Update app.py

app.py

@@ -544,4 +544,180 @@ async def predict(req: Request):
         return JSONResponse(
             status_code=500,
             content={"error": str(e), "trace": traceback.format_exc()},
-        )
+        )
+
+
+    # ============================================================
+# CORAL ORDINAL HELPERS (copied from your training script)
+# ============================================================
+
+def to_cumulative_targets_tf(y_true_int, K_):
+    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):
+    y_true = tf.reshape(y_true, [-1])
+    y_true = tf.cast(y_true, tf.int32)
+    T = to_cumulative_targets_tf(y_true, len(CLASSES))
+    bce = tf.nn.sigmoid_cross_entropy_with_logits(labels=T, logits=logits)
+    return tf.reduce_mean(tf.reduce_sum(bce, axis=1))
+
+def coral_probs_from_logits(logits):
+    sig = tf.math.sigmoid(logits)
+    left  = tf.concat([tf.ones_like(sig[:, :1]), sig], axis=1)
+    right = tf.concat([sig, tf.zeros_like(sig[:, :1])], axis=1)
+    return tf.clip_by_value(left - right, 1e-12, 1.0)
+
+@tf.function
+def ordinal_accuracy_metric(y_true, y_pred_logits):
+    y_true = tf.reshape(y_true, [-1])
+    y_true = tf.cast(y_true, tf.int32)
+    probs  = coral_probs_from_logits(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))
+
+    # ============================================================
+# RECREATE MODEL FROM BEST HYPERPARAMETERS
+# ============================================================
+
+def build_model_from_hparams(hp: dict):
+    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
+    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
+# ============================================================
+
+FINGERPRINT_CSV = "fingerprints_db.csv"   # <-- choose file name
+BEST_HP_JSON     = "best_params_and_metrics.json"
+
+
+def load_best_hparams():
+    with open(BEST_HP_JSON, "r") as f:
+        js = json.load(f)
+    return js["best_hyperparams"]
+
+
+def load_fingerprint_dataset():
+    df = pd.read_csv(FINGERPRINT_CSV)
+
+    # Must include: company, date, rating, and 21 features
+    y = df["rating"].map({c:i for i,c in enumerate(CLASSES)}).astype("int32").to_numpy()
+    X_raw = df[FEATURES].to_numpy().astype("float32")
+
+    # Fit imputer + scaler from full dataset
+    imp = SimpleImputer(strategy="median")
+    sc  = StandardScaler()
+
+    X_imp = imp.fit_transform(X_raw)
+    X_sc  = sc.fit_transform(X_imp)
+
+    return X_sc, y, imp, sc
+
+
+def retrain_model():
+    hp = load_best_hparams()
+    X, y, imp, sc = load_fingerprint_dataset()
+
+    model_new = build_model_from_hparams(hp)
+
+    es = tf.keras.callbacks.EarlyStopping(
+        monitor="loss", patience=15, restore_best_weights=True
+    )
+
+    model_new.fit(
+        X, y,
+        epochs=150,
+        batch_size=128,
+        verbose=1,
+        callbacks=[es]
+    )
+
+    # Update globals used by /predict
+    global model, IMPUTER, SCALER
+    model = model_new
+    IMPUTER = imp
+    SCALER  = sc
+
+    return True
+
+
+
+    # ============================================================
+# API ENDPOINT: APPEND + RETRAIN
+# ============================================================
+
+@app.post("/append_and_retrain")
+def append_and_retrain(payload: dict):
+    """
+    payload:
+    {
+        "company": "...",
+        "date": "2025-Q1",
+        "rating": "Mid-Top",
+        "features": { autosuf_oper: ..., improductiva: ..., ... }
+    }
+    """
+
+    company = payload.get("company")
+    date    = payload.get("date")
+    rating  = payload.get("rating")
+    feats   = payload.get("features", {})
+
+    if not company or not rating or len(feats) != len(FEATURES):
+        return {"ok": False, "error": "Invalid payload"}
+
+    # Append to CSV
+    df_new = pd.DataFrame([{**{"company": company,
+                               "date": date,
+                               "rating": rating},
+                            **feats}])
+
+    if os.path.exists(FINGERPRINT_CSV):
+        df = pd.read_csv(FINGERPRINT_CSV)
+        df = pd.concat([df, df_new], ignore_index=True)
+    else:
+        df = df_new
+
+    df.to_csv(FINGERPRINT_CSV, index=False)
+
+    # Retrain
+    retrain_model()
+
+    return {"ok": True, "message": "Fingerprint added and model retrained"}