inference.py

# SPDX-License-Identifier: Apache-2.0
# Copyright 2025 ISeeTheFuture

import os, json, argparse, warnings, joblib
import numpy as np
import pandas as pd
import torch
from typing import List

# === Features used at training time ===
FEATURE_COLS: List[str] = [
    "latitude","longitude","altitude",
    "accelerometer_x","accelerometer_y","accelerometer_z",
    "gyroscope_x","gyroscope_y","gyroscope_z",
    "compass"
]

# History window length for the model (uses the last 50 rows to predict the next row)
HIST_LEN_DEFAULT = 50                   # requires at least HIST_LEN+1 (=51) rows to produce one output

# === Default file locations (relative to the model repo root) ===
DEFAULT_WEIGHTS = "1753670088.7075965_lstm_corr.pth"
DEFAULT_SCALER_X = "scalers/1753670088.7075965_scaler_X.pkl"
DEFAULT_SCALER_Y = "scalers/1753670088.7075965_scaler_y.pkl"
DEFAULT_CONFIG   = "config.json"

# === Model class ===
from model import GPSCorrectionLSTM  # __init__(input_size, hidden_size=128, num_layers=2, dropout=0.3)

def load_config(cfg_path: str) -> dict:
    if os.path.exists(cfg_path):
        with open(cfg_path, "r") as f:
            return json.load(f)
    return {}

def build_model(input_size: int, cfg: dict) -> torch.nn.Module:
    """Instantiate the model with hyperparameters from config.json if available."""
    hidden_size = int(cfg.get("hidden_size", 128))
    num_layers  = int(cfg.get("num_layers", 2))
    dropout     = float(cfg.get("dropout", 0.3))
    # output_size was 2 in training (res_lat, res_lon); keep flexible if your class needs it.
    try:
        model = GPSCorrectionLSTM(input_size, hidden_size=hidden_size, num_layers=num_layers, dropout=dropout)
    except TypeError:
        model = GPSCorrectionLSTM(input_size, hidden_size=hidden_size, num_layers=num_layers)
    model.eval()
    return model

def load_scaler(path: str):
    """Load a Joblib scaler if present; otherwise continue without scaling."""
    if os.path.exists(path):
        return joblib.load(path)
    warnings.warn(f"[WARN] scaler not found: {path}. Proceeding without scaling.")
    return None

def load_df_from_csv(path: str) -> pd.DataFrame:
    """Load CSV, sort by timestamp if present, and validate feature columns."""
    df = pd.read_csv(path)
    if "timestamp" in df.columns:
        df = df.sort_values("timestamp")
    missing = [c for c in FEATURE_COLS if c not in df.columns]
    if missing:
        raise ValueError(f"CSV is missing columns: {missing}")
    return df.reset_index(drop=True)

def scale_window(X_win: np.ndarray, scaler_X):
    """Apply feature scaler to a single (T,F) window if provided."""
    if scaler_X is None:
        return X_win
    T, F = X_win.shape
    return scaler_X.transform(X_win.reshape(-1, F)).reshape(T, F)

def inverse_y(y: np.ndarray, scaler_y):
    """Inverse-transform a single (2,) or (3,) prediction if a target scaler is provided."""
    if scaler_y is None:
        return y
    return scaler_y.inverse_transform(y.reshape(1, -1)).reshape(-1)

def predict_next_residual(model: torch.nn.Module, X_win_tf: np.ndarray, device: str = "cpu") -> np.ndarray:
    """Predict next-step residual [res_lat, res_lon(, res_alt?)] from a (HIST_LEN,F) window."""
    x = torch.from_numpy(X_win_tf.astype(np.float32)).unsqueeze(0).to(device)  # (1, T, F)
    with torch.no_grad():
        y = model(x).squeeze(0).detach().cpu().numpy()
    return y  # shape: (2,) or (3,)

# python inference.py --csv samples/sample.csv
def main():
    ap = argparse.ArgumentParser(
        description="Rolling inference for next-step GNSS residuals using an LSTM model. "
                    "Uses the last HIST_LEN rows to predict the next row. "
                    "If the CSV has N rows and N >= HIST_LEN+1, this script outputs corrected coordinates "
                    "for rows [HIST_LEN ... N-1] (i.e., 51st to last)."
    )
    src = ap.add_mutually_exclusive_group(required=True)
    src.add_argument("--json", type=str, help="JSON string of shape [T, F]")
    src.add_argument("--json-file", type=str, help="Path to a JSON file (shape [T, F])")
    src.add_argument("--csv", type=str, help="Path to a CSV with columns: " + ",".join(FEATURE_COLS))

    ap.add_argument("--weights", default=DEFAULT_WEIGHTS, help="Model weights (state_dict or full model object).")
    ap.add_argument("--scaler-x", default=DEFAULT_SCALER_X, help="Feature scaler (Joblib).")
    ap.add_argument("--scaler-y", default=DEFAULT_SCALER_Y, help="Target scaler (Joblib).")
    ap.add_argument("--config",   default=DEFAULT_CONFIG,   help="Model hyperparameters (config.json).")
    ap.add_argument("--hist-len", type=int, default=HIST_LEN_DEFAULT, help="History window length used by the model (default: 50).")

    args = ap.parse_args()

    # 1) Load input
    if args.json:
        arr = np.asarray(json.loads(args.json), dtype=np.float32)
        timestamps = None
    elif args.json_file:
        with open(args.json_file, "r") as f:
            arr = np.asarray(json.load(f), dtype=np.float32)
        timestamps = None
    else:
        df = load_df_from_csv(args.csv)
        arr = df[FEATURE_COLS].to_numpy(dtype=np.float32)
        timestamps = df["timestamp"].to_numpy() if "timestamp" in df.columns else None

    T, F = arr.shape
    H = int(args.hist_len)
    if F != len(FEATURE_COLS):
        raise ValueError(f"Input feature dimension must be {len(FEATURE_COLS)}, got {F}.")

    # 2) Build & load model and scalers
    device = "cuda" if torch.cuda.is_available() else "cpu"
    cfg = load_config(args.config)
    model = build_model(input_size=F, cfg=cfg).to(device)

    state = torch.load(args.weights, map_location=device)
    try:
        model.load_state_dict(state)
    except Exception:
        model = state.to(device)
        model.eval()

    scaler_X = load_scaler(args.scaler_x)
    scaler_y = load_scaler(args.scaler_y)

    results = []

    # Rolling inference for indices i = H .. T-1
    # Each step uses arr[i-H : i] as input, and adds residual to noisy GNSS at i.
    for i in range(H, T):
        X_win = arr[i - H : i, :]                    # (H, F)
        X_win_tf = scale_window(X_win, scaler_X)
        y_pred = predict_next_residual(model, X_win_tf, device=device)  # (2,) or (3,)
        y_pred_deg = inverse_y(y_pred, scaler_y)

        res_lat = float(y_pred_deg[0])
        res_lon = float(y_pred_deg[1])
        # Noisy GNSS at step i (the "next" row after the window)
        noisy_lat = float(arr[i, 0])
        noisy_lon = float(arr[i, 1])

        out = {
            "index": int(i),  # 0-based row index in the input
            "noisy_next_lat_deg": noisy_lat,
            "noisy_next_lon_deg": noisy_lon,
            "pred_residual_lat_deg": res_lat,
            "pred_residual_lon_deg": res_lon,
            "corrected_next_lat_deg": noisy_lat + res_lat,
            "corrected_next_lon_deg": noisy_lon + res_lon,
        }
        # If model outputs altitude residual too
        if y_pred_deg.shape[0] >= 3:
            res_alt = float(y_pred_deg[2])
            noisy_alt = float(arr[i, 2])
            out.update({
                "noisy_next_alt_m": noisy_alt,
                "pred_residual_alt": res_alt,
                "corrected_next_alt_m": noisy_alt + res_alt
            })
        if timestamps is not None:
            out["timestamp"] = float(timestamps[i])
        results.append(out)

    print(json.dumps({
        "history_len": H,
        "total_rows": T,
        "outputs": results
    }, ensure_ascii=False, indent=2))

if __name__ == "__main__":
    main()