app.py

# app.py
import os, json, glob
from typing import Any, Dict, List, Optional

import numpy as np
import tensorflow as tf
from fastapi import FastAPI, Request
from fastapi.middleware.cors import CORSMiddleware

# ----------------- CONFIG -----------------
DEFAULT_MODEL_CANDIDATES = ["best_model.h5", "best_model.keras"]
DEFAULT_IMPUTER_CANDIDATES = ["imputer.joblib", "imputer.pkl", "imputer.sav"]
DEFAULT_SCALER_CANDIDATES  = ["scaler.joblib",  "scaler.pkl",  "scaler.sav"]
DEFAULT_STATS_PATH = "means_std.json"

CLASSES = ["Top", "Mid-Top", "Mid", "Mid-Low", "Low"]  # index 0=Top ... 4=Low
APPLY_CORAL_MONOTONE = True  # nudge thresholds to be non-increasing before decode
# ------------------------------------------

HERE = os.path.dirname(os.path.abspath(__file__))


# ---------- utilities: robust file resolving & logging ----------
def resolve_first(*names: str) -> Optional[str]:
    """Return absolute path to the first existing file among provided names
    by checking HERE, CWD, then recursive matches."""
    for base in (HERE, os.getcwd()):
        for n in names:
            p = os.path.join(base, n)
            if os.path.isfile(p):
                return p
    # recursive fallback (handles subfolders)
    patterns: List[str] = []
    for n in names:
        patterns += [os.path.join(HERE, "**", n),
                     os.path.join(os.getcwd(), "**", n)]
    for pat in patterns:
        for p in glob.glob(pat, recursive=True):
            if os.path.isfile(p):
                return p
    return None


def describe_dir():
    try:
        print("CWD:", os.getcwd())
        print("Repo dir (HERE):", HERE)
        print("Repo listing:", os.listdir(HERE))
    except Exception as e:
        print("listdir error:", e)


def load_joblib(label: str, candidates: List[str]):
    import joblib
    print(f"Looking for {label} among: {candidates}")
    describe_dir()
    path = resolve_first(*candidates)
    if not path:
        print(f"⚠️  {label} not found.")
        return None
    try:
        print(f"Loading {label} from {path} ({os.path.getsize(path)} bytes)")
    except Exception:
        print(f"Loading {label} from {path}")
    try:
        return joblib.load(path)
    except Exception as e:
        print(f"⚠️  Failed to load {label}: {repr(e)}")
        return None


def load_model_robust() -> tf.keras.Model:
    print("Resolving model...")
    # env override supported
    env_model = os.getenv("MODEL_PATH")
    if env_model:
        candidates = [env_model]
    else:
        candidates = DEFAULT_MODEL_CANDIDATES
    path = resolve_first(*candidates)
    if not path:
        raise FileNotFoundError(f"Model file not found. Tried: {candidates}")
    print(f"Loading model from {path} ({os.path.getsize(path)} bytes)")
    # We don't need custom objects for inference; compile=False is safer
    return tf.keras.models.load_model(path, compile=False)


def load_means_std(stats_path: Optional[str]) -> Optional[Dict[str, Dict[str, float]]]:
    path = stats_path or os.getenv("STATS_PATH") or DEFAULT_STATS_PATH
    path = resolve_first(path) if path else None
    if not path:
        print("⚠️  means_std.json not found.")
        return None
    print(f"Loading means/std from {path} ({os.path.getsize(path)} bytes)")
    with open(path, "r") as f:
        return json.load(f)


# ---------- numeric coercion ----------
def coerce_float(val: Any) -> float:
    """Accepts numeric, or locale strings like '49.709,14' -> 49709.14"""
    if isinstance(val, (int, float)):
        return float(val)
    s = str(val).strip()
    if s == "":
        raise ValueError("empty")
    s = s.replace(" ", "")
    has_dot = "." in s
    has_comma = "," in s
    if has_dot and has_comma:
        last_dot = s.rfind(".")
        last_comma = s.rfind(",")
        if last_comma > last_dot:
            s = s.replace(".", "")
            s = s.replace(",", ".")
        else:
            s = s.replace(",", "")
    elif has_comma and not has_dot:
        s = s.replace(",", ".")
    return float(s)


def z_manual(val: Any, mean: float, sd: float) -> float:
    try:
        v = coerce_float(val)
    except Exception:
        return 0.0
    if not sd:
        return 0.0
    return (v - mean) / sd


# ---------- CORAL decoding ----------
def coral_probs_from_logits(logits_np: np.ndarray, monotone: bool = False) -> np.ndarray:
    """
    logits: (N, K-1) cumulative logits.
    If monotone=True, enforce non-increasing thresholds per sample before decode.
    """
    logits = np.asarray(logits_np, dtype=np.float32)

    if monotone:
        # clamp each row to be non-increasing: t1 >= t2 >= t3 >= ...
        # for Top=0 best to Low=4 worst, cumulative boundary logits
        for i in range(logits.shape[0]):
            row = logits[i]
            # make it non-increasing by cumulative minimum from left to right
            for j in range(1, row.shape[0]):
                if row[j] > row[j - 1]:
                    row[j] = row[j - 1]
            logits[i] = row

    sig = 1.0 / (1.0 + np.exp(-logits))  # sigmoid
    left = np.concatenate([np.ones((sig.shape[0], 1), dtype=np.float32), sig], axis=1)
    right = np.concatenate([sig, np.zeros((sig.shape[0], 1), dtype=np.float32)], axis=1)
    probs = np.clip(left - right, 1e-12, 1.0)
    return probs


# ---------- FastAPI app ----------
app = FastAPI(title="Static Fingerprint API", version="1.1.0")

app.add_middleware(
    CORSMiddleware,
    allow_origins=["*"],
    allow_credentials=False,
    allow_methods=["*"],
    allow_headers=["*"],
)

print("Loading model / imputer / scaler...")
model = load_model_robust()
imputer = load_joblib("imputer", DEFAULT_IMPUTER_CANDIDATES)
scaler  = load_joblib("scaler",  DEFAULT_SCALER_CANDIDATES)
stats   = load_means_std(os.getenv("STATS_PATH"))

# Feature order:
# Prefer scaler.feature_names_in_ if present (sklearn >=1.0),
# else imputer.feature_names_in_,
# else the order in means_std.json,
# else fail loudly.
if hasattr(scaler, "feature_names_in_"):
    FEATURES: List[str] = list(scaler.feature_names_in_)
    print("FEATURES from scaler.feature_names_in_")
elif hasattr(imputer, "feature_names_in_"):
    FEATURES = list(imputer.feature_names_in_)
    print("FEATURES from imputer.feature_names_in_")
elif isinstance(stats, dict):
    FEATURES = list(stats.keys())
    print("FEATURES from means_std.json order")
else:
    raise RuntimeError("Cannot determine feature order. Provide scaler/imputer with feature_names_in_ or a means_std.json.")

print("Feature order:", FEATURES)
print("Artifacts present:",
      {"imputer": imputer is not None, "scaler": scaler is not None, "stats": stats is not None})


@app.get("/")
def root():
    return {
        "message": "Static Fingerprint API is running.",
        "try": ["GET /health", "POST /predict", "POST /echo"],
    }


@app.get("/health")
def health():
    return {
        "status": "ok",
        "features": FEATURES,
        "classes": CLASSES,
        "artifacts": {
            "imputer": bool(imputer is not None),
            "scaler": bool(scaler is not None),
            "means_std": bool(stats is not None),
        },
    }


@app.post("/echo")
async def echo(req: Request):
    payload = await req.json()
    return {"received": payload}


def preprocess_payload_to_X(payload: Dict[str, Any]) -> Dict[str, Any]:
    """
    Returns dict with:
      - X: np.ndarray shape (1, n_features) ready for model
      - z_scores: dict feature -> z value (if available)
      - missing: list of features not provided
      - used: dict feature -> raw value used (after imputation)
    """
    missing: List[str] = []
    used_vals: List[float] = []
    z_scores: Dict[str, float] = {}
    used_raw: Dict[str, float] = {}

    # Build raw feature vector in correct order
    raw_vec: List[float] = []
    for f in FEATURES:
        if f in payload:
            v = coerce_float(payload[f])
        else:
            missing.append(f)
            v = np.nan  # let imputer handle it (median), or we'll fill below
        raw_vec.append(v)

    raw = np.array([raw_vec], dtype=np.float32)

    # Impute if available
    if imputer is not None:
        raw_imp = imputer.transform(raw)
    else:
        # If no imputer, simple median fill using means_std or zero
        raw_imp = raw.copy()
        for j, f in enumerate(FEATURES):
            if np.isnan(raw_imp[0, j]):
                if stats and f in stats:
                    raw_imp[0, j] = stats[f].get("mean", 0.0)
                else:
                    raw_imp[0, j] = 0.0

    # Scale if available
    if scaler is not None:
        X = scaler.transform(raw_imp).astype(np.float32)
        # we can still compute z-scores from scaler if it exposes scale_ and mean_
        if hasattr(scaler, "mean_") and hasattr(scaler, "scale_"):
            for j, f in enumerate(FEATURES):
                mu = float(scaler.mean_[j])
                sd = float(scaler.scale_[j])
                z = 0.0 if sd == 0 else (float(raw_imp[0, j]) - mu) / sd
                z_scores[f] = float(z)
    else:
        # manual z-score using means_std.json
        if not stats:
            raise RuntimeError("No scaler and no means_std.json — cannot standardize.")
        z_list: List[float] = []
        for j, f in enumerate(FEATURES):
            mu = float(stats[f]["mean"])
            sd = float(stats[f]["std"])
            z = z_manual(raw_imp[0, j], mu, sd)
            z_list.append(z)
            z_scores[f] = float(z)
        X = np.array([z_list], dtype=np.float32)

    # capture used raw values (after imputation)
    for j, f in enumerate(FEATURES):
        used_val = float(raw_imp[0, j])
        used_raw[f] = used_val
        used_vals.append(used_val)

    return {
        "X": X,
        "z_scores": z_scores,
        "missing": missing,
        "used": used_raw,
    }


@app.post("/predict")
async def predict(req: Request):
    payload = await req.json()
    if not isinstance(payload, dict):
        return {"error": "Expected a JSON object mapping feature -> value."}

    prep = preprocess_payload_to_X(payload)
    X: np.ndarray = prep["X"]

    raw = model.predict(X, verbose=0)

    # CORAL (K-1) vs softmax (K)
    debug: Dict[str, Any] = {"raw_shape": list(raw.shape)}
    if raw.ndim == 2 and raw.shape[1] == (len(CLASSES) - 1):
        decode_mode = "auto_coral_monotone" if APPLY_CORAL_MONOTONE else "auto_coral"
        probs = coral_probs_from_logits(raw, monotone=APPLY_CORAL_MONOTONE)[0]
    else:
        decode_mode = "auto_softmax"
        probs = raw[0]
        s = float(np.sum(probs))
        if s > 0:
            probs = probs / s
    debug["decode_mode"] = decode_mode
    debug["raw_first_row"] = [float(x) for x in np.array(raw[0]).ravel().tolist()]

    pred_idx = int(np.argmax(probs))

    return {
        "input_ok": (len(prep["missing"]) == 0),
        "missing": prep["missing"],
        "used_raw": prep["used"],        # values after imputation
        "z_scores": prep["z_scores"],    # standardized (from scaler or stats)
        "probabilities": {CLASSES[i]: float(probs[i]) for i in range(len(CLASSES))},
        "predicted_state": CLASSES[pred_idx],
        "debug": debug,
    }