Spaces:
Sleeping
Sleeping
File size: 31,593 Bytes
a850728 b8edb00 6be85f7 261c687 b8edb00 a46e832 261c687 a850728 648b4ca 5f4bae5 a46e832 5f4bae5 a46e832 dbb37d9 a46e832 dbb37d9 a46e832 5f4bae5 dbb37d9 5f4bae5 a46e832 5f4bae5 a46e832 5f4bae5 a46e832 5f4bae5 dbb37d9 5f4bae5 dbb37d9 a46e832 5f4bae5 6363de7 e65910c a46e832 e65910c a46e832 5f4bae5 a46e832 5f4bae5 a46e832 a850728 a46e832 5f4bae5 a46e832 5f4bae5 a46e832 5f4bae5 a46e832 c7c0f5c a46e832 261c687 e65910c a46e832 a850728 a46e832 261c687 e65910c a46e832 a850728 a46e832 a850728 a46e832 a850728 a46e832 dbb37d9 e65910c a46e832 40255e4 a46e832 40255e4 a46e832 5f4bae5 a46e832 e65910c a46e832 e65910c 5f4bae5 a46e832 5f4bae5 a850728 a46e832 5f4bae5 a46e832 5f4bae5 a46e832 a850728 c7c0f5c a850728 c7c0f5c a850728 c71e704 465e1a1 c71e704 a850728 c71e704 a850728 c71e704 a850728 c71e704 a850728 c71e704 c7c0f5c a850728 c7c0f5c a46e832 c7c0f5c a46e832 c7c0f5c a46e832 c7c0f5c a46e832 5f4bae5 a46e832 a850728 5f4bae5 c7c0f5c a46e832 5f4bae5 c7c0f5c 5f4bae5 a46e832 04ccf8e a46e832 04ccf8e a850728 04ccf8e a46e832 04ccf8e a850728 04ccf8e a46e832 04ccf8e 130812d 04ccf8e a46e832 04ccf8e c7c0f5c 04ccf8e 8109a99 130812d 465e1a1 04ccf8e 130812d 1122e44 130812d 04ccf8e 130812d 04ccf8e 130812d 8109a99 130812d 7dc78bf 130812d 1122e44 130812d 1122e44 130812d 1122e44 130812d 1122e44 130812d 04ccf8e 130812d 04ccf8e 130812d 04ccf8e c7c0f5c a850728 04ccf8e 4b96a3d 04ccf8e 8109a99 04ccf8e 4b96a3d 04ccf8e 4b96a3d a46e832 4b96a3d 04ccf8e f92c118 6be85f7 dfe0810 f92c118 dfe0810 f92c118 6be85f7 f92c118 dfe0810 f92c118 dfe0810 f92c118 6be85f7 dfe0810 6be85f7 dfe0810 6be85f7 f92c118 dfe0810 6be85f7 f92c118 dfe0810 f92c118 6be85f7 f92c118 6be85f7 dfe0810 6be85f7 f92c118 dfe0810 f92c118 dfe0810 f92c118 6be85f7 f92c118 dfe0810 f92c118 dfe0810 f92c118 dfe0810 f92c118 dfe0810 f92c118 dfe0810 f92c118 dfe0810 f92c118 dfe0810 f92c118 dfe0810 f92c118 dfe0810 f92c118 6be85f7 f92c118 6be85f7 f92c118 dfe0810 6be85f7 f92c118 6be85f7 dfe0810 6be85f7 f92c118 dfe0810 f92c118 6be85f7 f92c118 dfe0810 f92c118 dfe0810 6be85f7 dfe0810 6be85f7 dfe0810 6be85f7 f92c118 dfe0810 f92c118 dfe0810 f92c118 dfe0810 f92c118 6be85f7 dfe0810 6be85f7 f92c118 dfe0810 f92c118 dfe0810 f92c118 dfe0810 f92c118 dfe0810 f92c118 340d80c 1b75554 340d80c |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 |
import os, json, io, traceback
from typing import Any, Dict, List, Optional
import pandas as pd
from sklearn.impute import SimpleImputer
from sklearn.preprocessing import StandardScaler
import numpy as np
import tensorflow as tf
from fastapi import FastAPI, Request
from fastapi.middleware.cors import CORSMiddleware
from fastapi.responses import JSONResponse
# ---------- SHAP optional import ----------
try:
import shap
SHAP_AVAILABLE = True
except ImportError:
SHAP_AVAILABLE = False
# ----------------- CONFIG -----------------
MODEL_PATH = os.getenv("MODEL_PATH", "best_model.h5")
STATS_PATH = os.getenv("STATS_PATH", "means_std.json")
IMPUTER_CANDIDATES = ["imputer.joblib", "imputer.pkl", "imputer.sav"]
SCALER_CANDIDATES = ["scaler.joblib", "scaler.pkl", "scaler.sav"]
CLASSES = ["Top", "Mid-Top", "Mid", "Mid-Low", "Low"]
# ⛔ DO NOT CHANGE: exact order used in training
FEATURES: List[str] = [
"autosuf_oper",
"improductiva",
"gastos_fin_over_avg_cart",
"_equity",
"grado_absorcion",
"_cartera_bruta",
"gastos_oper_over_ing_oper",
"cartera_vencida_ratio",
"roe_pre_tax",
"_assets",
"_liab",
"equity_over_assets",
"_margen_bruto",
"prov_over_cartera",
"gastos_oper_over_cart",
"ing_cartera_over_ing_total",
"debt_to_equity",
"prov_gasto_over_cart",
"cov_improductiva",
"rend_cart_over_avg_cart",
"roa_pre_tax",
]
# ------------------------------------------
# --------- helpers: I/O + numeric coercion ---------
def coerce_float(val: Any) -> float:
"""
Accepts numeric, or strings like:
"49.709,14" -> 49709.14
"49,709.14" -> 49709.14
"0,005" -> 0.005
"""
if isinstance(val, (int, float, np.number)):
return float(val)
s = str(val).strip()
if s == "":
raise ValueError("empty")
s = s.replace(" ", "")
has_dot, has_comma = "." in s, "," in s
if has_dot and has_comma:
# Decide decimal by last occurrence
if s.rfind(",") > s.rfind("."):
s = s.replace(".", "")
s = s.replace(",", ".")
else:
s = s.replace(",", "")
elif has_comma and not has_dot:
s = s.replace(",", ".")
# else leave as-is
return float(s)
def load_json(path: str) -> dict:
with open(path, "r") as f:
return json.load(f)
def load_joblib_if_exists(candidates: List[str]):
"""
Try loading a joblib/pickle artifact (imputer/scaler).
Returns (obj, path_str or None, error_str or None).
"""
for name in candidates:
p = os.path.join(os.getcwd(), name)
if os.path.isfile(p):
try:
# Import inside to avoid hard dependency if not used
import joblib # type: ignore
with open(p, "rb") as fh:
obj = joblib.load(fh)
return obj, p, None
except Exception as e:
return None, p, f"{type(e).__name__}({e})"
return None, None, None
# --------- model / artifacts load ---------
print("Loading model / imputer / scaler...")
# Model
model = tf.keras.models.load_model(MODEL_PATH, compile=False)
# Imputer
imputer, imputer_path, imputer_err = load_joblib_if_exists(IMPUTER_CANDIDATES)
if imputer_path and imputer_err:
print(f"⚠️ Failed to load imputer from {imputer_path}: {imputer_err}")
elif imputer:
print(f"Loaded imputer from {imputer_path}")
else:
print("⚠️ No imputer found — skipping median imputation.")
# Scaler
scaler, scaler_path, scaler_err = load_joblib_if_exists(SCALER_CANDIDATES)
if scaler_path and scaler_err:
print(f"⚠️ Failed to load scaler from {scaler_path}: {scaler_err}")
elif scaler:
print(f"Loaded scaler from {scaler_path}")
else:
print("⚠️ No scaler found — using manual z-scoring if stats are available.")
# Stats (means/std) for fallback manual z-score
stats: Dict[str, Dict[str, float]] = {}
if os.path.isfile(STATS_PATH):
stats = load_json(STATS_PATH)
print(f"Loaded means/std from {STATS_PATH}")
else:
print("⚠️ No means_std.json found — manual z-scoring will be unavailable if scaler missing.")
# --------- decoding for CORAL vs softmax ---------
def coral_probs_from_logits(logits_np: np.ndarray) -> np.ndarray:
"""
(N, K-1) logits -> (N, K) probabilities for CORAL ordinal output.
"""
logits = tf.convert_to_tensor(logits_np, dtype=tf.float32)
sig = tf.math.sigmoid(logits) # (N, K-1)
left = tf.concat([tf.ones_like(sig[:, :1]), sig], axis=1)
right = tf.concat([sig, tf.zeros_like(sig[:, :1])], axis=1)
probs = tf.clip_by_value(left - right, 1e-12, 1.0)
# normalize row-wise just in case
probs = probs / tf.reduce_sum(probs, axis=1, keepdims=True)
return probs.numpy()
def decode_logits(raw: np.ndarray) -> (np.ndarray, str):
"""
raw: (1, M) array
Returns (probs (K,), mode_str).
Detects CORAL (M=K-1) vs Softmax (M=K).
"""
if raw.ndim != 2:
raise ValueError(f"Unexpected raw shape {raw.shape}")
M = raw.shape[1]
K = len(CLASSES)
if M == K - 1:
# CORAL logits
probs = coral_probs_from_logits(raw)[0]
return probs, "auto_coral"
elif M == K:
# Softmax or unnormalized scores
row = raw[0]
exps = np.exp(row - np.max(row))
probs = exps / np.sum(exps)
return probs, "auto_softmax"
else:
# Fallback: normalize across whatever is there
row = raw[0]
s = float(np.sum(np.abs(row)))
probs = (row / s) if s > 0 else np.ones_like(row) / len(row)
return probs, f"fallback_M{M}_K{K}"
# --------- preprocessing pipeline ---------
def build_raw_vector(payload: Dict[str, Any]) -> np.ndarray:
"""
Build raw feature vector in exact training order.
Missing -> np.nan (imputer will handle if available).
Values coerced to float robustly.
"""
vals = []
for f in FEATURES:
if f in payload:
try:
vals.append(coerce_float(payload[f]))
except Exception:
vals.append(np.nan)
else:
vals.append(np.nan)
return np.array(vals, dtype=np.float32)
def apply_imputer_if_any(x: np.ndarray) -> np.ndarray:
if imputer is not None:
# imputer expects 2D
return imputer.transform(x.reshape(1, -1)).astype(np.float32)[0]
# fallback: replace NaNs with feature means from stats if available, else 0
out = x.copy()
for i, f in enumerate(FEATURES):
if np.isnan(out[i]):
if f in stats and "mean" in stats[f]:
out[i] = float(stats[f]["mean"])
else:
out[i] = 0.0
return out
def apply_scaling_or_stats(raw_vec: np.ndarray) -> (np.ndarray, Dict[str, float], str):
"""
Returns (z_vec, z_detail_dict, mode_str)
- If scaler present: scaler.transform
- Else: manual (x-mean)/std using stats
"""
if scaler is not None:
z = scaler.transform(raw_vec.reshape(1, -1)).astype(np.float32)[0]
z_detail = {f: float(z[i]) for i, f in enumerate(FEATURES)}
return z, z_detail, "sklearn_scaler"
else:
z = np.zeros_like(raw_vec, dtype=np.float32)
z_detail: Dict[str, float] = {}
for i, f in enumerate(FEATURES):
mean = stats.get(f, {}).get("mean", 0.0)
sd = stats.get(f, {}).get("std", 1.0)
if not sd:
sd = 1.0
z[i] = (raw_vec[i] - mean) / sd
z_detail[f] = float(z[i])
return z, z_detail, "manual_stats"
# --------- SHAP model wrapper & explainer ---------
def model_proba_from_z(z_batch_np: np.ndarray) -> np.ndarray:
"""
Wrapper for SHAP: takes (N, n_features) in z-space and returns (N, K) probabilities.
"""
raw = model.predict(z_batch_np, verbose=0)
if raw.ndim != 2:
raise ValueError(f"Unexpected raw shape from model: {raw.shape}")
N, M = raw.shape
K = len(CLASSES)
if M == K - 1:
# CORAL
probs = coral_probs_from_logits(raw) # (N, K)
elif M == K:
# Softmax or scores
exps = np.exp(raw - np.max(raw, axis=1, keepdims=True))
probs = exps / np.sum(exps, axis=1, keepdims=True)
else:
# Fallback normalize
s = np.sum(np.abs(raw), axis=1, keepdims=True)
probs = np.divide(raw, s, out=np.ones_like(raw) / max(M, 1), where=(s > 0))
return probs
EXPLAINER = None
if SHAP_AVAILABLE:
try:
# Background: 50 "average" institutions at z=0
BACKGROUND_Z = np.zeros((50, len(FEATURES)), dtype=np.float32)
EXPLAINER = shap.KernelExplainer(model_proba_from_z, BACKGROUND_Z)
print("SHAP KernelExplainer initialized.")
except Exception as e:
EXPLAINER = None
print("⚠️ Failed to initialize SHAP explainer:", repr(e))
else:
print("SHAP not installed; explanations disabled.")
# ----------------- FastAPI -----------------
app = FastAPI(title="Static Fingerprint API", version="1.2.0")
app.add_middleware(
CORSMiddleware,
allow_origins=["*"],
allow_credentials=False,
allow_methods=["*"],
allow_headers=["*"],
)
@app.get("/")
def root():
return {
"message": "Static Fingerprint API is running.",
"try": ["GET /health", "POST /predict", "POST /debug/z"],
}
@app.get("/health")
def health():
stats_keys = []
try:
if os.path.isfile(STATS_PATH):
stats_keys = list(load_json(STATS_PATH).keys())
except Exception:
pass
return {
"status": "ok",
"classes": CLASSES,
"features_training_order": FEATURES,
"features_in_means_std": stats_keys,
"model_file": MODEL_PATH,
"imputer": bool(imputer),
"scaler": bool(scaler),
"stats_available": bool(stats),
"shap_available": bool(EXPLAINER is not None),
}
@app.post("/debug/z")
async def debug_z(req: Request):
try:
payload = await req.json()
if not isinstance(payload, dict):
return JSONResponse(status_code=400, content={"error": "Expected JSON object"})
raw = build_raw_vector(payload)
raw_imp = apply_imputer_if_any(raw)
z, z_detail, mode = apply_scaling_or_stats(raw_imp)
rows = []
for i, f in enumerate(FEATURES):
rows.append({
"feature": f,
"input_value": None if np.isnan(raw[i]) else float(raw[i]),
"imputed_value": float(raw_imp[i]),
"z": float(z[i]),
"mean": stats.get(f, {}).get("mean", None),
"std": stats.get(f, {}).get("std", None),
})
return {"preprocess_mode": mode, "rows": rows}
except Exception as e:
return JSONResponse(status_code=500, content={"error": str(e), "trace": traceback.format_exc()})
@app.post("/predict")
async def predict(req: Request):
"""
Body: JSON object mapping feature -> numeric value (strings with commas/points ok).
Missing features are imputed if imputer present; else filled with means (if stats) or 0.
Returns:
- probabilities per state
- predicted_state
- z_scores (per feature, after imputation & scaling pipeline)
- shap: per-class explanations if available
"""
try:
payload = await req.json()
if not isinstance(payload, dict):
return JSONResponse(
status_code=400,
content={"error": "Expected JSON object"},
)
# ---------- 1) Preprocess: raw -> imputed -> z ----------
raw_vec = build_raw_vector(payload) # (21,) may contain NaNs
raw_imp = apply_imputer_if_any(raw_vec) # impute missing
z_vec, z_detail, z_mode = apply_scaling_or_stats(raw_imp)
# ---------- 2) Model prediction ----------
X = z_vec.reshape(1, -1).astype(np.float32)
raw_logits = model.predict(X, verbose=0)
probs, decode_mode = decode_logits(raw_logits)
pred_idx = int(np.argmax(probs))
probs_dict = {CLASSES[i]: float(probs[i]) for i in range(len(CLASSES))}
missing = [f for i, f in enumerate(FEATURES) if np.isnan(raw_vec[i])]
# ---------- 3) SHAP explanations (all classes) ----------
shap_block: Dict[str, Any] = {"available": False}
if EXPLAINER is not None and SHAP_AVAILABLE:
try:
X_z = z_vec.reshape(1, -1).astype(np.float32)
shap_vals = EXPLAINER.shap_values(X_z, nsamples=50)
all_classes: Dict[str, Dict[str, float]] = {}
# ---------- CASE 1: SHAP returns list (usual multi-class) ----------
if isinstance(shap_vals, list):
for k, class_name in enumerate(CLASSES):
if k >= len(shap_vals):
continue
arr = np.array(shap_vals[k], dtype=float) # shape (N, D) or (D,)
# reduce to a 1D (D,) vector for the first sample
if arr.ndim == 2 and arr.shape[0] >= 1 and arr.shape[1] == len(FEATURES):
vec = arr[0, :]
elif arr.ndim == 1 and arr.shape[0] == len(FEATURES):
vec = arr
else:
# shape we don't know how to handle for this class
continue
all_classes[class_name] = {
FEATURES[i]: float(vec[i]) for i in range(len(FEATURES))
}
if all_classes:
shap_block = {
"available": True,
"mode": "per_class",
"explained_classes": list(all_classes.keys()),
"all_classes": all_classes,
}
else:
shap_block = {
"available": False,
"error": "No per-class SHAP vectors matched expected shape.",
}
# ---------- CASE 2: SHAP returns a numpy array ----------
else:
arr = np.array(shap_vals, dtype=float)
# (1, D, K) <-- THIS IS YOUR (1, 21, 5) CASE
if (
arr.ndim == 3
and arr.shape[0] == 1
and arr.shape[1] == len(FEATURES)
and arr.shape[2] == len(CLASSES)
):
# first sample, loop over classes on last axis
for k, class_name in enumerate(CLASSES):
vec = arr[0, :, k] # (D,)
all_classes[class_name] = {
FEATURES[i]: float(vec[i]) for i in range(len(FEATURES))
}
shap_block = {
"available": True,
"mode": "per_class",
"explained_classes": list(all_classes.keys()),
"all_classes": all_classes,
}
# (1, K, D)
elif (
arr.ndim == 3
and arr.shape[0] == 1
and arr.shape[1] == len(CLASSES)
and arr.shape[2] == len(FEATURES)
):
for k, class_name in enumerate(CLASSES):
vec = arr[0, k, :] # (D,)
all_classes[class_name] = {
FEATURES[i]: float(vec[i]) for i in range(len(FEATURES))
}
shap_block = {
"available": True,
"mode": "per_class",
"explained_classes": list(all_classes.keys()),
"all_classes": all_classes,
}
# (K, D)
elif (
arr.ndim == 2
and arr.shape[0] == len(CLASSES)
and arr.shape[1] == len(FEATURES)
):
for k, class_name in enumerate(CLASSES):
vec = arr[k, :] # (D,)
all_classes[class_name] = {
FEATURES[i]: float(vec[i]) for i in range(len(FEATURES))
}
shap_block = {
"available": True,
"mode": "per_class",
"explained_classes": list(all_classes.keys()),
"all_classes": all_classes,
}
# Single-vector fallback: (1, D) or (D,)
elif arr.ndim == 2 and arr.shape[0] == 1 and arr.shape[1] == len(FEATURES):
vec = arr[0, :] # (D,)
shap_block = {
"available": True,
"mode": "single_class",
"explained_class": CLASSES[pred_idx],
"values": {
FEATURES[i]: float(vec[i]) for i in range(len(FEATURES))
},
}
elif arr.ndim == 1 and arr.shape[0] == len(FEATURES):
vec = arr # (D,)
shap_block = {
"available": True,
"mode": "single_class",
"explained_class": CLASSES[pred_idx],
"values": {
FEATURES[i]: float(vec[i]) for i in range(len(FEATURES))
},
}
else:
shap_block = {
"available": False,
"error": f"Unexpected SHAP array shape {arr.shape}",
}
except Exception as e:
shap_block = {
"available": False,
"error": str(e),
"trace": traceback.format_exc(),
}
# ---------- 4) Build response ----------
return {
"input_ok": (len(missing) == 0),
"missing": missing,
"preprocess": {
"imputer": bool(imputer),
"scaler": bool(scaler),
"z_mode": z_mode,
},
"z_scores": z_detail, # per feature
"probabilities": probs_dict, # per state
"predicted_state": CLASSES[pred_idx],
"shap": shap_block,
"debug": {
"raw_shape": list(raw_logits.shape),
"decode_mode": decode_mode,
"raw_first_row": [float(v) for v in raw_logits[0]],
},
}
except Exception as e:
return JSONResponse(
status_code=500,
content={"error": str(e), "trace": traceback.format_exc()},
)
# ============================================================
# CORAL ORDINAL HELPERS (from training script)
# (we do NOT redefine coral_probs_from_logits here to avoid
# clashing with the one already used by decode_logits)
# ============================================================
def to_cumulative_targets_tf(y_true_int, K_):
"""
y_true_int: (N,) integer targets 0..K-1
returns (N, K_-1) with t_k = 1[y >= k], k = 1..K-1
"""
y = tf.reshape(y_true_int, [-1])
y = tf.cast(y, tf.int32)
thresholds = tf.range(1, K_, dtype=tf.int32)
T = tf.cast(tf.greater_equal(y[:, None], thresholds[None, :]), tf.float32)
return T
def coral_loss_tf(y_true, logits):
"""
CORAL ordinal loss implemented in TF:
y_true: (N,) or (N,1) with integer labels 0..K-1
logits: (N, K-1)
"""
y_true = tf.reshape(y_true, [-1])
y_true = tf.cast(y_true, tf.int32)
T = to_cumulative_targets_tf(y_true, len(CLASSES)) # (N, K-1)
bce = tf.nn.sigmoid_cross_entropy_with_logits(labels=T, logits=logits)
return tf.reduce_mean(tf.reduce_sum(bce, axis=1))
# ---------- TF helper (pure TF CORAL probs) ----------
def _coral_probs_from_logits_tf(logits_tf: tf.Tensor) -> tf.Tensor:
"""
Pure-TF version of CORAL probability transform, used in metric.
logits_tf: (N, K-1)
returns (N, K) probabilities
"""
sig = tf.math.sigmoid(logits_tf)
left = tf.concat([tf.ones_like(sig[:, :1]), sig], axis=1)
right = tf.concat([sig, tf.zeros_like(sig[:, :1])], axis=1)
probs = tf.clip_by_value(left - right, 1e-12, 1.0)
return probs
@tf.function
def ordinal_accuracy_metric(y_true, y_pred_logits):
"""
Exact class accuracy for CORAL outputs (same idea as training script).
"""
y_true = tf.reshape(y_true, [-1])
y_true = tf.cast(y_true, tf.int32)
probs = _coral_probs_from_logits_tf(y_pred_logits)
y_pred = tf.argmax(probs, axis=1, output_type=tf.int32)
return tf.reduce_mean(tf.cast(tf.equal(y_true, y_pred), tf.float32))
# ============================================================
# IMPORTS FOR RETRAINING / DATA MGMT
# (Ok to import here; Python allows imports anywhere in file)
# ============================================================
import pandas as pd
from sklearn.impute import SimpleImputer
from sklearn.preprocessing import StandardScaler
# ============================================================
# LETTER → 5-CLASS GROUP MAPPING (same logic as training code)
# ============================================================
def letter_to_group(letter: str):
"""
Converts raw rating letters (AAA, A-, BBB+, BB-, etc.)
into the 5 ordinal groups used by the model:
Top, Mid-Top, Mid, Mid-Low, Low
"""
if letter is None:
return None
s = str(letter).strip().upper()
if s == "":
return None
# Normalise duals like "AA / AA+" by taking the stronger one
s_clean = s.replace(" ", "")
if "/" in s_clean:
order = [
"E","D","C-","C","C+",
"B-","B","B+","BB-","BB","BB+",
"BBB-","BBB","BBB+",
"A-","A","A+",
"AA-","AA","AA+",
"AAA-","AAA"
]
parts = [p for p in s_clean.split("/") if p]
idxs = [order.index(p) for p in parts if p in order]
if idxs:
s = order[max(idxs)] # stronger (higher index)
else:
s = parts[0]
# Group boundaries (as in your training script)
g1 = {"AAA","AAA-","AA+","AA"} # Top
g2 = {"AA-","A+","A","A-"} # Mid-Top
g3 = {"BBB+","BBB","BBB-","BB+"} # Mid
g4 = {"BB","BB-","B+","B","B-"} # Mid-Low
g5 = {"C+","C","C-","D","E"} # Low
if s in g1: return "Top"
if s in g2: return "Mid-Top"
if s in g3: return "Mid"
if s in g4: return "Mid-Low"
if s in g5: return "Low"
return None
# ============================================================
# RECREATE MODEL FROM BEST HYPERPARAMETERS
# ============================================================
def build_model_from_hparams(hp: dict):
"""
Rebuilds the CORAL DNN with the same structure & hyperparameters
as in your training script.
"""
inputs = tf.keras.Input(shape=(len(FEATURES),))
x = inputs
n_hidden = hp["n_hidden"]
use_bn = hp["batchnorm"]
act = hp["activation"]
l2_reg = hp["l2"]
for i in range(1, n_hidden + 1):
units = hp[f"units_{i}"]
drop = hp[f"dropout_{i}"]
x = tf.keras.layers.Dense(
units,
activation=act,
kernel_regularizer=tf.keras.regularizers.l2(l2_reg)
)(x)
if use_bn:
x = tf.keras.layers.BatchNormalization()(x)
if drop > 0:
x = tf.keras.layers.Dropout(drop)(x)
# CORAL output: K-1 logits (K = len(CLASSES))
outputs = tf.keras.layers.Dense(len(CLASSES) - 1, activation=None)(x)
model = tf.keras.Model(inputs, outputs)
model.compile(
optimizer=tf.keras.optimizers.Adam(learning_rate=hp["lr"]),
loss=coral_loss_tf,
metrics=[ordinal_accuracy_metric],
)
return model
# ============================================================
# RETRAINING LOGIC + DATASET MGMT
# ============================================================
FINGERPRINT_CSV = "fingerprints_db.csv" # master DB file
BEST_HP_JSON = "best_params_and_metrics.json" # hyperparams JSON
def load_best_hparams():
"""
Loads best hyperparameters from your tuning JSON.
Expects JSON to contain key "best_hyperparams".
"""
with open(BEST_HP_JSON, "r") as f:
js = json.load(f)
return js["best_hyperparams"]
def load_fingerprint_dataset():
"""
Loads the full fingerprint DB from FINGERPRINT_CSV.
Expected columns (at minimum):
- QTR
- COMPANY
- Supervisor
- RATING_RAW
- 21 ratio features named exactly as in FEATURES
- rating_score (can be ignored for training)
We:
- derive RATING_GROUP (Top/Mid-Top/...) from RATING_RAW if missing
- drop rows with RATING_GROUP = NaN
- impute missing feature values with median
- scale with StandardScaler
"""
df = pd.read_csv(FINGERPRINT_CSV)
# Derive 5-class group if not already present
if "RATING_GROUP" not in df.columns:
df["RATING_GROUP"] = df["RATING_RAW"].apply(letter_to_group)
df = df[df["RATING_GROUP"].notna()].copy()
# y labels 0..4
class_to_id = {c: i for i, c in enumerate(CLASSES)}
y = df["RATING_GROUP"].map(class_to_id).astype("int32").to_numpy()
# X features
X_raw = df[FEATURES].to_numpy().astype("float32")
# Fit fresh imputer + scaler on full dataset
imp = SimpleImputer(strategy="median")
sc = StandardScaler()
X_imp = imp.fit_transform(X_raw)
X_sc = sc.fit_transform(X_imp).astype("float32")
return X_sc, y, imp, sc
def retrain_model():
"""
Retrains the model on the current fingerprints_db.csv
using the fixed best hyperparameters.
- Rebuilds the model
- Fits on full (X_sc, y)
- Updates global model/imputer/scaler
- Rebuilds SHAP explainer to stay in sync
"""
print(">>> RETRAIN: loading dataset")
hp = load_best_hparams()
X, y, imp, sc = load_fingerprint_dataset()
print(">>> RETRAIN: building model from best hparams")
model_new = build_model_from_hparams(hp)
print(">>> RETRAIN: fitting on fingerprint DB")
es = tf.keras.callbacks.EarlyStopping(
monitor="loss",
patience=15,
restore_best_weights=True,
verbose=1
)
model_new.fit(
X, y,
epochs=150,
batch_size=128,
callbacks=[es],
verbose=1,
)
# Update global model + preprocessors used by /predict
global model, imputer, scaler
model = model_new
imputer = imp
scaler = sc
# Rebuild SHAP explainer so explanations match new model
global EXPLAINER
if SHAP_AVAILABLE:
try:
BACKGROUND_Z = np.zeros((50, len(FEATURES)), dtype=np.float32)
EXPLAINER = shap.KernelExplainer(model_proba_from_z, BACKGROUND_Z)
print("SHAP explainer rebuilt after retrain.")
except Exception as e:
EXPLAINER = None
print("⚠️ Failed to rebuild SHAP explainer:", repr(e))
print(">>> RETRAIN COMPLETE")
return True
# ============================================================
# API ENDPOINT: APPEND + RETRAIN
# ============================================================
@app.post("/append_and_retrain")
def append_and_retrain(payload: dict):
"""
Appends a new fingerprint row to fingerprints_db.csv
and retrains the model.
Expected payload:
{
"qtr": "2014Q4",
"company": "COAC Ambato Ltda",
"supervisor": "SEPS",
"rating_raw": "B",
"features": {
"autosuf_oper": 0.536154555,
"improductiva": null,
"gastos_fin_over_avg_cart": 1.200803646,
"_equity": ...,
...
"roa_pre_tax": 1.580296249
}
}
- rating_raw is the letter rating (AAA, A-, BBB+, BB-, ...)
- we derive RATING_GROUP (Top / Mid-Top / Mid / Mid-Low / Low)
using the same logic as in the training script.
"""
qtr = payload.get("qtr")
company = payload.get("company")
supervisor = payload.get("supervisor")
rating_raw = payload.get("rating_raw")
feats = payload.get("features", {})
if not qtr or not company or not rating_raw:
return {"ok": False, "error": "Missing qtr/company/rating_raw"}
if set(feats.keys()) != set(FEATURES):
return {"ok": False, "error": "features must contain all 21 ratio names"}
rating_group = letter_to_group(rating_raw)
if rating_group is None:
return {"ok": False, "error": f"Cannot map rating_raw '{rating_raw}' to 5-class group"}
# Build new row matching your CSV schema
row = {
"QTR": qtr,
"COMPANY": company,
"Supervisor": supervisor,
"RATING_RAW": rating_raw,
"RATING_GROUP": rating_group,
**feats,
"rating_score": None # optional, can be filled later
}
# Append row to CSV
if os.path.exists(FINGERPRINT_CSV):
df = pd.read_csv(FINGERPRINT_CSV)
df = pd.concat([df, pd.DataFrame([row])], ignore_index=True)
else:
df = pd.DataFrame([row])
df.to_csv(FINGERPRINT_CSV, index=False)
# Retrain model on full updated DB
retrain_model()
return {"ok": True, "message": "Fingerprint appended and model retrained"}
@app.get("/debug/db_head")
def debug_db_head(n: int = 5):
import os
import pandas as pd
if not os.path.exists(FINGERPRINT_CSV):
return {
"exists": False,
"message": f"{FINGERPRINT_CSV} not found in current working dir."
}
df = pd.read_csv(FINGERPRINT_CSV)
return {
"exists": True,
"file": FINGERPRINT_CSV,
"rows": int(len(df)),
"head": df.head(n).to_dict(orient="records"),
"columns": list(df.columns),
}
import pandas as pd # make sure this is at the top of the file if not already
# from here: after append_and_retrain
@app.get("/debug/db_tail")
def debug_db_tail(n: int = 10):
"""
Returns the last n rows of fingerprints_db.csv so you can verify
that new points are really being appended inside the container.
"""
if not os.path.exists(FINGERPRINT_CSV):
return {"ok": False, "error": f"{FINGERPRINT_CSV} not found"}
try:
df = pd.read_csv(FINGERPRINT_CSV)
except Exception as e:
return {"ok": False, "error": f"Failed to read CSV: {e}"}
tail = df.tail(n)
return {
"ok": True,
"rows": tail.to_dict(orient="records"),
"n_rows_total": int(df.shape[0]),
"n_returned": int(tail.shape[0]),
} |