ble-ai-localizer/app/train_executor.py

import os
import json
import time
import joblib
import pandas as pd
import numpy as np
from pathlib import Path
from datetime import datetime
from sklearn.neighbors import KNeighborsClassifier, KNeighborsRegressor

# Import delle utility esistenti
from .logger_utils import log_msg as log
from .csv_config import load_gateway_features_csv

def process_train_jobs():
    """Monitora ed esegue i job di addestramento salvando backup cronologici."""
    JOBS_DIR = Path("/data/train/train_jobs")
    JOBS_DIR.mkdir(parents=True, exist_ok=True)
    
    job_files = list(JOBS_DIR.glob("*.lock"))
    if not job_files:
        return

    for job_path in job_files:
        try:
            log(f"[TRAIN-CORE] Rilevato nuovo job: {job_path.name}")
            
            with open(job_path, "r") as f:
                job = json.load(f)
            
            campagna = job["campaign"]
            knn_cfg = job["knn"]
            nan_fill = job["nan_fill"]
            gw_csv = job["gateways_csv"]
            
            # --- GENERAZIONE NOME FILE CON TIMESTAMP ---
            now_str = datetime.now().strftime("%Y%m%d_%H%M%S")
            model_filename = f"model_camp_{campagna}_{now_str}.joblib"
            model_path = Path("/data/model") / model_filename
            
            # Caricamento Gateway con normalizzazione MAC
            gws = load_gateway_features_csv(gw_csv)
            gateways_order = [g.mac.lower().strip() for g in gws]
            log(f"[TRAIN-CORE] Feature vector: {len(gateways_order)} gateway caricati da {gw_csv}")

            # Analisi file campioni
            samples_dir = Path("/data/train/samples")
            sample_files = list(samples_dir.glob(f"{campagna}_*.csv"))
            log(f"[TRAIN-CORE] Analisi di {len(sample_files)} file per campagna '{campagna}'")
            
            X_list, y_z, y_xy = [], [], []
            for fp in sample_files:
                try:
                    df = pd.read_csv(fp, sep=";")
                    if df.empty: continue
                    # Normalizziamo le colonne del DF in minuscolo
                    df.columns = [c.lower().strip() for c in df.columns]
                    row = df.iloc[0]
                    
                    features = []
                    for gw in gateways_order:
                        val = row.get(gw)
                        # Gestione esplicita di 'nan' stringa o NaN numerico
                        if val is not None and str(val).lower() != 'nan' and not pd.isna(val):
                            features.append(float(val))
                        else:
                            features.append(float(nan_fill))
                    
                    X_list.append(features)
                    y_z.append(int(round(float(row.get("z")))))
                    y_xy.append([float(row.get("x")), float(row.get("y"))])
                except Exception as e:
                    log(f"[TRAIN-CORE] Errore nel file {fp.name}: {e}")
                    continue

            if not X_list:
                log(f"[TRAIN-CORE] ❌ ABORTO: Nessun dato matchato tra gateway.csv e fingerprint!")
                job_path.unlink()
                continue

            X = np.array(X_list)
            # Calcolo copertura reale
            matches_per_point = [np.sum(np.array(vec) > nan_fill) for vec in X_list]
            avg_match = np.mean(matches_per_point)
            log(f"[TRAIN-CORE] Dataset pronto. Punti: {len(X_list)}. Media match Gateway: {avg_match:.2f}/{len(gateways_order)}")

            # --- LOG PARAMETRI ADDESTRAMENTO ---
            k_val = int(knn_cfg.get('k', 5))
            w_val = knn_cfg.get('weights', 'distance')
            m_val = knn_cfg.get('metric', 'euclidean')
            log(f"[TRAIN-CORE] Fitting KNN -> k: {k_val}, weights: {w_val}, metric: {m_val}, nan_fill: {nan_fill}")

            # Fitting
            floor_clf = KNeighborsClassifier(
                n_neighbors=k_val, weights=w_val, metric=m_val
            ).fit(X, np.array(y_z))
            
            models_xy = {}
            for z in np.unique(y_z):
                idx = np.where(np.array(y_z) == z)[0]
                # k_xy non può essere superiore al numero di campioni per piano
                current_k_xy = min(k_val, len(idx))
                models_xy[int(z)] = KNeighborsRegressor(
                    n_neighbors=current_k_xy, weights=w_val, metric=m_val
                ).fit(X[idx], np.array(y_xy)[idx])

            # Salvataggio
            model_pkg = {
                "floor_clf": floor_clf, "xy_by_floor": models_xy,
                "gateways_order": gateways_order, "nan_fill": nan_fill,
                "knn_params": {"k": k_val, "weights": w_val, "metric": m_val},
                "created_at": datetime.now().isoformat(), "campaign": campagna
            }
            
            Path("/data/model").mkdir(parents=True, exist_ok=True)
            joblib.dump(model_pkg, model_path)
            log(f"[TRAIN-CORE] ✅ Addestramento COMPLETATO: {model_filename}")

        except Exception as e:
            log(f"[TRAIN-CORE] ❌ ERRORE CRITICO: {str(e)}")
        finally:
            if job_path.exists(): job_path.unlink()

def run_train_monitor():
    while True:
        process_train_jobs()
        time.sleep(5)