{"metadata":{"kernelspec":{"language":"python","display_name":"Python 3","name":"python3"},"language_info":{"name":"python","version":"3.11.13","mimetype":"text/x-python","codemirror_mode":{"name":"ipython","version":3},"pygments_lexer":"ipython3","nbconvert_exporter":"python","file_extension":".py"},"kaggle":{"accelerator":"none","dataSources":[{"sourceId":101849,"databundleVersionId":13093295,"sourceType":"competition"},{"sourceId":9629432,"sourceType":"datasetVersion","datasetId":5846888},{"sourceId":13121320,"sourceType":"datasetVersion","datasetId":8304548},{"sourceId":13128023,"sourceType":"datasetVersion","datasetId":8311982},{"sourceId":564250,"sourceType":"modelInstanceVersion","modelInstanceId":426187,"modelId":443655}],"dockerImageVersionId":31089,"isInternetEnabled":false,"language":"python","sourceType":"notebook","isGpuEnabled":false}},"nbformat_minor":4,"nbformat":4,"cells":[{"cell_type":"code","source":"!pip install --no-index --find-links=/kaggle/input/ariel-2024-pqdm pqdm","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-09-22T15:19:17.161384Z","iopub.execute_input":"2025-09-22T15:19:17.161692Z","iopub.status.idle":"2025-09-22T15:19:23.140582Z","shell.execute_reply.started":"2025-09-22T15:19:17.161668Z","shell.execute_reply":"2025-09-22T15:19:23.139052Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import os\nimport time\nimport itertools\nimport multiprocessing as mp\n\nimport numpy as np\nimport pandas as pd\nimport pandas.api.types\n\nimport torch\nimport torch.nn as nn\nimport torch.nn.functional as F\nfrom torch.utils.data import DataLoader, TensorDataset, random_split\n\nimport matplotlib.pyplot as plt\n\nfrom tqdm import tqdm\nfrom pqdm.threads import pqdm\n\nfrom sklearn.model_selection import train_test_split\nfrom sklearn.preprocessing import StandardScaler\nfrom sklearn.metrics import mean_squared_error\n\nfrom astropy.stats import sigma_clip\nfrom scipy.signal import savgol_filter\nfrom scipy.optimize import minimize\nimport scipy.stats\nfrom sklearn.metrics import mean_absolute_error, mean_squared_error, r2_score\nfrom sklearn.preprocessing import StandardScaler\nimport pandas as pd\nimport numpy as np\nfrom sklearn.metrics import mean_absolute_error, mean_squared_error, r2_score\n\nROOT_PATH = \"/kaggle/input/ariel-data-challenge-2025\"\nMODE = \"test\"\n\n__t0 = time.perf_counter()\n\nclass Config:\n    # FEATURES = ['transit_depth', 'Rs', 'i']\n    FEATURES = ['transit_depth', 'Rs', 'i', 'P', 'planet_temp', 'sma']\n    # FEATURES = ['transit_depth', 'Rs', 'Ms', 'Ts', 'Mp', 'e', 'P', 'sma', 'i']\n    DATA_PATH = '/kaggle/input/ariel-data-challenge-2025'\n    DATASET = \"test\"\n    load_data = False  # Set to True to load from disk, False to generate\n    DEBUG = False\n    load_model = False\n\n    SCALE = 0.96\n    SIGMA = 0.00055\n    \n    CUT_INF = 39\n    CUT_SUP = 321\n    \n    SENSOR_CONFIG = {\n        \"AIRS-CH0\": {\n            \"raw_shape\": [11250, 32, 356],\n            \"calibrated_shape\": [1, 32, CUT_SUP - CUT_INF],\n            \"linear_corr_shape\": (6, 32, 356),\n            \"dt_pattern\": (0.1, 4.5), \n            \"binning\": 30\n        },\n        \"FGS1\": {\n            \"raw_shape\": [135000, 32, 32],\n            \"calibrated_shape\": [1, 32, 32],\n            \"linear_corr_shape\": (6, 32, 32),\n            \"dt_pattern\": (0.1, 0.1),\n            \"binning\": 30 * 12\n        }\n    }\n    \n    MODEL_PHASE_DETECTION_SLICE = slice(30, 140)\n    MODEL_OPTIMIZATION_DELTA = 11 # 9\n    MODEL_POLYNOMIAL_DEGREE = 3\n    \n    N_JOBS = 3\n\nclass ParticipantVisibleError(Exception):\n    pass\n\ndef score(\n    solution: pd.DataFrame,\n    submission: pd.DataFrame,\n    row_id_column_name: str,\n    naive_mean: float,\n    naive_sigma: float,\n    fsg_sigma_true: float = 1e-6,\n    airs_sigma_true: float = 1e-5,\n    fgs_weight: float = 1,\n) -> float:\n    \"\"\"\n    This is a Gaussian Log Likelihood based metric. For a submission, which contains the predicted mean (x_hat) and variance (x_hat_std),\n    we calculate the Gaussian Log-likelihood (GLL) value to the provided ground truth (x). We treat each pair of x_hat,\n    x_hat_std as a 1D gaussian, meaning there will be 283 1D gaussian distributions, hence 283 values for each test spectrum,\n    the GLL value for one spectrum is the sum of all of them.\n\n    Inputs:\n        - solution: Ground Truth spectra (from test set)\n            - shape: (nsamples, n_wavelengths)\n        - submission: Predicted spectra and errors (from participants)\n            - shape: (nsamples, n_wavelengths*2)\n        naive_mean: (float) mean from the train set.\n        naive_sigma: (float) standard deviation from the train set.\n        fsg_sigma_true: (float) standard deviation from the FSG1 instrument for the test set.\n        airs_sigma_true: (float) standard deviation from the AIRS instrument for the test set.\n        fgs_weight: (float) relative weight of the fgs channel\n    \"\"\"\n\n    del solution[row_id_column_name]\n    del submission[row_id_column_name]\n\n    if submission.min().min() < 0:\n        raise ParticipantVisibleError('Negative values in the submission')\n    for col in submission.columns:\n        if not pandas.api.types.is_numeric_dtype(submission[col]):\n            raise ParticipantVisibleError(f'Submission column {col} must be a number')\n\n    n_wavelengths = len(solution.columns)\n    if len(submission.columns) != n_wavelengths * 2:\n        raise ParticipantVisibleError('Wrong number of columns in the submission')\n\n    y_pred = submission.iloc[:, :n_wavelengths].values\n    # Set a non-zero minimum sigma pred to prevent division by zero errors.\n    sigma_pred = np.clip(submission.iloc[:, n_wavelengths:].values, a_min=10**-15, a_max=None)\n    sigma_true = np.append(\n        np.array(\n            [\n                fsg_sigma_true,\n            ]\n        ),\n        np.ones(n_wavelengths - 1) * airs_sigma_true,\n    )\n    y_true = solution.values\n\n    GLL_pred = scipy.stats.norm.logpdf(y_true, loc=y_pred, scale=sigma_pred)\n    GLL_true = scipy.stats.norm.logpdf(y_true, loc=y_true, scale=sigma_true * np.ones_like(y_true))\n    GLL_mean = scipy.stats.norm.logpdf(y_true, loc=naive_mean * np.ones_like(y_true), scale=naive_sigma * np.ones_like(y_true))\n\n    # normalise the score, right now it becomes a matrix instead of a scalar.\n    ind_scores = (GLL_pred - GLL_mean) / (GLL_true - GLL_mean)\n\n    weights = np.append(np.array([fgs_weight]), np.ones(len(solution.columns) - 1))\n    weights = weights * np.ones_like(ind_scores)\n    submit_score = np.average(ind_scores, weights=weights)\n    return float(np.clip(submit_score, 0.0, 1.0))\n\ndef _phase_detector_signal(signal, cfg):\n    sl = cfg.MODEL_PHASE_DETECTION_SLICE\n    min_idx = int(np.argmin(signal[sl])) + sl.start\n    s1 = signal[:min_idx]; s2 = signal[min_idx:]\n    if s1.size < 3 or s2.size < 3:\n        return 0, len(signal) - 1\n    g1 = np.gradient(s1); g1_max = np.max(g1) if np.size(g1) else 0.0\n    g2 = np.gradient(s2); g2_max = np.max(g2) if np.size(g2) else 0.0\n    if g1_max != 0: g1 /= g1_max\n    if g2_max != 0: g2 /= g2_max\n    phase1 = int(np.argmin(g1)); phase2 = int(np.argmax(g2)) + min_idx\n    return phase1, phase2\n\ndef estimate_sigma_fgs(preprocessed_data, cfg):\n    \"\"\"Возвращает вектор sigma_1 (для FGS1) длиной N_planets — мягкий множитель к cfg.SIGMA.\"\"\"\n    sig_rel = []\n    delta = cfg.MODEL_OPTIMIZATION_DELTA\n    eps = 1e-12\n    for single in preprocessed_data:\n        # фазы по AIRS белой кривой — так же, как в модели\n        air_white = savgol_filter(single[:, 1:].mean(axis=1), 20, 2)\n        p1, p2 = _phase_detector_signal(air_white, cfg)\n        p1 = max(delta, p1)\n        p2 = min(len(air_white) - delta - 1, p2)\n\n        fgs = single[:, 0]\n        oot = (fgs[: p1 - delta] if p1 - delta > 0 else np.empty(0, fgs.dtype))\n        if p2 + delta < fgs.size:\n            oot = np.concatenate([oot, fgs[p2 + delta :]])\n        inn = fgs[p1 + delta : max(p1 + delta, p2 - delta)]\n\n        if oot.size == 0 or inn.size == 0:\n            sig_rel.append(np.nan); continue\n\n        n_oot, n_in = len(oot), len(inn)\n        var_oot = np.nanvar(oot, ddof=1)\n        var_in  = np.nanvar(inn, ddof=1)\n        oot_mean = float(np.nanmean(oot)) if np.isfinite(np.nanmean(oot)) else float(np.nanmean(fgs))\n        # относительная неопределённость глубины (в тех же ед., что s)\n        sigma_rel = np.sqrt(var_oot / max(n_oot,1) + var_in / max(n_in,1)) / max(oot_mean, eps)\n        sig_rel.append(sigma_rel)\n\n    s = np.asarray(sig_rel, dtype=float)\n    mask = np.isfinite(s) & (s > 0)\n    med = float(np.nanmedian(s[mask])) if mask.any() else 1.0\n\n    # мягкий множитель: корень, и узкий клип, чтобы не рисковать\n    k = np.ones_like(s)\n    if med > 0 and np.isfinite(med):\n        k[mask] = np.sqrt(s[mask] / med)\n    k = np.clip(k, 0.8, 1.25)  # ±20–25% от базовой σ\n\n    return k * cfg.SIGMA\n\ndef estimate_sigma_air(preprocessed_data, cfg):\n    \"\"\"Возвращает вектор sigma_air длиной N_planets — мягкий множитель к cfg.SIGMA для всех AIRS-каналов.\"\"\"\n    sig_rel = []\n    delta = cfg.MODEL_OPTIMIZATION_DELTA\n    eps = 1e-12\n\n    for single in preprocessed_data:\n        # белая кривая AIRS на бинированных данных (после всех твоих весов по λ)\n        white = np.nanmean(single[:, 1:], axis=1)         # (n_bins,)\n        white_s = savgol_filter(white, 20, 2)             # для фаз\n\n        p1, p2 = _phase_detector_signal(white_s, cfg)\n        p1 = max(delta, p1)\n        p2 = min(len(white) - delta - 1, p2)\n\n        oot_left = white[: p1 - delta] if p1 - delta > 0 else np.empty(0, white.dtype)\n        oot_right = white[p2 + delta :] if (p2 + delta) < white.size else np.empty(0, white.dtype)\n        oot = np.concatenate([oot_left, oot_right]) if (oot_left.size + oot_right.size) else oot_left\n        inn = white[p1 + delta : max(p1 + delta, p2 - delta)]\n\n        if oot.size == 0 or inn.size == 0:\n            sig_rel.append(np.nan); continue\n\n        n_oot, n_in = len(oot), len(inn)\n        var_oot = np.nanvar(oot, ddof=1)\n        var_in  = np.nanvar(inn, ddof=1)\n        oot_mean = float(np.nanmean(oot)) if np.isfinite(np.nanmean(oot)) else float(np.nanmean(white))\n\n        sigma_rel = np.sqrt(var_oot / max(n_oot,1) + var_in / max(n_in,1)) / max(oot_mean, eps)\n        sig_rel.append(sigma_rel)\n\n    s = np.asarray(sig_rel, dtype=float)\n    mask = np.isfinite(s) & (s > 0)\n    med = float(np.nanmedian(s[mask])) if mask.any() else 1.0\n\n    # мягкий множитель вокруг медианы\n    k = np.ones_like(s)\n    if med > 0 and np.isfinite(med):\n        k[mask] = np.sqrt(s[mask] / med)\n    k = np.clip(k, 0.90, 1.20)  # ±10%–20%\n\n    return k * cfg.SIGMA\n\nclass SignalProcessor:\n    def __init__(self, config):\n        self.cfg = config\n        self.adc_info = pd.read_csv(f\"{self.cfg.DATA_PATH}/adc_info.csv\")\n        self.planet_ids = pd.read_csv(f'{self.cfg.DATA_PATH}/{self.cfg.DATASET}_star_info.csv', index_col='planet_id').index.astype(int)\n\n    def _apply_linear_corr(self, linear_corr, signal):\n\n        coeffs = np.flip(linear_corr, axis=0)      # shape: (D, X, Y), D — старшая степень сначала\n        x = signal.astype(np.float64, copy=False)  # считаем в float64 для стабильности\n        out = np.empty_like(x, dtype=np.float64)\n        out[...] = coeffs[0]  # broadcast (X,Y) -> (T,X,Y)\n        for k in range(1, coeffs.shape[0]):\n            np.multiply(out, x, out=out)  # in-place умножение\n            out += coeffs[k]              # broadcast (X,Y)\n\n        return out.astype(signal.dtype, copy=False)\n\n    def _calibrate_single_signal(self, planet_id, sensor):\n        \"\"\"\n        Калибровка single-node сигнала.\n        Политика масок: DEAD — маскируем, HOT — НЕ маскируем (оставляем в данных).\n        \"\"\"\n        sensor_cfg = self.cfg.SENSOR_CONFIG[sensor]\n    \n        # --- load ---\n        signal = pd.read_parquet(\n            f\"{self.cfg.DATA_PATH}/{self.cfg.DATASET}/{planet_id}/{sensor}_signal_0.parquet\"\n        ).to_numpy()\n        dark = pd.read_parquet(\n            f\"{self.cfg.DATA_PATH}/{self.cfg.DATASET}/{planet_id}/{sensor}_calibration_0/dark.parquet\"\n        ).to_numpy()\n        dead = pd.read_parquet(\n            f\"{self.cfg.DATA_PATH}/{self.cfg.DATASET}/{planet_id}/{sensor}_calibration_0/dead.parquet\"\n        ).to_numpy()\n        flat = pd.read_parquet(\n            f\"{self.cfg.DATA_PATH}/{self.cfg.DATASET}/{planet_id}/{sensor}_calibration_0/flat.parquet\"\n        ).to_numpy()\n        linear_corr = pd.read_parquet(\n            f\"{self.cfg.DATA_PATH}/{self.cfg.DATASET}/{planet_id}/{sensor}_calibration_0/linear_corr.parquet\"\n        ).values.astype(np.float64).reshape(sensor_cfg[\"linear_corr_shape\"])\n    \n        # --- reshape & ADC ---\n        signal = signal.reshape(sensor_cfg[\"raw_shape\"])\n        gain = self.adc_info[f\"{sensor}_adc_gain\"].iloc[0]\n        offset = self.adc_info[f\"{sensor}_adc_offset\"].iloc[0]\n        signal = signal / gain + offset  # сохраняем твою формулу\n    \n        # HOT только для мониторинга, не для маскирования\n        hot = sigma_clip(dark, sigma=5, maxiters=5).mask\n    \n        # --- crop per sensor ---\n        if sensor == \"AIRS-CH0\":\n            signal = signal[:, :, self.cfg.CUT_INF : self.cfg.CUT_SUP]\n            linear_corr = linear_corr[:, :, self.cfg.CUT_INF : self.cfg.CUT_SUP]\n            dark = dark[:, self.cfg.CUT_INF : self.cfg.CUT_SUP]\n            dead = dead[:, self.cfg.CUT_INF : self.cfg.CUT_SUP]\n            flat = flat[:, self.cfg.CUT_INF : self.cfg.CUT_SUP]\n            hot = hot[:, self.cfg.CUT_INF : self.cfg.CUT_SUP]  # только для логов\n    \n        if sensor == \"FGS1\":\n            y0, y1, x0, x1 = 10, 22, 10, 22\n            signal = signal[:, y0:y1, x0:x1]\n            dark   = dark[y0:y1, x0:x1]\n            dead   = dead[y0:y1, x0:x1]\n            flat   = flat[y0:y1, x0:x1]\n            linear_corr = linear_corr[:, y0:y1, x0:x1]\n            hot    = hot[y0:y1, x0:x1]  # только для логов\n    \n        # --- non-neg clamp before linearity corr (как у тебя) ---\n        np.maximum(signal, 0, out=signal)\n    \n        # --- linearity correction ---\n        if sensor == \"FGS1\":\n            signal = self._apply_linear_corr(linear_corr, signal)\n        elif sensor == \"AIRS-CH0\":\n            sl = (slice(None), slice(10, 22), slice(None))  # T, Y, λ\n            signal[sl] = self._apply_linear_corr(linear_corr[:, 10:22, :], signal[sl])\n        else:\n            signal = self._apply_linear_corr(linear_corr, signal)\n    \n        # --- dark subtraction с учётом паттерна интеграций ---\n        base_dt, increment = sensor_cfg[\"dt_pattern\"]\n        even_scale = base_dt\n        odd_scale  = base_dt + increment\n        signal[::2]  -= dark * even_scale\n        signal[1::2] -= dark * odd_scale\n    \n        # --- APPLY FLAT (HOT-KEEP: не включаем hot в маску!) ---\n        if sensor == \"FGS1\":\n            flat_roi = flat.astype(signal.dtype, copy=False).copy()      # (12,12)\n            bad = (dead) | ~np.isfinite(flat_roi) | (flat_roi == 0)      # ← ТОЛЬКО dead/invalid\n            flat_roi[bad] = np.nan\n            signal /= flat_roi\n    \n        elif sensor == \"AIRS-CH0\":\n            y0, y1 = 10, 22\n            flat_roi = flat[y0:y1, :].astype(signal.dtype, copy=False).copy()  # (12, λ)\n            bad = (dead[y0:y1, :]) | ~np.isfinite(flat_roi) | (flat_roi == 0)  # ← ТОЛЬКО dead/invalid\n            flat_roi[bad] = np.nan\n            signal[:, y0:y1, :] /= flat_roi\n    \n        else:\n            flat2 = flat.astype(signal.dtype, copy=False).copy()\n            bad2 = (dead) | ~np.isfinite(flat2) | (flat2 == 0)                  # ← ТОЛЬКО dead/invalid\n            flat2[bad2] = np.nan\n            signal /= flat2\n        # --- END FLAT ---\n    \n        # (опционально) логируем метрики hot/dead\n        if getattr(self.cfg, \"LOG_HOT_STATS\", False):\n            if not hasattr(self, \"stats\"):\n                self.stats = []\n            self.stats.append({\n                \"planet_id\": int(planet_id),\n                \"sensor\": sensor,\n                \"hot_frac\": float(np.mean(hot)),\n                \"dead_frac\": float(np.mean(dead)),\n            })\n    \n        return signal\n\n    def _preprocess_calibrated_signal(self, calibrated_signal, sensor):\n        sensor_cfg = self.cfg.SENSOR_CONFIG[sensor]\n        binning = sensor_cfg[\"binning\"]\n\n        if sensor == \"AIRS-CH0\":\n            signal_roi = calibrated_signal[:, 10:22, :]\n        elif sensor == \"FGS1\":\n            signal_roi = calibrated_signal[:, 10:22, 10:22]\n            signal_roi = signal_roi.reshape(signal_roi.shape[0], -1)\n        \n        mean_signal = np.nanmean(signal_roi, axis=1)\n\n        cds_signal = mean_signal[1::2] - mean_signal[0::2]\n\n        n_bins = cds_signal.shape[0] // binning\n        binned = np.array([\n            cds_signal[j*binning : (j+1)*binning].mean(axis=0) \n            for j in range(n_bins)\n        ])\n\n        if sensor == \"AIRS-CH0\":\n            q_lo = np.nanpercentile(binned, 5.0, axis=1, keepdims=True)    # (n_bins, 1)\n            q_hi = np.nanpercentile(binned, 95.0, axis=1, keepdims=True)   # (n_bins, 1)\n            np.clip(binned, q_lo, q_hi, out=binned)\n\n        if sensor == \"FGS1\":\n            binned = binned.reshape((binned.shape[0], 1))\n\n        if sensor == \"AIRS-CH0\":\n            var = np.nanvar(binned, axis=0, ddof=1)                 # (λ, )\n            med = np.nanmedian(var)\n            safe_var = np.where(~np.isfinite(var) | (var <= 0), med if (np.isfinite(med) and med > 0) else 1.0, var)\n            w = 1.0 / safe_var\n\n            lo, hi = np.nanpercentile(w, 5.0), np.nanpercentile(w, 95.0)\n            if np.isfinite(lo) and np.isfinite(hi) and lo < hi:\n                w = np.clip(w, lo, hi)\n\n            M = binned.shape[1]\n            s = np.nansum(w)\n            if np.isfinite(s) and s > 0:\n                w = w * (M / s)\n            else:\n                w = np.ones_like(w)\n\n            binned *= w[None, :]\n\n\n        return binned\n\n    def _process_planet_sensor(self, args):\n        planet_id, sensor = args['planet_id'], args['sensor']\n        calibrated = self._calibrate_single_signal(planet_id, sensor)\n        preprocessed = self._preprocess_calibrated_signal(calibrated, sensor)\n        return preprocessed\n\n    def process_all_data(self):\n        args_fgs1 = [dict(planet_id=planet_id, sensor=\"FGS1\") for planet_id in self.planet_ids]\n        preprocessed_fgs1 = pqdm(args_fgs1, self._process_planet_sensor, n_jobs=self.cfg.N_JOBS)\n\n        args_airs_ch0 = [dict(planet_id=planet_id, sensor=\"AIRS-CH0\") for planet_id in self.planet_ids]\n        preprocessed_airs_ch0 = pqdm(args_airs_ch0, self._process_planet_sensor, n_jobs=self.cfg.N_JOBS)\n\n        preprocessed_signal = np.concatenate(\n            [np.stack(preprocessed_fgs1), np.stack(preprocessed_airs_ch0)], axis=2\n        )\n        return preprocessed_signal\n    \n\nclass TransitModel:\n    def __init__(self, config):\n        self.cfg = config\n\n    def _phase_detector(self, signal):\n        search_slice = self.cfg.MODEL_PHASE_DETECTION_SLICE\n        min_index = np.argmin(signal[search_slice]) + search_slice.start\n        \n        signal1 = signal[:min_index]\n        signal2 = signal[min_index:]\n\n        grad1 = np.gradient(signal1)\n        grad1 /= grad1.max()\n        \n        grad2 = np.gradient(signal2)\n        grad2 /= grad2.max()\n\n        phase1 = np.argmin(grad1)\n        phase2 = np.argmax(grad2) + min_index\n\n        return phase1, phase2\n    \n    def _objective_function(self, s, signal, phase1, phase2):\n        delta = self.cfg.MODEL_OPTIMIZATION_DELTA\n        power = self.cfg.MODEL_POLYNOMIAL_DEGREE\n\n        if phase1 - delta <= 0 or phase2 + delta >= len(signal) or phase2 - delta - (phase1 + delta) < 5:\n            delta = 2\n\n        y = np.concatenate([\n            signal[: phase1 - delta],\n            signal[phase1 + delta : phase2 - delta] * (1 + s),\n            signal[phase2 + delta :]\n        ])\n        x = np.arange(len(y))\n\n        coeffs = np.polyfit(x, y, deg=power)\n        poly = np.poly1d(coeffs)\n        error = np.abs(poly(x) - y).mean()\n        \n        return error\n\n    def predict(self, single_preprocessed_signal):\n        signal_1d = single_preprocessed_signal[:, 1:].mean(axis=1)\n        signal_1d = savgol_filter(signal_1d, 23, 2)\n        \n        phase1, phase2 = self._phase_detector(signal_1d)\n\n        phase1 = max(self.cfg.MODEL_OPTIMIZATION_DELTA, phase1)\n        phase2 = min(len(signal_1d) - self.cfg.MODEL_OPTIMIZATION_DELTA - 1, phase2)    \n\n        result = minimize(\n            fun=self._objective_function,\n            x0=[0.0001],\n            args=(signal_1d, phase1, phase2),\n            method=\"Nelder-Mead\"\n        )\n        \n        return result.x[0]\n\n    def predict_all(self, preprocessed_signals):\n        predictions = [\n            self.predict(preprocessed_signal)\n            for preprocessed_signal in tqdm(preprocessed_signals)\n        ]\n        return np.array(predictions) * self.cfg.SCALE\n\nStarInfo = pd.read_csv(ROOT_PATH + f\"/{MODE}_star_info.csv\")\nStarInfo[\"planet_id\"] = StarInfo[\"planet_id\"].astype(int)\nPlanetIds = StarInfo[\"planet_id\"].tolist()\nStarInfo = StarInfo.set_index(\"planet_id\")\n\nclass SubmissionGenerator:\n    def __init__(self, config):\n        self.cfg = config\n        self.sample_submission = pd.read_csv(\"/kaggle/input/ariel-data-challenge-2025/sample_submission.csv\", index_col=\"planet_id\")\n\n    def create(self, predictions1, predictions, sigma_fgs=None, sigma_air=None):\n        planet_ids = self.sample_submission.index\n        n_mu = self.sample_submission.shape[1] // 2  # 283\n\n        preds = np.asarray(predictions, dtype=float).reshape(-1)\n        mu = np.tile(preds.reshape(-1, 1), (1, n_mu))\n        mu = np.clip(mu, 0, None)\n\n        sigmas = np.full_like(mu, self.cfg.SIGMA, dtype=float)\n        if sigma_fgs is not None:\n            sigma_fgs = np.asarray(sigma_fgs, dtype=float).reshape(-1)\n            sigmas[:, 0] = np.clip(sigma_fgs, 1e-6, 0.1)\n        if sigma_air is not None:\n            # sigma_air should already have the correct shape (num_planets, 282)\n            sigma_air_processed = np.asarray(sigma_air, dtype=float)\n            sigmas[:, 1:] = np.clip(sigma_air_processed, 1e-6, 0.1)\n\n        submission_df = pd.DataFrame(\n            np.concatenate([mu, sigmas], axis=1),\n            columns=self.sample_submission.columns,\n            index=planet_ids\n        )\n        submission_df.iloc[:, 0] = predictions\n        submission_df.iloc[:, 1:283] = predictions1\n        submission_df.to_csv(\"submission.csv\")\n        \n        return submission_df\n\nclass SEBlock(nn.Module):\n    def __init__(self, dim, reduction=16):\n        super().__init__()\n        self.fc1 = nn.Linear(dim, dim // reduction, bias=False)\n        self.fc2 = nn.Linear(dim // reduction, dim, bias=False)\n        self.act = nn.SiLU()   # Or nn.ReLU()\n\n    def forward(self, x):\n        # Compute channel-wise attention\n        w = x.mean(dim=0, keepdim=True)      # Global context (mean across batch)\n        w = self.act(self.fc1(w))\n        w = torch.sigmoid(self.fc2(w))\n        return x * w   # Rescale input\n\nclass AttentionBlock(nn.Module):\n    def __init__(self, dim, num_heads=4, dropout=0.1):\n        super().__init__()\n        self.attn = nn.MultiheadAttention(embed_dim=dim, num_heads=num_heads, batch_first=True)\n        self.norm = nn.LayerNorm(dim)\n        self.dropout = nn.Dropout(dropout)\n\n    def forward(self, x):\n        # Expect shape [batch, seq_len, dim], so expand if only [batch, dim]\n        if x.dim() == 2:\n            x = x.unsqueeze(1)   # -> [batch, 1, dim]\n        attn_out, _ = self.attn(x, x, x)\n        out = self.norm(x + self.dropout(attn_out))\n        return out.squeeze(1)    # Back to [batch, dim]\n\nclass ResidualBlock2(nn.Module):\n    def __init__(self, dim, p=0.2):\n        super().__init__()\n        self.fc1 = nn.Linear(dim, dim)\n        self.fc2 = nn.Linear(dim, dim)\n        self.relu = nn.ReLU()\n        self.dropout = nn.Dropout(p)\n\n    def forward(self, x):\n        identity = x\n        out = self.relu(self.fc1(x))\n        out = self.dropout(out)\n        out = self.fc2(out)\n        return self.relu(out + identity)\n\nclass ResNetMLP2(nn.Module):\n    def __init__(self, input_dim=3, hidden_dim=128, output_dim=282, num_blocks=3, dropout_rate=0.2):\n        super().__init__()\n        self.input_layer = nn.Linear(input_dim, hidden_dim)\n        self.blocks = nn.Sequential(*[ResidualBlock2(hidden_dim, p=dropout_rate) for _ in range(num_blocks)])\n        self.output_layer = nn.Linear(hidden_dim, output_dim)\n\n    def forward(self, x):\n        x = self.input_layer(x)\n        x = self.blocks(x)\n        x = self.output_layer(x)\n        return x\n        \nimport torch.distributions as D\n\n# This is now an MDN model, but with the same class name for easy integration\nclass ResNetMLP_WithUncertainty(nn.Module):\n    def __init__(self, input_dim, hidden_dim, output_dim, num_blocks, dropout_rate, n_gaussians=3):\n        super().__init__()\n        self.n_gaussians = n_gaussians\n        self.output_dim = output_dim\n        \n        # The main body of the network\n        self.blocks = nn.Sequential(\n            nn.Linear(input_dim, hidden_dim),\n            *[ResidualBlock2(hidden_dim, p=dropout_rate) for _ in range(num_blocks)]\n        )\n        \n        # Three separate heads for the mixture parameters\n        self.pi_head = nn.Linear(hidden_dim, output_dim * n_gaussians)\n        self.mu_head = nn.Linear(hidden_dim, output_dim * n_gaussians)\n        self.sigma_head = nn.Linear(hidden_dim, output_dim * n_gaussians)\n\n    def forward(self, x):\n        x = self.blocks(x)\n        \n        # Reshape and apply activations to the outputs\n        pi_logits = self.pi_head(x).view(-1, self.output_dim, self.n_gaussians)\n        pi = torch.softmax(pi_logits, dim=2)\n        \n        mu = self.mu_head(x).view(-1, self.output_dim, self.n_gaussians)\n        \n        log_sigma = self.sigma_head(x).view(-1, self.output_dim, self.n_gaussians)\n        sigma = torch.exp(log_sigma)\n        \n        return pi, mu, sigma\n\n# This is now the MDN loss function, with the same name for easy integration\ndef nll_loss(pi, mu, sigma, targets):\n    \"\"\"\n    Negative Log-Likelihood loss for a Gaussian Mixture Model.\n    \"\"\"\n    targets = targets.unsqueeze(2) # Reshape targets for mixture calculation\n    normal_dist = D.Normal(mu, sigma)\n    log_probs = normal_dist.log_prob(targets)\n    log_sum_exp = torch.logsumexp(torch.log(pi) + log_probs, dim=2)\n    return -torch.mean(log_sum_exp)\n    \ndef train_resnet_model(X_train, y_train, X_val, y_val, model_params, training_params):\n    \"\"\"\n    Trains the ResNetMLP_WithUncertainty model with weight decay and early stopping.\n    \"\"\"\n    device = torch.device(\"cuda\" if torch.cuda.is_available() else \"cpu\")\n    print(f\"Using device: {device}\")\n\n    train_dataset = TensorDataset(torch.tensor(X_train, dtype=torch.float64), torch.tensor(y_train, dtype=torch.float64))\n    val_dataset = TensorDataset(torch.tensor(X_val, dtype=torch.float64), torch.tensor(y_val, dtype=torch.float64))\n\n    train_loader = DataLoader(train_dataset, batch_size=training_params['batch_size'], shuffle=True)\n    val_loader = DataLoader(val_dataset, batch_size=training_params['batch_size'], shuffle=False)\n\n    model = ResNetMLP_WithUncertainty(**model_params).double().to(device)\n    # loss_function = nn.MSELoss()\n\n    # NEW: Add weight_decay to the optimizer from training_params\n\n    optimizer = torch.optim.Adam(\n        model.parameters(),\n        lr=training_params['learning_rate'],\n        weight_decay=training_params.get('weight_decay', 0) # Defaults to 0 if not provided\n    )\n\n    # Add scheduler\n    scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(\n        optimizer,\n        mode='min',        # we want to minimize validation loss\n        factor=0.5,        # shrink LR by half each time\n        patience=5,        # wait 5 epochs without improvement\n        verbose=True       # prints a message when LR is reduced\n    )\n\n    best_val_loss = float('inf')\n    epochs = training_params['epochs']\n\n    # NEW: Initialize variables for early stopping\n    patience = training_params.get('patience', 20) # Defaults to 20\n    epochs_no_improve = 0\n\n    # --- Training Loop ---\n    for epoch in range(epochs):\n        model.train()\n        running_train_loss = 0.0\n        for inputs, targets in train_loader:\n            inputs, targets = inputs.to(device), targets.to(device)\n            optimizer.zero_grad()\n\n            pi_outputs, mu_outputs, sigma_outputs = model(inputs)\n\n            # NEW: Use the custom NLL loss function\n            loss = nll_loss(pi_outputs, mu_outputs, sigma_outputs, targets)\n\n            loss.backward()\n            optimizer.step()\n            running_train_loss += loss.item() * inputs.size(0)\n        epoch_train_loss = running_train_loss / len(train_loader.dataset)\n\n        # --- Validation Loop ---\n        model.eval()\n        running_val_loss = 0.0\n        with torch.no_grad():\n            for inputs, targets in val_loader:\n                inputs, targets = inputs.to(device), targets.to(device)\n                \n                # FIX: Unpack the three MDN outputs\n                pi_outputs, mu_outputs, sigma_outputs = model(inputs)\n                \n                # FIX: Call the loss function with the correct arguments\n                loss = nll_loss(pi_outputs, mu_outputs, sigma_outputs, targets)\n        \n                running_val_loss += loss.item() * inputs.size(0)\n        epoch_val_loss = running_val_loss / len(val_loader.dataset)\n\n        print(f\"Epoch {epoch+1}/{epochs} | Train Loss: {epoch_train_loss:.6f} | Val Loss: {epoch_val_loss:.6f}\")\n\n        # Step the scheduler with the latest validation loss\n        scheduler.step(epoch_val_loss)\n\n        # NEW: Early stopping logic\n        if epoch_val_loss < best_val_loss:\n            best_val_loss = epoch_val_loss\n            torch.save(model.state_dict(), training_params['save_path'])\n            print(f\"Validation loss decreased. Model saved to {training_params['save_path']}\")\n            epochs_no_improve = 0 # Reset counter\n        else:\n            epochs_no_improve += 1\n            print(f\"Validation loss did not improve for {epochs_no_improve} epoch(s).\")\n        \n        if epochs_no_improve >= patience:\n            print(f\"Early stopping triggered after {patience} epochs with no improvement.\")\n            break # Exit the training loop\n\n    print(\"Training finished.\")\n\ndef train_new_model(hidden_dim=256, num_blocks=35, dropout_rate=0.1):\n    \"\"\"\n    This function handles the entire training pipeline:\n    1. Generates REAL transit_depth features from the training data.\n    2. Trains the ResNetMLP_WithUncertainty model.\n    3. Saves and returns the trained model.\n    \"\"\"\n    # === A. Generate REAL features for training ===\n    train_config = Config()\n    train_config.DATASET = \"train\"\n    if not Config.load_data:\n        # print(\"Generating real transit_depth features for training set...\")\n        # processor = SignalProcessor(train_config)\n        # preprocessed_train_data = processor.process_all_data()\n        # # Save preprocessed_train_data for reproducibility/debugging\n        # np.save(\"preprocessed_train_data.npy\", preprocessed_train_data)\n        print(\"Loading real transit_depth features for training set...\")\n        preprocessed_train_data = np.load('/kaggle/input/ariel-2025-result-2/results_sv43/preprocessed_train_data.npy')\n        print(f\"Preprocessed training data shape: {preprocessed_train_data.shape}\")\n        model_1d = TransitModel(train_config)\n        train_predictions_1d = model_1d.predict_all(preprocessed_train_data)\n        \n        train_star_info = pd.read_csv(f\"{train_config.DATA_PATH}/train_star_info.csv\")\n        train_targets = pd.read_csv(f\"{train_config.DATA_PATH}/train.csv\")\n        planet_ids = pd.read_csv(f'{train_config.DATA_PATH}/{train_config.DATASET}_star_info.csv', index_col='planet_id').index.astype(int)\n\n        predictions_df = pd.DataFrame({\n            \"planet_id\": planet_ids,\n            \"transit_depth\": train_predictions_1d\n        })\n        \n        # === B. Prepare X and y for training ===\n        input_df = pd.merge(predictions_df, train_star_info, on=\"planet_id\", how=\"left\")\n        input_df['scaled_depth'] = input_df['transit_depth'] / input_df['Rs']\n        # Using the simplified formula T_eq ∝ T_star * sqrt(R_star / a)\n        input_df['planet_temp'] = input_df['Ts'] * np.sqrt(input_df['Rs'] / (2 * input_df['sma']))\n        input_df['stellar_density'] = input_df['Ms'] / (input_df['Rs']**3)\n        input_df['stellar_gravity'] = input_df['Ms'] / (input_df['Rs']**2)\n\n        X = input_df[Config.FEATURES].values.astype(np.float64)\n        \n        full_train_df = pd.merge(input_df, train_targets, on=\"planet_id\", how=\"left\")\n        # Save full_train_df for reproducibility/debugging\n        full_train_df.to_csv(\"full_train_df.csv\", index=False)\n        target_cols = [f'wl_{i}' for i in range(2, 284)]\n        y = full_train_df[target_cols].values.astype(np.float64)\n        # Save X and y for reproducibility/debugging\n        np.save(\"X_train_features.npy\", X)\n        np.save(\"y_train_targets.npy\", y)\n    else:\n        # Load X and y if they exist (for reproducibility/debugging)\n        # Define the directory containing the files\n        data_dir = '/kaggle/input/ariel-2025-data/data_f9'\n        # Pick X and y file from the directory\n        X_file = os.path.join(data_dir, 'X_train_features.npy')\n        y_file = os.path.join(data_dir, 'y_train_targets.npy')\n        # Load preprocessed_train_data for sigma estimation in CV\n        preprocessed_train_data = np.load(os.path.join(data_dir, 'preprocessed_train_data.npy'))\n        # features = ['transit_depth', 'Rs', 'i']\n        if os.path.exists(X_file) and os.path.exists(y_file):\n            X = np.load(X_file)\n            y = np.load(y_file)\n        else:\n            raise FileNotFoundError(f\"Could not find {X_file} or {y_file}. Please check your data path or set Config.load_data = False.\")\n    # Standard scaling for X and y\n    scaler_X = StandardScaler()\n    scaler_y = StandardScaler()\n    X_scaled = scaler_X.fit_transform(X)\n    y_scaled = scaler_y.fit_transform(y)\n    # Print summary statistics for X and y\n    X_df = pd.DataFrame(X_scaled, columns=Config.FEATURES)\n    y_df = pd.DataFrame(y_scaled, columns=[f'wl_{i}' for i in range(2, 284)])\n    # print(\"X feature summary (scaled):\")\n    # print(X_df.describe())\n    # print(\"\\ny target summary (scaled):\")\n    # print(y_df.describe())\n\n    all_scores = []\n    all_mae = []\n    all_rmse = []\n    all_r2 = []\n    all_models = []\n    fold = 1\n    model_params = {\n        'input_dim': len(Config.FEATURES),  # Always use the current feature count\n        'hidden_dim': hidden_dim,\n        'output_dim': 282,\n        'num_blocks': num_blocks,\n        'dropout_rate': dropout_rate\n    }\n    training_params = {\n        'learning_rate': 0.001,\n        'epochs': 5 if Config.DEBUG else 500,\n        'batch_size': 64,\n        'save_path': 'best_model_airs_cv.pth',\n        'patience': 20,\n        'weight_decay': 1e-5\n    }\n    from sklearn.model_selection import StratifiedKFold # Import the new class\n    \n    # 1. Create a binned version of your target variable to stratify on.\n    # We can use the mean of the target spectrum for each planet as a representative value.\n    y_mean = y.mean(axis=1)\n    # pd.cut creates discrete bins from the continuous values\n    strata = pd.cut(y_mean, bins=10, labels=False, duplicates='drop')\n    \n    # 2. Use StratifiedKFold instead of KFold\n    n_splits = 10\n    # Use StratifiedKFold and pass the strata to the split() method\n    skf = StratifiedKFold(n_splits=n_splits, shuffle=True, random_state=42)\n    \n    # The CV loop now uses skf.split(X, strata)\n    for train_index, val_index in skf.split(X_scaled, strata):\n    \n    # for train_index, val_index in kf.split(X_scaled):\n        print(f\"\\n--- Fold {fold}/{n_splits} ---\")\n        X_train, X_val = X_scaled[train_index], X_scaled[val_index]\n        y_train, y_val = y_scaled[train_index], y_scaled[val_index]\n        \n        # 1. Get the preprocessed time-series data for the validation fold\n        preprocessed_data_val = preprocessed_train_data[val_index]\n        \n        # 2. Calculate the statistical sigmas using this validation data slice\n        # Make sure you have a config object available (e.g., train_config)\n        sigma_fgs_val = estimate_sigma_fgs(preprocessed_data_val, train_config)\n        sigma_air_val = estimate_sigma_air(preprocessed_data_val, train_config)\n        \n        # Update save_path for each fold to avoid overwriting\n        training_params['save_path'] = f'best_model_airs_cv_fold{fold}.pth'\n        train_resnet_model(X_train, y_train, X_val, y_val, model_params, training_params)\n        trained_model = ResNetMLP_WithUncertainty(**model_params).double()\n        model_path = training_params['save_path']\n        trained_model.load_state_dict(torch.load(model_path, map_location=torch.device('cpu')))\n        trained_model.eval()\n        # Pass the newly calculated sigmas to the evaluation function\n        val_score, mae, rmse, r2 = evaluate_model_on_val(\n            trained_model, \n            X_val, y_val, \n            X_train, y_train, \n            scaler_y, \n            sigma_fgs_val # Pass the calculated FGS sigma\n        )\n        print(f\"Fold {fold} - Validation Score: {val_score:.5f}, MAE: {mae:.5f}, RMSE: {rmse:.5f}, R2: {r2:.5f}\")\n        all_scores.append(val_score)\n        all_mae.append(mae)\n        all_rmse.append(rmse)\n        all_r2.append(r2)\n        all_models.append(trained_model)\n        fold += 1\n    avg_score = np.mean(all_scores)\n    avg_mae = np.mean(all_mae)\n    avg_rmse = np.mean(all_rmse)\n    avg_r2 = np.mean(all_r2)\n    print(f\"\\n==== Cross-Validation Results (averaged over {n_splits} folds) ====\")\n    print(f\"Average Validation Score (custom metric): {avg_score:.5f} ± {np.std(all_scores):.5f}\")\n    print(f\"Average MAE: {avg_mae:.5f} ± {np.std(all_mae):.5f}, Average RMSE: {avg_rmse:.5f} ± {np.std(all_rmse):.5f}, Average R2: {avg_r2:.5f} ± {np.std(all_r2):.5f}\")\n\n    # Save scalers for reproducibility and later use\n    import joblib\n    joblib.dump(scaler_X, 'scaler_X.joblib')\n    joblib.dump(scaler_y, 'scaler_y.joblib')\n    # Return models and scalers for use in prediction\n    return all_models, scaler_X, scaler_y, avg_score, avg_mae, avg_rmse, avg_r2\n\n\n# --- Hyperparameter sweep function ---\ndef sweep_hyperparameters():\n    \"\"\"\n    Sweeps through all combinations of hidden_dim, num_blocks, and dropout_rate,\n    calls train_new_model for each, and returns a DataFrame of results.\n    \"\"\"\n    import pandas as pd\n    import itertools\n    results = []\n    hidden_dim_list, num_blocks_list, dropout_rate_list = [64, 128, 256], [75, 80, 85], [0.1, 0.2, 0.3]\n    for hidden_dim, num_blocks, dropout_rate in itertools.product(hidden_dim_list, num_blocks_list, dropout_rate_list):\n        print(f\"\\n=== Training with hidden_dim={hidden_dim}, num_blocks={num_blocks}, dropout_rate={dropout_rate} ===\")\n        try:\n            _, _, _, mean_mae, mean_rmse, mean_r2 = train_new_model(hidden_dim=hidden_dim, num_blocks=num_blocks, dropout_rate=dropout_rate)\n            results.append({\n                'hidden_dim': hidden_dim,\n                'num_blocks': num_blocks,\n                'dropout_rate': dropout_rate,\n                'mean_mae': mean_mae,\n                'mean_rmse': mean_rmse,\n                'mean_r2': mean_r2\n            })\n            print(f\"Results so far: {results[-1]}\")\n        except Exception as e:\n            print(f\"Error for hidden_dim={hidden_dim}, num_blocks={num_blocks}, dropout_rate={dropout_rate}: {e}\")\n            results.append({\n                'hidden_dim': hidden_dim,\n                'num_blocks': num_blocks,\n                'dropout_rate': dropout_rate,\n                'mean_mae': None,\n                'mean_rmse': None,\n                'mean_r2': None,\n                'error': str(e)\n            })\n    df = pd.DataFrame(results)\n    df.to_csv(\"hyperparameter_sweep_results.csv\", index=False)\n    print(\"\\nHyperparameter sweep results:\")\n    print(df)\n    return df\n\n# ADD THIS HELPER FUNCTION to your script\ndef get_final_preds_from_mdn(pi, mu, sigma):\n    \"\"\"\n    Calculates the final mean and sigma from the MDN's mixture parameters.\n    \"\"\"\n    # Final mean (mu) is the weighted average of the component means\n    final_mu = torch.sum(pi * mu, dim=2)\n\n    # Final variance is calculated from the mixture components\n    variance_within = torch.sum(pi * (sigma**2), dim=2)\n    variance_between = torch.sum(pi * (mu**2), dim=2) - final_mu**2\n    final_sigma = torch.sqrt(variance_within + variance_between)\n    \n    return final_mu, final_sigma\n\n# REPLACE your old evaluate_model_on_val with this new version\ndef evaluate_model_on_val(trained_model, X_val, y_val, X_train, y_train, scaler_y, sigma_fgs_val):\n    \"\"\"\n    Evaluates the MDN-trained model on the validation set.\n    \"\"\"\n    trained_model.eval()\n    device = torch.device(\"cuda\" if torch.cuda.is_available() else \"cpu\")\n    trained_model.to(device)\n    X_val_tensor = torch.tensor(X_val, dtype=torch.float64).to(device)\n    \n    with torch.no_grad():\n        # 1. Get the three mixture parameter outputs from the MDN model\n        pi_val_scaled, mu_val_scaled, sigma_val_scaled = trained_model(X_val_tensor)\n        \n        # 2. Calculate the final mean and sigma using the helper function\n        y_pred_val_scaled, y_sigma_val_scaled = get_final_preds_from_mdn(\n            pi_val_scaled, mu_val_scaled, sigma_val_scaled\n        )\n\n        y_pred_val_scaled = y_pred_val_scaled.cpu().numpy()\n        y_sigma_val_scaled = y_sigma_val_scaled.cpu().numpy()\n\n    # 3. Inverse transform all predictions and targets to their original scale\n    y_pred_val = scaler_y.inverse_transform(y_pred_val_scaled)\n    y_val_orig = scaler_y.inverse_transform(y_val)\n    y_sigma_val = y_sigma_val_scaled * scaler_y.scale_\n\n    # 4. Prepare the 'solution' DataFrame\n    target_cols = [f'wl_{i}' for i in range(2, 284)] # For AIRS channels\n    val_df = pd.DataFrame(y_val_orig, columns=target_cols)\n    val_df['planet_id'] = np.arange(len(val_df))\n\n    # 5. Prepare the 'submission' DataFrame\n    pred_sigmas = np.zeros_like(y_pred_val)\n    pred_sigmas[:, 0] = sigma_fgs_val\n    pred_sigmas[:, 1:] = y_sigma_val[:, 1:]\n\n    pred_full = np.concatenate([y_pred_val, pred_sigmas], axis=1)\n    # ... The rest of the function is the same as your previous version ...\n    pred_full_cols = target_cols + [f'sigma_{i}' for i in range(2, 284)]\n    pred_full_df = pd.DataFrame(pred_full, columns=pred_full_cols)\n    pred_full_df['planet_id'] = np.arange(len(pred_full_df))\n    y_train_orig = scaler_y.inverse_transform(y_train)\n    naive_mean = y_train_orig.mean()\n    naive_sigma = y_train_orig.std()\n    val_score = score(\n        solution=val_df,\n        submission=pred_full_df,\n        row_id_column_name='planet_id',\n        naive_mean=naive_mean,\n        naive_sigma=naive_sigma\n    )\n    y_true_flat = y_val_orig.flatten()\n    y_pred_flat = y_pred_val.flatten()\n    mae = mean_absolute_error(y_true_flat, y_pred_flat)\n    rmse = mean_squared_error(y_true_flat, y_pred_flat, squared=False)\n    r2 = r2_score(y_true_flat, y_pred_flat)\n    \n    print(f\"Validation Score (custom metric): {val_score:.5f}\")\n    print(f\"MAE: {mae:.5f}, RMSE: {rmse:.5f}, R2: {r2:.5f}\")\n    \n    return val_score, mae, rmse, r2\n    \ndef load_cv_models_and_scalers(directory):\n    \"\"\"\n    Loads all cross-validation models and scalers from the specified directory.\n    Args:\n        directory (str): Path to the directory containing model and scaler files.\n    Returns:\n        all_models (list): List of loaded models.\n        scaler_X: Loaded X scaler.\n        scaler_y: Loaded y scaler.\n    \"\"\"\n    import os\n    import joblib\n    # Load scalers\n    scaler_X = joblib.load(os.path.join(directory, 'scaler_X.joblib'))\n    scaler_y = joblib.load(os.path.join(directory, 'scaler_y.joblib'))\n\n    # Load all CV models\n    all_models = []\n    model_params = {\n        'input_dim': len(Config.FEATURES),  # Always use the current feature count\n        'hidden_dim': 256,\n        'output_dim': 282,\n        'num_blocks': 35,\n        'dropout_rate': 0.1\n    }\n    for fold in range(1, 11):\n        model = ResNetMLP_WithUncertainty(**model_params).double()\n        model_path = os.path.join(directory, f'best_model_airs_cv_fold{fold}.pth')\n        model.load_state_dict(torch.load(model_path, map_location=torch.device('cpu')))\n        model.eval()\n        all_models.append(model)\n    return all_models, scaler_X, scaler_y\n\nconfig = Config()\n# sweep_hyperparameters()\nif config.load_model:\n    all_models, scaler_X, scaler_y = load_cv_models_and_scalers('/kaggle/input/ariel-2025-result-2/results_v39')\nelse:\n    all_models, scaler_X, scaler_y, avg_score, avg_mae, avg_rmse, avg_r2 = train_new_model(hidden_dim=128, num_blocks=30, dropout_rate=0.3)\nsignal_processor = SignalProcessor(config)\npreprocessed_data = signal_processor.process_all_data()\n\nmodel = TransitModel(config)\npredictions = model.predict_all(preprocessed_data)\nsigma_fgs_vec = estimate_sigma_fgs(preprocessed_data, config)\nsigma_air_vec = estimate_sigma_air(preprocessed_data, config)\n\npredictions_df = pd.DataFrame({\n    \"planet_id\": PlanetIds,\n    \"transit_depth\": predictions\n})\n\ninput_df = pd.merge(predictions_df, StarInfo, on=\"planet_id\", how=\"left\")\ninput_df['scaled_depth'] = input_df['transit_depth'] / input_df['Rs']\ninput_df['planet_temp'] = input_df['Ts'] * np.sqrt(input_df['Rs'] / (2 * input_df['sma']))\ninput_df['stellar_density'] = input_df['Ms'] / (input_df['Rs']**3)\ninput_df['stellar_gravity'] = input_df['Ms'] / (input_df['Rs']**2)\n\nX = input_df[Config.FEATURES].values.astype(np.float64)\nX_scaled = scaler_X.transform(X)\nX_tensor = torch.tensor(X_scaled, dtype=torch.float64)\n\nall_means_scaled = []\nall_sigmas_scaled = []\n\nwith torch.no_grad():\n    for model in all_models:\n        pi, mu, sigma = model(X_tensor)\n\n        # Final mean (mu) is the weighted average of the component means\n        final_mu = torch.sum(pi * mu, dim=2)\n    \n        # Final variance is calculated from the mixture components\n        variance_within = torch.sum(pi * (sigma**2), dim=2)\n        variance_between = torch.sum(pi * (mu**2), dim=2) - final_mu**2\n        final_sigma = torch.sqrt(variance_within + variance_between)\n        \n        all_means_scaled.append(final_mu.numpy())\n        all_sigmas_scaled.append(final_sigma.numpy())\n\n# Average the means from all models\nfinal_mean_scaled = np.mean(all_means_scaled, axis=0)\n\n# Combine the uncertainties (root mean square of the sigmas)\nfinal_sigma_scaled = np.sqrt(np.mean(np.square(all_sigmas_scaled), axis=0))\n\n\n# --- Inverse transform to get final values for submission ---\npredictions1 = scaler_y.inverse_transform(final_mean_scaled)\n# Remember to only scale the sigma, not shift it\npredictions1_sigma = final_sigma_scaled * scaler_y.scale_","metadata":{"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import os\nimport time\nimport itertools\nimport multiprocessing as mp\n\nimport numpy as np\nimport pandas as pd\nimport pandas.api.types\n\nimport torch\nimport torch.nn as nn\nimport torch.nn.functional as F\nfrom torch.utils.data import DataLoader, TensorDataset, random_split\n\nimport matplotlib.pyplot as plt\n\nfrom tqdm import tqdm\nfrom pqdm.threads import pqdm\n\nfrom sklearn.model_selection import train_test_split\nfrom sklearn.preprocessing import StandardScaler\nfrom sklearn.metrics import mean_squared_error\n\nfrom astropy.stats import sigma_clip\nfrom scipy.signal import savgol_filter\nfrom scipy.optimize import minimize\nimport scipy.stats\nfrom sklearn.metrics import mean_absolute_error, mean_squared_error, r2_score\nfrom sklearn.preprocessing import StandardScaler\nimport pandas as pd\nimport numpy as np\nfrom sklearn.metrics import mean_absolute_error, mean_squared_error, r2_score\n\nROOT_PATH = \"/kaggle/input/ariel-data-challenge-2025\"\nMODE = \"test\"\n\n__t0 = time.perf_counter()\n\nclass Config:\n    # FEATURES = ['transit_depth', 'Rs', 'i']\n    FEATURES = ['transit_depth', 'Rs', 'i', 'P', 'sma']\n    # FEATURES = ['transit_depth', 'Rs', 'Ms', 'Ts', 'Mp', 'e', 'P', 'sma', 'i']\n    DATA_PATH = '/kaggle/input/ariel-data-challenge-2025'\n    DATASET = \"test\"\n    load_data = False  # Set to True to load from disk, False to generate\n    DEBUG = False\n    load_model = False\n\n    SCALE = 0.96\n    SIGMA = 0.00055\n    \n    CUT_INF = 39\n    CUT_SUP = 321\n    \n    SENSOR_CONFIG = {\n        \"AIRS-CH0\": {\n            \"raw_shape\": [11250, 32, 356],\n            \"calibrated_shape\": [1, 32, CUT_SUP - CUT_INF],\n            \"linear_corr_shape\": (6, 32, 356),\n            \"dt_pattern\": (0.1, 4.5), \n            \"binning\": 30\n        },\n        \"FGS1\": {\n            \"raw_shape\": [135000, 32, 32],\n            \"calibrated_shape\": [1, 32, 32],\n            \"linear_corr_shape\": (6, 32, 32),\n            \"dt_pattern\": (0.1, 0.1),\n            \"binning\": 30 * 12\n        }\n    }\n    \n    MODEL_PHASE_DETECTION_SLICE = slice(30, 140)\n    MODEL_OPTIMIZATION_DELTA = 11 # 9\n    MODEL_POLYNOMIAL_DEGREE = 3\n    \n    N_JOBS = 3\n\nclass ParticipantVisibleError(Exception):\n    pass\n\ndef score(\n    solution: pd.DataFrame,\n    submission: pd.DataFrame,\n    row_id_column_name: str,\n    naive_mean: float,\n    naive_sigma: float,\n    fsg_sigma_true: float = 1e-6,\n    airs_sigma_true: float = 1e-5,\n    fgs_weight: float = 1,\n) -> float:\n    \"\"\"\n    This is a Gaussian Log Likelihood based metric. For a submission, which contains the predicted mean (x_hat) and variance (x_hat_std),\n    we calculate the Gaussian Log-likelihood (GLL) value to the provided ground truth (x). We treat each pair of x_hat,\n    x_hat_std as a 1D gaussian, meaning there will be 283 1D gaussian distributions, hence 283 values for each test spectrum,\n    the GLL value for one spectrum is the sum of all of them.\n\n    Inputs:\n        - solution: Ground Truth spectra (from test set)\n            - shape: (nsamples, n_wavelengths)\n        - submission: Predicted spectra and errors (from participants)\n            - shape: (nsamples, n_wavelengths*2)\n        naive_mean: (float) mean from the train set.\n        naive_sigma: (float) standard deviation from the train set.\n        fsg_sigma_true: (float) standard deviation from the FSG1 instrument for the test set.\n        airs_sigma_true: (float) standard deviation from the AIRS instrument for the test set.\n        fgs_weight: (float) relative weight of the fgs channel\n    \"\"\"\n\n    del solution[row_id_column_name]\n    del submission[row_id_column_name]\n\n    if submission.min().min() < 0:\n        raise ParticipantVisibleError('Negative values in the submission')\n    for col in submission.columns:\n        if not pandas.api.types.is_numeric_dtype(submission[col]):\n            raise ParticipantVisibleError(f'Submission column {col} must be a number')\n\n    n_wavelengths = len(solution.columns)\n    if len(submission.columns) != n_wavelengths * 2:\n        raise ParticipantVisibleError('Wrong number of columns in the submission')\n\n    y_pred = submission.iloc[:, :n_wavelengths].values\n    # Set a non-zero minimum sigma pred to prevent division by zero errors.\n    sigma_pred = np.clip(submission.iloc[:, n_wavelengths:].values, a_min=10**-15, a_max=None)\n    sigma_true = np.append(\n        np.array(\n            [\n                fsg_sigma_true,\n            ]\n        ),\n        np.ones(n_wavelengths - 1) * airs_sigma_true,\n    )\n    y_true = solution.values\n\n    GLL_pred = scipy.stats.norm.logpdf(y_true, loc=y_pred, scale=sigma_pred)\n    GLL_true = scipy.stats.norm.logpdf(y_true, loc=y_true, scale=sigma_true * np.ones_like(y_true))\n    GLL_mean = scipy.stats.norm.logpdf(y_true, loc=naive_mean * np.ones_like(y_true), scale=naive_sigma * np.ones_like(y_true))\n\n    # normalise the score, right now it becomes a matrix instead of a scalar.\n    ind_scores = (GLL_pred - GLL_mean) / (GLL_true - GLL_mean)\n\n    weights = np.append(np.array([fgs_weight]), np.ones(len(solution.columns) - 1))\n    weights = weights * np.ones_like(ind_scores)\n    submit_score = np.average(ind_scores, weights=weights)\n    return float(np.clip(submit_score, 0.0, 1.0))\n\ndef _phase_detector_signal(signal, cfg):\n    sl = cfg.MODEL_PHASE_DETECTION_SLICE\n    min_idx = int(np.argmin(signal[sl])) + sl.start\n    s1 = signal[:min_idx]; s2 = signal[min_idx:]\n    if s1.size < 3 or s2.size < 3:\n        return 0, len(signal) - 1\n    g1 = np.gradient(s1); g1_max = np.max(g1) if np.size(g1) else 0.0\n    g2 = np.gradient(s2); g2_max = np.max(g2) if np.size(g2) else 0.0\n    if g1_max != 0: g1 /= g1_max\n    if g2_max != 0: g2 /= g2_max\n    phase1 = int(np.argmin(g1)); phase2 = int(np.argmax(g2)) + min_idx\n    return phase1, phase2\n\ndef estimate_sigma_fgs(preprocessed_data, cfg):\n    \"\"\"Возвращает вектор sigma_1 (для FGS1) длиной N_planets — мягкий множитель к cfg.SIGMA.\"\"\"\n    sig_rel = []\n    delta = cfg.MODEL_OPTIMIZATION_DELTA\n    eps = 1e-12\n    for single in preprocessed_data:\n        # фазы по AIRS белой кривой — так же, как в модели\n        air_white = savgol_filter(single[:, 1:].mean(axis=1), 20, 2)\n        p1, p2 = _phase_detector_signal(air_white, cfg)\n        p1 = max(delta, p1)\n        p2 = min(len(air_white) - delta - 1, p2)\n\n        fgs = single[:, 0]\n        oot = (fgs[: p1 - delta] if p1 - delta > 0 else np.empty(0, fgs.dtype))\n        if p2 + delta < fgs.size:\n            oot = np.concatenate([oot, fgs[p2 + delta :]])\n        inn = fgs[p1 + delta : max(p1 + delta, p2 - delta)]\n\n        if oot.size == 0 or inn.size == 0:\n            sig_rel.append(np.nan); continue\n\n        n_oot, n_in = len(oot), len(inn)\n        var_oot = np.nanvar(oot, ddof=1)\n        var_in  = np.nanvar(inn, ddof=1)\n        oot_mean = float(np.nanmean(oot)) if np.isfinite(np.nanmean(oot)) else float(np.nanmean(fgs))\n        # относительная неопределённость глубины (в тех же ед., что s)\n        sigma_rel = np.sqrt(var_oot / max(n_oot,1) + var_in / max(n_in,1)) / max(oot_mean, eps)\n        sig_rel.append(sigma_rel)\n\n    s = np.asarray(sig_rel, dtype=float)\n    mask = np.isfinite(s) & (s > 0)\n    med = float(np.nanmedian(s[mask])) if mask.any() else 1.0\n\n    # мягкий множитель: корень, и узкий клип, чтобы не рисковать\n    k = np.ones_like(s)\n    if med > 0 and np.isfinite(med):\n        k[mask] = np.sqrt(s[mask] / med)\n    k = np.clip(k, 0.8, 1.25)  # ±20–25% от базовой σ\n\n    return k * cfg.SIGMA\n\ndef estimate_sigma_air(preprocessed_data, cfg):\n    \"\"\"Возвращает вектор sigma_air длиной N_planets — мягкий множитель к cfg.SIGMA для всех AIRS-каналов.\"\"\"\n    sig_rel = []\n    delta = cfg.MODEL_OPTIMIZATION_DELTA\n    eps = 1e-12\n\n    for single in preprocessed_data:\n        # белая кривая AIRS на бинированных данных (после всех твоих весов по λ)\n        white = np.nanmean(single[:, 1:], axis=1)         # (n_bins,)\n        white_s = savgol_filter(white, 20, 2)             # для фаз\n\n        p1, p2 = _phase_detector_signal(white_s, cfg)\n        p1 = max(delta, p1)\n        p2 = min(len(white) - delta - 1, p2)\n\n        oot_left = white[: p1 - delta] if p1 - delta > 0 else np.empty(0, white.dtype)\n        oot_right = white[p2 + delta :] if (p2 + delta) < white.size else np.empty(0, white.dtype)\n        oot = np.concatenate([oot_left, oot_right]) if (oot_left.size + oot_right.size) else oot_left\n        inn = white[p1 + delta : max(p1 + delta, p2 - delta)]\n\n        if oot.size == 0 or inn.size == 0:\n            sig_rel.append(np.nan); continue\n\n        n_oot, n_in = len(oot), len(inn)\n        var_oot = np.nanvar(oot, ddof=1)\n        var_in  = np.nanvar(inn, ddof=1)\n        oot_mean = float(np.nanmean(oot)) if np.isfinite(np.nanmean(oot)) else float(np.nanmean(white))\n\n        sigma_rel = np.sqrt(var_oot / max(n_oot,1) + var_in / max(n_in,1)) / max(oot_mean, eps)\n        sig_rel.append(sigma_rel)\n\n    s = np.asarray(sig_rel, dtype=float)\n    mask = np.isfinite(s) & (s > 0)\n    med = float(np.nanmedian(s[mask])) if mask.any() else 1.0\n\n    # мягкий множитель вокруг медианы\n    k = np.ones_like(s)\n    if med > 0 and np.isfinite(med):\n        k[mask] = np.sqrt(s[mask] / med)\n    k = np.clip(k, 0.90, 1.20)  # ±10%–20%\n\n    return k * cfg.SIGMA\n\nclass SignalProcessor:\n    def __init__(self, config):\n        self.cfg = config\n        self.adc_info = pd.read_csv(f\"{self.cfg.DATA_PATH}/adc_info.csv\")\n        self.planet_ids = pd.read_csv(f'{self.cfg.DATA_PATH}/{self.cfg.DATASET}_star_info.csv', index_col='planet_id').index.astype(int)\n\n    def _apply_linear_corr(self, linear_corr, signal):\n\n        coeffs = np.flip(linear_corr, axis=0)      # shape: (D, X, Y), D — старшая степень сначала\n        x = signal.astype(np.float64, copy=False)  # считаем в float64 для стабильности\n        out = np.empty_like(x, dtype=np.float64)\n        out[...] = coeffs[0]  # broadcast (X,Y) -> (T,X,Y)\n        for k in range(1, coeffs.shape[0]):\n            np.multiply(out, x, out=out)  # in-place умножение\n            out += coeffs[k]              # broadcast (X,Y)\n\n        return out.astype(signal.dtype, copy=False)\n\n    def _calibrate_single_signal(self, planet_id, sensor):\n        \"\"\"\n        Калибровка single-node сигнала.\n        Политика масок: DEAD — маскируем, HOT — НЕ маскируем (оставляем в данных).\n        \"\"\"\n        sensor_cfg = self.cfg.SENSOR_CONFIG[sensor]\n    \n        # --- load ---\n        signal = pd.read_parquet(\n            f\"{self.cfg.DATA_PATH}/{self.cfg.DATASET}/{planet_id}/{sensor}_signal_0.parquet\"\n        ).to_numpy()\n        dark = pd.read_parquet(\n            f\"{self.cfg.DATA_PATH}/{self.cfg.DATASET}/{planet_id}/{sensor}_calibration_0/dark.parquet\"\n        ).to_numpy()\n        dead = pd.read_parquet(\n            f\"{self.cfg.DATA_PATH}/{self.cfg.DATASET}/{planet_id}/{sensor}_calibration_0/dead.parquet\"\n        ).to_numpy()\n        flat = pd.read_parquet(\n            f\"{self.cfg.DATA_PATH}/{self.cfg.DATASET}/{planet_id}/{sensor}_calibration_0/flat.parquet\"\n        ).to_numpy()\n        linear_corr = pd.read_parquet(\n            f\"{self.cfg.DATA_PATH}/{self.cfg.DATASET}/{planet_id}/{sensor}_calibration_0/linear_corr.parquet\"\n        ).values.astype(np.float64).reshape(sensor_cfg[\"linear_corr_shape\"])\n    \n        # --- reshape & ADC ---\n        signal = signal.reshape(sensor_cfg[\"raw_shape\"])\n        gain = self.adc_info[f\"{sensor}_adc_gain\"].iloc[0]\n        offset = self.adc_info[f\"{sensor}_adc_offset\"].iloc[0]\n        signal = signal / gain + offset  # сохраняем твою формулу\n    \n        # HOT только для мониторинга, не для маскирования\n        hot = sigma_clip(dark, sigma=5, maxiters=5).mask\n    \n        # --- crop per sensor ---\n        if sensor == \"AIRS-CH0\":\n            signal = signal[:, :, self.cfg.CUT_INF : self.cfg.CUT_SUP]\n            linear_corr = linear_corr[:, :, self.cfg.CUT_INF : self.cfg.CUT_SUP]\n            dark = dark[:, self.cfg.CUT_INF : self.cfg.CUT_SUP]\n            dead = dead[:, self.cfg.CUT_INF : self.cfg.CUT_SUP]\n            flat = flat[:, self.cfg.CUT_INF : self.cfg.CUT_SUP]\n            hot = hot[:, self.cfg.CUT_INF : self.cfg.CUT_SUP]  # только для логов\n    \n        if sensor == \"FGS1\":\n            y0, y1, x0, x1 = 10, 22, 10, 22\n            signal = signal[:, y0:y1, x0:x1]\n            dark   = dark[y0:y1, x0:x1]\n            dead   = dead[y0:y1, x0:x1]\n            flat   = flat[y0:y1, x0:x1]\n            linear_corr = linear_corr[:, y0:y1, x0:x1]\n            hot    = hot[y0:y1, x0:x1]  # только для логов\n    \n        # --- non-neg clamp before linearity corr (как у тебя) ---\n        np.maximum(signal, 0, out=signal)\n    \n        # --- linearity correction ---\n        if sensor == \"FGS1\":\n            signal = self._apply_linear_corr(linear_corr, signal)\n        elif sensor == \"AIRS-CH0\":\n            sl = (slice(None), slice(10, 22), slice(None))  # T, Y, λ\n            signal[sl] = self._apply_linear_corr(linear_corr[:, 10:22, :], signal[sl])\n        else:\n            signal = self._apply_linear_corr(linear_corr, signal)\n    \n        # --- dark subtraction с учётом паттерна интеграций ---\n        base_dt, increment = sensor_cfg[\"dt_pattern\"]\n        even_scale = base_dt\n        odd_scale  = base_dt + increment\n        signal[::2]  -= dark * even_scale\n        signal[1::2] -= dark * odd_scale\n    \n        # --- APPLY FLAT (HOT-KEEP: не включаем hot в маску!) ---\n        if sensor == \"FGS1\":\n            flat_roi = flat.astype(signal.dtype, copy=False).copy()      # (12,12)\n            bad = (dead) | ~np.isfinite(flat_roi) | (flat_roi == 0)      # ← ТОЛЬКО dead/invalid\n            flat_roi[bad] = np.nan\n            signal /= flat_roi\n    \n        elif sensor == \"AIRS-CH0\":\n            y0, y1 = 10, 22\n            flat_roi = flat[y0:y1, :].astype(signal.dtype, copy=False).copy()  # (12, λ)\n            bad = (dead[y0:y1, :]) | ~np.isfinite(flat_roi) | (flat_roi == 0)  # ← ТОЛЬКО dead/invalid\n            flat_roi[bad] = np.nan\n            signal[:, y0:y1, :] /= flat_roi\n    \n        else:\n            flat2 = flat.astype(signal.dtype, copy=False).copy()\n            bad2 = (dead) | ~np.isfinite(flat2) | (flat2 == 0)                  # ← ТОЛЬКО dead/invalid\n            flat2[bad2] = np.nan\n            signal /= flat2\n        # --- END FLAT ---\n    \n        # (опционально) логируем метрики hot/dead\n        if getattr(self.cfg, \"LOG_HOT_STATS\", False):\n            if not hasattr(self, \"stats\"):\n                self.stats = []\n            self.stats.append({\n                \"planet_id\": int(planet_id),\n                \"sensor\": sensor,\n                \"hot_frac\": float(np.mean(hot)),\n                \"dead_frac\": float(np.mean(dead)),\n            })\n    \n        return signal\n\n    def _preprocess_calibrated_signal(self, calibrated_signal, sensor):\n        sensor_cfg = self.cfg.SENSOR_CONFIG[sensor]\n        binning = sensor_cfg[\"binning\"]\n\n        if sensor == \"AIRS-CH0\":\n            signal_roi = calibrated_signal[:, 10:22, :]\n        elif sensor == \"FGS1\":\n            signal_roi = calibrated_signal[:, 10:22, 10:22]\n            signal_roi = signal_roi.reshape(signal_roi.shape[0], -1)\n        \n        mean_signal = np.nanmean(signal_roi, axis=1)\n\n        cds_signal = mean_signal[1::2] - mean_signal[0::2]\n\n        n_bins = cds_signal.shape[0] // binning\n        binned = np.array([\n            cds_signal[j*binning : (j+1)*binning].mean(axis=0) \n            for j in range(n_bins)\n        ])\n\n        if sensor == \"AIRS-CH0\":\n            q_lo = np.nanpercentile(binned, 5.0, axis=1, keepdims=True)    # (n_bins, 1)\n            q_hi = np.nanpercentile(binned, 95.0, axis=1, keepdims=True)   # (n_bins, 1)\n            np.clip(binned, q_lo, q_hi, out=binned)\n\n        if sensor == \"FGS1\":\n            binned = binned.reshape((binned.shape[0], 1))\n\n        if sensor == \"AIRS-CH0\":\n            var = np.nanvar(binned, axis=0, ddof=1)                 # (λ, )\n            med = np.nanmedian(var)\n            safe_var = np.where(~np.isfinite(var) | (var <= 0), med if (np.isfinite(med) and med > 0) else 1.0, var)\n            w = 1.0 / safe_var\n\n            lo, hi = np.nanpercentile(w, 5.0), np.nanpercentile(w, 95.0)\n            if np.isfinite(lo) and np.isfinite(hi) and lo < hi:\n                w = np.clip(w, lo, hi)\n\n            M = binned.shape[1]\n            s = np.nansum(w)\n            if np.isfinite(s) and s > 0:\n                w = w * (M / s)\n            else:\n                w = np.ones_like(w)\n\n            binned *= w[None, :]\n\n\n        return binned\n\n    def _process_planet_sensor(self, args):\n        planet_id, sensor = args['planet_id'], args['sensor']\n        calibrated = self._calibrate_single_signal(planet_id, sensor)\n        preprocessed = self._preprocess_calibrated_signal(calibrated, sensor)\n        return preprocessed\n\n    def process_all_data(self):\n        args_fgs1 = [dict(planet_id=planet_id, sensor=\"FGS1\") for planet_id in self.planet_ids]\n        preprocessed_fgs1 = pqdm(args_fgs1, self._process_planet_sensor, n_jobs=self.cfg.N_JOBS)\n\n        args_airs_ch0 = [dict(planet_id=planet_id, sensor=\"AIRS-CH0\") for planet_id in self.planet_ids]\n        preprocessed_airs_ch0 = pqdm(args_airs_ch0, self._process_planet_sensor, n_jobs=self.cfg.N_JOBS)\n\n        preprocessed_signal = np.concatenate(\n            [np.stack(preprocessed_fgs1), np.stack(preprocessed_airs_ch0)], axis=2\n        )\n        return preprocessed_signal\n    \n\nclass TransitModel:\n    def __init__(self, config):\n        self.cfg = config\n\n    def _phase_detector(self, signal):\n        search_slice = self.cfg.MODEL_PHASE_DETECTION_SLICE\n        min_index = np.argmin(signal[search_slice]) + search_slice.start\n        \n        signal1 = signal[:min_index]\n        signal2 = signal[min_index:]\n\n        grad1 = np.gradient(signal1)\n        grad1 /= grad1.max()\n        \n        grad2 = np.gradient(signal2)\n        grad2 /= grad2.max()\n\n        phase1 = np.argmin(grad1)\n        phase2 = np.argmax(grad2) + min_index\n\n        return phase1, phase2\n    \n    def _objective_function(self, s, signal, phase1, phase2):\n        delta = self.cfg.MODEL_OPTIMIZATION_DELTA\n        power = self.cfg.MODEL_POLYNOMIAL_DEGREE\n\n        if phase1 - delta <= 0 or phase2 + delta >= len(signal) or phase2 - delta - (phase1 + delta) < 5:\n            delta = 2\n\n        y = np.concatenate([\n            signal[: phase1 - delta],\n            signal[phase1 + delta : phase2 - delta] * (1 + s),\n            signal[phase2 + delta :]\n        ])\n        x = np.arange(len(y))\n\n        coeffs = np.polyfit(x, y, deg=power)\n        poly = np.poly1d(coeffs)\n        error = np.abs(poly(x) - y).mean()\n        \n        return error\n\n    def predict(self, single_preprocessed_signal):\n        signal_1d = single_preprocessed_signal[:, 1:].mean(axis=1)\n        signal_1d = savgol_filter(signal_1d, 23, 2)\n        \n        phase1, phase2 = self._phase_detector(signal_1d)\n\n        phase1 = max(self.cfg.MODEL_OPTIMIZATION_DELTA, phase1)\n        phase2 = min(len(signal_1d) - self.cfg.MODEL_OPTIMIZATION_DELTA - 1, phase2)    \n\n        result = minimize(\n            fun=self._objective_function,\n            x0=[0.0001],\n            args=(signal_1d, phase1, phase2),\n            method=\"Nelder-Mead\"\n        )\n        \n        return result.x[0]\n\n    def predict_all(self, preprocessed_signals):\n        predictions = [\n            self.predict(preprocessed_signal)\n            for preprocessed_signal in tqdm(preprocessed_signals)\n        ]\n        return np.array(predictions) * self.cfg.SCALE\n\nStarInfo = pd.read_csv(ROOT_PATH + f\"/{MODE}_star_info.csv\")\nStarInfo[\"planet_id\"] = StarInfo[\"planet_id\"].astype(int)\nPlanetIds = StarInfo[\"planet_id\"].tolist()\nStarInfo = StarInfo.set_index(\"planet_id\")\n\nclass SubmissionGenerator:\n    def __init__(self, config):\n        self.cfg = config\n        self.sample_submission = pd.read_csv(\"/kaggle/input/ariel-data-challenge-2025/sample_submission.csv\", index_col=\"planet_id\")\n\n    def create(self, predictions1, predictions, sigma_fgs=None, sigma_air=None):\n        planet_ids = self.sample_submission.index\n        n_mu = self.sample_submission.shape[1] // 2  # 283\n\n        preds = np.asarray(predictions, dtype=float).reshape(-1)\n        mu = np.tile(preds.reshape(-1, 1), (1, n_mu))\n        mu = np.clip(mu, 0, None)\n\n        sigmas = np.full_like(mu, self.cfg.SIGMA, dtype=float)\n        if sigma_fgs is not None:\n            sigma_fgs = np.asarray(sigma_fgs, dtype=float).reshape(-1)\n            sigmas[:, 0] = np.clip(sigma_fgs, 1e-6, 0.1)\n        if sigma_air is not None:\n            # sigma_air should already have the correct shape (num_planets, 282)\n            sigma_air_processed = np.asarray(sigma_air, dtype=float)\n            sigmas[:, 1:] = np.clip(sigma_air_processed, 1e-6, 0.1)\n\n        submission_df = pd.DataFrame(\n            np.concatenate([mu, sigmas], axis=1),\n            columns=self.sample_submission.columns,\n            index=planet_ids\n        )\n        submission_df.iloc[:, 0] = predictions\n        submission_df.iloc[:, 1:283] = predictions1\n        submission_df.to_csv(\"submission.csv\")\n        \n        return submission_df\n\nclass SEBlock(nn.Module):\n    def __init__(self, dim, reduction=16):\n        super().__init__()\n        self.fc1 = nn.Linear(dim, dim // reduction, bias=False)\n        self.fc2 = nn.Linear(dim // reduction, dim, bias=False)\n        self.act = nn.SiLU()   # Or nn.ReLU()\n\n    def forward(self, x):\n        # Compute channel-wise attention\n        w = x.mean(dim=0, keepdim=True)      # Global context (mean across batch)\n        w = self.act(self.fc1(w))\n        w = torch.sigmoid(self.fc2(w))\n        return x * w   # Rescale input\n\nclass AttentionBlock(nn.Module):\n    def __init__(self, dim, num_heads=4, dropout=0.1):\n        super().__init__()\n        self.attn = nn.MultiheadAttention(embed_dim=dim, num_heads=num_heads, batch_first=True)\n        self.norm = nn.LayerNorm(dim)\n        self.dropout = nn.Dropout(dropout)\n\n    def forward(self, x):\n        # Expect shape [batch, seq_len, dim], so expand if only [batch, dim]\n        if x.dim() == 2:\n            x = x.unsqueeze(1)   # -> [batch, 1, dim]\n        attn_out, _ = self.attn(x, x, x)\n        out = self.norm(x + self.dropout(attn_out))\n        return out.squeeze(1)    # Back to [batch, dim]\n\nclass ResidualBlock2(nn.Module):\n    def __init__(self, dim, p=0.2):\n        super().__init__()\n        self.fc1 = nn.Linear(dim, dim)\n        self.fc2 = nn.Linear(dim, dim)\n        self.relu = nn.ReLU()\n        self.dropout = nn.Dropout(p)\n\n    def forward(self, x):\n        identity = x\n        out = self.relu(self.fc1(x))\n        out = self.dropout(out)\n        out = self.fc2(out)\n        return self.relu(out + identity)\n\nclass ResNetMLP2(nn.Module):\n    def __init__(self, input_dim=3, hidden_dim=128, output_dim=282, num_blocks=3, dropout_rate=0.2):\n        super().__init__()\n        self.input_layer = nn.Linear(input_dim, hidden_dim)\n        self.blocks = nn.Sequential(*[ResidualBlock2(hidden_dim, p=dropout_rate) for _ in range(num_blocks)])\n        self.output_layer = nn.Linear(hidden_dim, output_dim)\n\n    def forward(self, x):\n        x = self.input_layer(x)\n        x = self.blocks(x)\n        x = self.output_layer(x)\n        return x\n        \nimport torch.distributions as D\n\n# This is now an MDN model, but with the same class name for easy integration\nclass ResNetMLP_WithUncertainty(nn.Module):\n    def __init__(self, input_dim, hidden_dim, output_dim, num_blocks, dropout_rate, n_gaussians=3):\n        super().__init__()\n        self.n_gaussians = n_gaussians\n        self.output_dim = output_dim\n        \n        # The main body of the network\n        self.blocks = nn.Sequential(\n            nn.Linear(input_dim, hidden_dim),\n            *[ResidualBlock2(hidden_dim, p=dropout_rate) for _ in range(num_blocks)]\n        )\n        \n        # Three separate heads for the mixture parameters\n        self.pi_head = nn.Linear(hidden_dim, output_dim * n_gaussians)\n        self.mu_head = nn.Linear(hidden_dim, output_dim * n_gaussians)\n        self.sigma_head = nn.Linear(hidden_dim, output_dim * n_gaussians)\n\n    def forward(self, x):\n        x = self.blocks(x)\n        \n        # Reshape and apply activations to the outputs\n        pi_logits = self.pi_head(x).view(-1, self.output_dim, self.n_gaussians)\n        pi = torch.softmax(pi_logits, dim=2)\n        \n        mu = self.mu_head(x).view(-1, self.output_dim, self.n_gaussians)\n        \n        log_sigma = self.sigma_head(x).view(-1, self.output_dim, self.n_gaussians)\n        sigma = torch.exp(log_sigma)\n        \n        return pi, mu, sigma\n\n# This is now the MDN loss function, with the same name for easy integration\ndef nll_loss(pi, mu, sigma, targets):\n    \"\"\"\n    Negative Log-Likelihood loss for a Gaussian Mixture Model.\n    \"\"\"\n    targets = targets.unsqueeze(2) # Reshape targets for mixture calculation\n    normal_dist = D.Normal(mu, sigma)\n    log_probs = normal_dist.log_prob(targets)\n    log_sum_exp = torch.logsumexp(torch.log(pi) + log_probs, dim=2)\n    return -torch.mean(log_sum_exp)\n    \ndef train_resnet_model(X_train, y_train, X_val, y_val, model_params, training_params):\n    \"\"\"\n    Trains the ResNetMLP_WithUncertainty model with weight decay and early stopping.\n    \"\"\"\n    device = torch.device(\"cuda\" if torch.cuda.is_available() else \"cpu\")\n    print(f\"Using device: {device}\")\n\n    train_dataset = TensorDataset(torch.tensor(X_train, dtype=torch.float64), torch.tensor(y_train, dtype=torch.float64))\n    val_dataset = TensorDataset(torch.tensor(X_val, dtype=torch.float64), torch.tensor(y_val, dtype=torch.float64))\n\n    train_loader = DataLoader(train_dataset, batch_size=training_params['batch_size'], shuffle=True)\n    val_loader = DataLoader(val_dataset, batch_size=training_params['batch_size'], shuffle=False)\n\n    model = ResNetMLP_WithUncertainty(**model_params).double().to(device)\n    # loss_function = nn.MSELoss()\n\n    # NEW: Add weight_decay to the optimizer from training_params\n\n    optimizer = torch.optim.Adam(\n        model.parameters(),\n        lr=training_params['learning_rate'],\n        weight_decay=training_params.get('weight_decay', 0) # Defaults to 0 if not provided\n    )\n\n    # Add scheduler\n    scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(\n        optimizer,\n        mode='min',        # we want to minimize validation loss\n        factor=0.5,        # shrink LR by half each time\n        patience=5,        # wait 5 epochs without improvement\n        verbose=True       # prints a message when LR is reduced\n    )\n\n    best_val_loss = float('inf')\n    epochs = training_params['epochs']\n\n    # NEW: Initialize variables for early stopping\n    patience = training_params.get('patience', 20) # Defaults to 20\n    epochs_no_improve = 0\n\n    # --- Training Loop ---\n    for epoch in range(epochs):\n        model.train()\n        running_train_loss = 0.0\n        for inputs, targets in train_loader:\n            inputs, targets = inputs.to(device), targets.to(device)\n            optimizer.zero_grad()\n\n            pi_outputs, mu_outputs, sigma_outputs = model(inputs)\n\n            # NEW: Use the custom NLL loss function\n            loss = nll_loss(pi_outputs, mu_outputs, sigma_outputs, targets)\n\n            loss.backward()\n            optimizer.step()\n            running_train_loss += loss.item() * inputs.size(0)\n        epoch_train_loss = running_train_loss / len(train_loader.dataset)\n\n        # --- Validation Loop ---\n        model.eval()\n        running_val_loss = 0.0\n        with torch.no_grad():\n            for inputs, targets in val_loader:\n                inputs, targets = inputs.to(device), targets.to(device)\n                \n                # FIX: Unpack the three MDN outputs\n                pi_outputs, mu_outputs, sigma_outputs = model(inputs)\n                \n                # FIX: Call the loss function with the correct arguments\n                loss = nll_loss(pi_outputs, mu_outputs, sigma_outputs, targets)\n        \n                running_val_loss += loss.item() * inputs.size(0)\n        epoch_val_loss = running_val_loss / len(val_loader.dataset)\n\n        print(f\"Epoch {epoch+1}/{epochs} | Train Loss: {epoch_train_loss:.6f} | Val Loss: {epoch_val_loss:.6f}\")\n\n        # Step the scheduler with the latest validation loss\n        scheduler.step(epoch_val_loss)\n\n        # NEW: Early stopping logic\n        if epoch_val_loss < best_val_loss:\n            best_val_loss = epoch_val_loss\n            torch.save(model.state_dict(), training_params['save_path'])\n            print(f\"Validation loss decreased. Model saved to {training_params['save_path']}\")\n            epochs_no_improve = 0 # Reset counter\n        else:\n            epochs_no_improve += 1\n            print(f\"Validation loss did not improve for {epochs_no_improve} epoch(s).\")\n        \n        if epochs_no_improve >= patience:\n            print(f\"Early stopping triggered after {patience} epochs with no improvement.\")\n            break # Exit the training loop\n\n    print(\"Training finished.\")\n\ndef train_new_model(hidden_dim=256, num_blocks=35, dropout_rate=0.1):\n    \"\"\"\n    This function handles the entire training pipeline:\n    1. Generates REAL transit_depth features from the training data.\n    2. Trains the ResNetMLP_WithUncertainty model.\n    3. Saves and returns the trained model.\n    \"\"\"\n    # === A. Generate REAL features for training ===\n    train_config = Config()\n    train_config.DATASET = \"train\"\n    if not Config.load_data:\n        # print(\"Generating real transit_depth features for training set...\")\n        # processor = SignalProcessor(train_config)\n        # preprocessed_train_data = processor.process_all_data()\n        # # Save preprocessed_train_data for reproducibility/debugging\n        # np.save(\"preprocessed_train_data.npy\", preprocessed_train_data)\n        print(\"Loading real transit_depth features for training set...\")\n        preprocessed_train_data = np.load('/kaggle/input/ariel-2025-result-2/results_sv43/preprocessed_train_data.npy')\n        print(f\"Preprocessed training data shape: {preprocessed_train_data.shape}\")\n        model_1d = TransitModel(train_config)\n        train_predictions_1d = model_1d.predict_all(preprocessed_train_data)\n        \n        train_star_info = pd.read_csv(f\"{train_config.DATA_PATH}/train_star_info.csv\")\n        train_targets = pd.read_csv(f\"{train_config.DATA_PATH}/train.csv\")\n        planet_ids = pd.read_csv(f'{train_config.DATA_PATH}/{train_config.DATASET}_star_info.csv', index_col='planet_id').index.astype(int)\n\n        predictions_df = pd.DataFrame({\n            \"planet_id\": planet_ids,\n            \"transit_depth\": train_predictions_1d\n        })\n        \n        # === B. Prepare X and y for training ===\n        input_df = pd.merge(predictions_df, train_star_info, on=\"planet_id\", how=\"left\")\n        input_df['scaled_depth'] = input_df['transit_depth'] / input_df['Rs']\n        # Using the simplified formula T_eq ∝ T_star * sqrt(R_star / a)\n        input_df['planet_temp'] = input_df['Ts'] * np.sqrt(input_df['Rs'] / (2 * input_df['sma']))\n        input_df['stellar_density'] = input_df['Ms'] / (input_df['Rs']**3)\n        input_df['stellar_gravity'] = input_df['Ms'] / (input_df['Rs']**2)\n\n        X = input_df[Config.FEATURES].values.astype(np.float64)\n        \n        full_train_df = pd.merge(input_df, train_targets, on=\"planet_id\", how=\"left\")\n        # Save full_train_df for reproducibility/debugging\n        full_train_df.to_csv(\"full_train_df.csv\", index=False)\n        target_cols = [f'wl_{i}' for i in range(2, 284)]\n        y = full_train_df[target_cols].values.astype(np.float64)\n        # Save X and y for reproducibility/debugging\n        np.save(\"X_train_features.npy\", X)\n        np.save(\"y_train_targets.npy\", y)\n    else:\n        # Load X and y if they exist (for reproducibility/debugging)\n        # Define the directory containing the files\n        data_dir = '/kaggle/input/ariel-2025-data/data_f9'\n        # Pick X and y file from the directory\n        X_file = os.path.join(data_dir, 'X_train_features.npy')\n        y_file = os.path.join(data_dir, 'y_train_targets.npy')\n        # Load preprocessed_train_data for sigma estimation in CV\n        preprocessed_train_data = np.load(os.path.join(data_dir, 'preprocessed_train_data.npy'))\n        # features = ['transit_depth', 'Rs', 'i']\n        if os.path.exists(X_file) and os.path.exists(y_file):\n            X = np.load(X_file)\n            y = np.load(y_file)\n        else:\n            raise FileNotFoundError(f\"Could not find {X_file} or {y_file}. Please check your data path or set Config.load_data = False.\")\n    # Standard scaling for X and y\n    scaler_X = StandardScaler()\n    scaler_y = StandardScaler()\n    X_scaled = scaler_X.fit_transform(X)\n    y_scaled = scaler_y.fit_transform(y)\n    # Print summary statistics for X and y\n    X_df = pd.DataFrame(X_scaled, columns=Config.FEATURES)\n    y_df = pd.DataFrame(y_scaled, columns=[f'wl_{i}' for i in range(2, 284)])\n    # print(\"X feature summary (scaled):\")\n    # print(X_df.describe())\n    # print(\"\\ny target summary (scaled):\")\n    # print(y_df.describe())\n\n    all_scores = []\n    all_mae = []\n    all_rmse = []\n    all_r2 = []\n    all_models = []\n    fold = 1\n    model_params = {\n        'input_dim': len(Config.FEATURES),  # Always use the current feature count\n        'hidden_dim': hidden_dim,\n        'output_dim': 282,\n        'num_blocks': num_blocks,\n        'dropout_rate': dropout_rate\n    }\n    training_params = {\n        'learning_rate': 0.001,\n        'epochs': 5 if Config.DEBUG else 500,\n        'batch_size': 64,\n        'save_path': 'best_model_airs_cv.pth',\n        'patience': 20,\n        'weight_decay': 1e-5\n    }\n    from sklearn.model_selection import StratifiedKFold # Import the new class\n    \n    # 1. Create a binned version of your target variable to stratify on.\n    # We can use the mean of the target spectrum for each planet as a representative value.\n    y_mean = y.mean(axis=1)\n    # pd.cut creates discrete bins from the continuous values\n    strata = pd.cut(y_mean, bins=10, labels=False, duplicates='drop')\n    \n    # 2. Use StratifiedKFold instead of KFold\n    n_splits = 10\n    # Use StratifiedKFold and pass the strata to the split() method\n    skf = StratifiedKFold(n_splits=n_splits, shuffle=True, random_state=42)\n    \n    # The CV loop now uses skf.split(X, strata)\n    for train_index, val_index in skf.split(X_scaled, strata):\n    \n    # for train_index, val_index in kf.split(X_scaled):\n        print(f\"\\n--- Fold {fold}/{n_splits} ---\")\n        X_train, X_val = X_scaled[train_index], X_scaled[val_index]\n        y_train, y_val = y_scaled[train_index], y_scaled[val_index]\n        \n        # 1. Get the preprocessed time-series data for the validation fold\n        preprocessed_data_val = preprocessed_train_data[val_index]\n        \n        # 2. Calculate the statistical sigmas using this validation data slice\n        # Make sure you have a config object available (e.g., train_config)\n        sigma_fgs_val = estimate_sigma_fgs(preprocessed_data_val, train_config)\n        sigma_air_val = estimate_sigma_air(preprocessed_data_val, train_config)\n        \n        # Update save_path for each fold to avoid overwriting\n        training_params['save_path'] = f'best_model_airs_cv_fold{fold}.pth'\n        train_resnet_model(X_train, y_train, X_val, y_val, model_params, training_params)\n        trained_model = ResNetMLP_WithUncertainty(**model_params).double()\n        model_path = training_params['save_path']\n        trained_model.load_state_dict(torch.load(model_path, map_location=torch.device('cpu')))\n        trained_model.eval()\n        # Pass the newly calculated sigmas to the evaluation function\n        val_score, mae, rmse, r2 = evaluate_model_on_val(\n            trained_model, \n            X_val, y_val, \n            X_train, y_train, \n            scaler_y, \n            sigma_fgs_val # Pass the calculated FGS sigma\n        )\n        print(f\"Fold {fold} - Validation Score: {val_score:.5f}, MAE: {mae:.5f}, RMSE: {rmse:.5f}, R2: {r2:.5f}\")\n        all_scores.append(val_score)\n        all_mae.append(mae)\n        all_rmse.append(rmse)\n        all_r2.append(r2)\n        all_models.append(trained_model)\n        fold += 1\n    avg_score = np.mean(all_scores)\n    avg_mae = np.mean(all_mae)\n    avg_rmse = np.mean(all_rmse)\n    avg_r2 = np.mean(all_r2)\n    print(f\"\\n==== Cross-Validation Results (averaged over {n_splits} folds) ====\")\n    print(f\"Average Validation Score (custom metric): {avg_score:.5f} ± {np.std(all_scores):.5f}\")\n    print(f\"Average MAE: {avg_mae:.5f} ± {np.std(all_mae):.5f}, Average RMSE: {avg_rmse:.5f} ± {np.std(all_rmse):.5f}, Average R2: {avg_r2:.5f} ± {np.std(all_r2):.5f}\")\n\n    # Save scalers for reproducibility and later use\n    import joblib\n    joblib.dump(scaler_X, 'scaler_X.joblib')\n    joblib.dump(scaler_y, 'scaler_y.joblib')\n    # Return models and scalers for use in prediction\n    return all_models, scaler_X, scaler_y, avg_score, avg_mae, avg_rmse, avg_r2\n\n\n# --- Hyperparameter sweep function ---\ndef sweep_hyperparameters():\n    \"\"\"\n    Sweeps through all combinations of hidden_dim, num_blocks, and dropout_rate,\n    calls train_new_model for each, and returns a DataFrame of results.\n    \"\"\"\n    import pandas as pd\n    import itertools\n    results = []\n    hidden_dim_list, num_blocks_list, dropout_rate_list = [64, 128, 256], [75, 80, 85], [0.1, 0.2, 0.3]\n    for hidden_dim, num_blocks, dropout_rate in itertools.product(hidden_dim_list, num_blocks_list, dropout_rate_list):\n        print(f\"\\n=== Training with hidden_dim={hidden_dim}, num_blocks={num_blocks}, dropout_rate={dropout_rate} ===\")\n        try:\n            _, _, _, mean_mae, mean_rmse, mean_r2 = train_new_model(hidden_dim=hidden_dim, num_blocks=num_blocks, dropout_rate=dropout_rate)\n            results.append({\n                'hidden_dim': hidden_dim,\n                'num_blocks': num_blocks,\n                'dropout_rate': dropout_rate,\n                'mean_mae': mean_mae,\n                'mean_rmse': mean_rmse,\n                'mean_r2': mean_r2\n            })\n            print(f\"Results so far: {results[-1]}\")\n        except Exception as e:\n            print(f\"Error for hidden_dim={hidden_dim}, num_blocks={num_blocks}, dropout_rate={dropout_rate}: {e}\")\n            results.append({\n                'hidden_dim': hidden_dim,\n                'num_blocks': num_blocks,\n                'dropout_rate': dropout_rate,\n                'mean_mae': None,\n                'mean_rmse': None,\n                'mean_r2': None,\n                'error': str(e)\n            })\n    df = pd.DataFrame(results)\n    df.to_csv(\"hyperparameter_sweep_results.csv\", index=False)\n    print(\"\\nHyperparameter sweep results:\")\n    print(df)\n    return df\n\n# ADD THIS HELPER FUNCTION to your script\ndef get_final_preds_from_mdn(pi, mu, sigma):\n    \"\"\"\n    Calculates the final mean and sigma from the MDN's mixture parameters.\n    \"\"\"\n    # Final mean (mu) is the weighted average of the component means\n    final_mu = torch.sum(pi * mu, dim=2)\n\n    # Final variance is calculated from the mixture components\n    variance_within = torch.sum(pi * (sigma**2), dim=2)\n    variance_between = torch.sum(pi * (mu**2), dim=2) - final_mu**2\n    final_sigma = torch.sqrt(variance_within + variance_between)\n    \n    return final_mu, final_sigma\n\n# REPLACE your old evaluate_model_on_val with this new version\ndef evaluate_model_on_val(trained_model, X_val, y_val, X_train, y_train, scaler_y, sigma_fgs_val):\n    \"\"\"\n    Evaluates the MDN-trained model on the validation set.\n    \"\"\"\n    trained_model.eval()\n    device = torch.device(\"cuda\" if torch.cuda.is_available() else \"cpu\")\n    trained_model.to(device)\n    X_val_tensor = torch.tensor(X_val, dtype=torch.float64).to(device)\n    \n    with torch.no_grad():\n        # 1. Get the three mixture parameter outputs from the MDN model\n        pi_val_scaled, mu_val_scaled, sigma_val_scaled = trained_model(X_val_tensor)\n        \n        # 2. Calculate the final mean and sigma using the helper function\n        y_pred_val_scaled, y_sigma_val_scaled = get_final_preds_from_mdn(\n            pi_val_scaled, mu_val_scaled, sigma_val_scaled\n        )\n\n        y_pred_val_scaled = y_pred_val_scaled.cpu().numpy()\n        y_sigma_val_scaled = y_sigma_val_scaled.cpu().numpy()\n\n    # 3. Inverse transform all predictions and targets to their original scale\n    y_pred_val = scaler_y.inverse_transform(y_pred_val_scaled)\n    y_val_orig = scaler_y.inverse_transform(y_val)\n    y_sigma_val = y_sigma_val_scaled * scaler_y.scale_\n\n    # 4. Prepare the 'solution' DataFrame\n    target_cols = [f'wl_{i}' for i in range(2, 284)] # For AIRS channels\n    val_df = pd.DataFrame(y_val_orig, columns=target_cols)\n    val_df['planet_id'] = np.arange(len(val_df))\n\n    # 5. Prepare the 'submission' DataFrame\n    pred_sigmas = np.zeros_like(y_pred_val)\n    pred_sigmas[:, 0] = sigma_fgs_val\n    pred_sigmas[:, 1:] = y_sigma_val[:, 1:]\n\n    pred_full = np.concatenate([y_pred_val, pred_sigmas], axis=1)\n    # ... The rest of the function is the same as your previous version ...\n    pred_full_cols = target_cols + [f'sigma_{i}' for i in range(2, 284)]\n    pred_full_df = pd.DataFrame(pred_full, columns=pred_full_cols)\n    pred_full_df['planet_id'] = np.arange(len(pred_full_df))\n    y_train_orig = scaler_y.inverse_transform(y_train)\n    naive_mean = y_train_orig.mean()\n    naive_sigma = y_train_orig.std()\n    val_score = score(\n        solution=val_df,\n        submission=pred_full_df,\n        row_id_column_name='planet_id',\n        naive_mean=naive_mean,\n        naive_sigma=naive_sigma\n    )\n    y_true_flat = y_val_orig.flatten()\n    y_pred_flat = y_pred_val.flatten()\n    mae = mean_absolute_error(y_true_flat, y_pred_flat)\n    rmse = mean_squared_error(y_true_flat, y_pred_flat, squared=False)\n    r2 = r2_score(y_true_flat, y_pred_flat)\n    \n    print(f\"Validation Score (custom metric): {val_score:.5f}\")\n    print(f\"MAE: {mae:.5f}, RMSE: {rmse:.5f}, R2: {r2:.5f}\")\n    \n    return val_score, mae, rmse, r2\n    \ndef load_cv_models_and_scalers(directory):\n    \"\"\"\n    Loads all cross-validation models and scalers from the specified directory.\n    Args:\n        directory (str): Path to the directory containing model and scaler files.\n    Returns:\n        all_models (list): List of loaded models.\n        scaler_X: Loaded X scaler.\n        scaler_y: Loaded y scaler.\n    \"\"\"\n    import os\n    import joblib\n    # Load scalers\n    scaler_X = joblib.load(os.path.join(directory, 'scaler_X.joblib'))\n    scaler_y = joblib.load(os.path.join(directory, 'scaler_y.joblib'))\n\n    # Load all CV models\n    all_models = []\n    model_params = {\n        'input_dim': len(Config.FEATURES),  # Always use the current feature count\n        'hidden_dim': 256,\n        'output_dim': 282,\n        'num_blocks': 35,\n        'dropout_rate': 0.1\n    }\n    for fold in range(1, 11):\n        model = ResNetMLP_WithUncertainty(**model_params).double()\n        model_path = os.path.join(directory, f'best_model_airs_cv_fold{fold}.pth')\n        model.load_state_dict(torch.load(model_path, map_location=torch.device('cpu')))\n        model.eval()\n        all_models.append(model)\n    return all_models, scaler_X, scaler_y\n\nconfig = Config()\n# sweep_hyperparameters()\nif config.load_model:\n    all_models, scaler_X, scaler_y = load_cv_models_and_scalers('/kaggle/input/ariel-2025-result-2/results_v39')\nelse:\n    all_models, scaler_X, scaler_y, avg_score, avg_mae, avg_rmse, avg_r2 = train_new_model(hidden_dim=128, num_blocks=30, dropout_rate=0.3)\nsignal_processor = SignalProcessor(config)\npreprocessed_data = signal_processor.process_all_data()\n\nmodel = TransitModel(config)\npredictions = model.predict_all(preprocessed_data)\nsigma_fgs_vec = estimate_sigma_fgs(preprocessed_data, config)\nsigma_air_vec = estimate_sigma_air(preprocessed_data, config)\n\npredictions_df = pd.DataFrame({\n    \"planet_id\": PlanetIds,\n    \"transit_depth\": predictions\n})\n\ninput_df = pd.merge(predictions_df, StarInfo, on=\"planet_id\", how=\"left\")\ninput_df['scaled_depth'] = input_df['transit_depth'] / input_df['Rs']\ninput_df['planet_temp'] = input_df['Ts'] * np.sqrt(input_df['Rs'] / (2 * input_df['sma']))\ninput_df['stellar_density'] = input_df['Ms'] / (input_df['Rs']**3)\ninput_df['stellar_gravity'] = input_df['Ms'] / (input_df['Rs']**2)\n\nX = input_df[Config.FEATURES].values.astype(np.float64)\nX_scaled = scaler_X.transform(X)\nX_tensor = torch.tensor(X_scaled, dtype=torch.float64)\n\nall_means_scaled = []\nall_sigmas_scaled = []\n\nwith torch.no_grad():\n    for model in all_models:\n        pi, mu, sigma = model(X_tensor)\n\n        # Final mean (mu) is the weighted average of the component means\n        final_mu = torch.sum(pi * mu, dim=2)\n    \n        # Final variance is calculated from the mixture components\n        variance_within = torch.sum(pi * (sigma**2), dim=2)\n        variance_between = torch.sum(pi * (mu**2), dim=2) - final_mu**2\n        final_sigma = torch.sqrt(variance_within + variance_between)\n        \n        all_means_scaled.append(final_mu.numpy())\n        all_sigmas_scaled.append(final_sigma.numpy())\n\n# Average the means from all models\nfinal_mean_scaled = np.mean(all_means_scaled, axis=0)\n\n# Combine the uncertainties (root mean square of the sigmas)\nfinal_sigma_scaled = np.sqrt(np.mean(np.square(all_sigmas_scaled), axis=0))\n\n\n# --- Inverse transform to get final values for submission ---\npredictions2 = scaler_y.inverse_transform(final_mean_scaled)\n# Remember to only scale the sigma, not shift it\npredictions2_sigma = final_sigma_scaled * scaler_y.scale_\n\n\n","metadata":{"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# Combine the predictions into a single array\nall_predictions = np.array([predictions1, predictions2])\n\n# Calculate the mean across the models (axis=0)\nfinal_predictions = np.mean(all_predictions, axis=0)\n\n# Combine the predictions into a single array\nall_sigma = np.array([predictions1_sigma, predictions2_sigma])\n\n# Calculate the mean across the models (axis=0)\nfinal_sigmans = np.mean(all_sigma, axis=0)","metadata":{"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# --- Create the final submission file ---\nsubmission_generator = SubmissionGenerator(config)\n\n# Use the new model-based sigma for the AIRS channels\nsubmission = submission_generator.create(\n    predictions1=final_predictions, \n    predictions=predictions,       # This is still the FGS1 prediction from TransitModel\n    sigma_fgs=sigma_fgs_vec,       # Still use the statistical sigma for FGS1\n    sigma_air=final_sigmans   # Use the new, more accurate sigma for AIRS\n)\n\nprint(\"Submission file created using NLL model uncertainty.\")\nsubmission.head()","metadata":{"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# submission_generator = SubmissionGenerator(config)\n# submission = submission_generator.create(predictions1, predictions, sigma_fgs=sigma_fgs_vec, sigma_air=sigma_air_vec)\n\n\n# __t1 = time.perf_counter()\n# elapsed = __t1 - __t0\n# print(f\"[TIMING] total runtime: {elapsed:.2f} s ({elapsed/60:.2f} min)\")\npd.read_csv(\"submission.csv\")","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-09-22T15:32:36.96302Z","iopub.execute_input":"2025-09-22T15:32:36.963352Z","iopub.status.idle":"2025-09-22T15:32:36.997744Z","shell.execute_reply.started":"2025-09-22T15:32:36.963325Z","shell.execute_reply":"2025-09-22T15:32:36.996696Z"}},"outputs":[],"execution_count":null}]}