(y_true, y_pred, class_names)
    cm_path = save_confusion_matrix_figure(metrics["confusion_matrix"], model_name)

    summary = {
        "test_cross_entropy_loss": round(test_loss, 4),
        "test_accuracy": round(test_acc, 4),
        "test_f1_macro": metrics["f1_macro"],
        "test_f1_weighted": metrics["f1_weighted"],
        "device": str(device),
    }

    return summary, metrics["classification_report"], metrics["confusion_matrix"], cm_path