Update app.py

app.py

@@ -40,65 +40,82 @@ class TextRequest(BaseModel):
     text: str
 
 # =========================================================
-# OPERASI BEDAH FILE .KERAS (Solusi Definitif Error JSON)
+# JURUS 1: BYPASS ARSITEKTUR MANUAL UNTUK NER (Anti-Error LSTM)
+# =========================================================
+def build_manual_ner_model(vocab_size=7929, max_len=64, num_classes=8):
+    inputs = tf.keras.Input(shape=(max_len,), dtype='int32', name='tokens')
+    x = tf.keras.layers.Embedding(input_dim=vocab_size, output_dim=64, mask_zero=True, name='embedding')(inputs)
+    x = tf.keras.layers.SpatialDropout1D(0.3, name='spatial_dropout1d')(x)
+    
+    lstm1_f = tf.keras.layers.LSTM(128, return_sequences=True, dropout=0.4, name='forward_lstm')
+    lstm1_b = tf.keras.layers.LSTM(128, return_sequences=True, go_backwards=True, dropout=0.4, name='backward_lstm')
+    x = tf.keras.layers.Bidirectional(lstm1_f, backward_layer=lstm1_b, name='bilstm_1')(x)
+    
+    lstm2_f = tf.keras.layers.LSTM(64, return_sequences=True, dropout=0.4, name='forward_lstm_1')
+    lstm2_b = tf.keras.layers.LSTM(64, return_sequences=True, go_backwards=True, dropout=0.4, name='backward_lstm_1')
+    x = tf.keras.layers.Bidirectional(lstm2_f, backward_layer=lstm2_b, name='bilstm_2')(x)
+    
+    dense1 = tf.keras.layers.Dense(64, activation='relu', name='dense')
+    x = tf.keras.layers.TimeDistributed(dense1, name='proj')(x)
+    x = tf.keras.layers.Dropout(0.4, name='dropout')(x)
+    
+    dense2 = tf.keras.layers.Dense(num_classes, activation='linear', name='dense_1')
+    x = tf.keras.layers.TimeDistributed(dense2, name='logits')(x)
+    probs = tf.keras.layers.Activation('softmax', name='probabilities')(x)
+    
+    return tf.keras.Model(inputs=inputs, outputs=probs, name='BiLSTM_NER')
+
+def get_ner_model():
+    global ner_model, vocab
+    if ner_model is None:
+        with open("vocab.pkl", "rb") as f:
+            vocab = pickle.load(f)
+        # Load kerangka manual, isi dengan weights
+        ner_model = build_manual_ner_model()
+        ner_model.load_weights("receipt_ner_model.keras")
+    return ner_model, vocab
+
+# =========================================================
+# JURUS 2: BEDAH JSON UNTUK KLASIFIKASI (Menyelamatkan Kamus Kata)
 # =========================================================
 def patch_keras_model(original_path):
     patched_path = "patched_" + original_path
-    # Jika sudah pernah dibedah, langsung gunakan hasilnya
-    if os.path.exists(patched_path):
-        return patched_path
+    if os.path.exists(patched_path): return patched_path
     
-    print(f"Membedah {original_path} untuk menghapus quantization_config...")
     try:
         with zipfile.ZipFile(original_path, 'r') as zin:
             with zipfile.ZipFile(patched_path, 'w') as zout:
                 for item in zin.infolist():
                     content = zin.read(item.filename)
                     if item.filename == 'config.json':
-                        # Bongkar JSON-nya
                         config_dict = json.loads(content.decode('utf-8'))
-                        
-                        # Fungsi rekursif untuk mencari dan menghapus quantization_config di semua layer
                         def hapus_biang_kerok(d):
                             if isinstance(d, dict):
                                 d.pop('quantization_config', None)
-                                for k, v in d.items():
-                                    hapus_biang_kerok(v)
+                                for v in d.values(): hapus_biang_kerok(v)
                             elif isinstance(d, list):
-                                for i in d:
-                                    hapus_biang_kerok(i)
-                        
+                                for i in d: hapus_biang_kerok(i)
                         hapus_biang_kerok(config_dict)
-                        # Bungkus kembali jadi JSON
                         content = json.dumps(config_dict).encode('utf-8')
-                        
                     zout.writestr(item, content)
         return patched_path
     except Exception as e:
         print(f"Gagal membedah model: {e}")
-        return original_path # Fallback ke file asli jika gagal
-
-# =========================================================
-# LAZY LOAD MENGGUNAKAN MODEL YANG SUDAH DIBEDAH
-# =========================================================
-def get_ner_model():
-    global ner_model, vocab
-    if ner_model is None:
-        with open("vocab.pkl", "rb") as f:
-            vocab = pickle.load(f)
-        patched_file = patch_keras_model("receipt_ner_model.keras")
-        ner_model = tf.keras.models.load_model(patched_file, compile=False)
-    return ner_model, vocab
+        return original_path
 
 def get_clf_model():
     global clf_model, clf_encoder
     if clf_model is None:
         with open('penerjemah_kategori.pkl', 'rb') as file:
             clf_encoder = pickle.load(file)
+        # Bedah JSON lalu load utuh agar TextVectorization bekerja
         patched_file = patch_keras_model("model_strukly.keras")
         clf_model = tf.keras.models.load_model(patched_file, compile=False)
     return clf_model, clf_encoder
 
+# =========================================================
+# FUNGSI UTAMA
+# =========================================================
 def normalize_token(token: str):
     token = token.strip()
     if re.fullmatch(r"\d{1,3}([.,]\d{3})*", token): return PRICE_TOKEN
@@ -122,15 +139,14 @@ def predict_ner(request: TextRequest):
         
         if tokens:
             norm_tokens = [normalize_token(t) for t in tokens]
+            # Karena pakai arsitektur manual, mask tidak perlu dibuat eksplisit (otomatis di layer Embedding)
             X = np.zeros((1, MAX_SEQ_LEN), dtype=np.int32)
-            mask = np.zeros((1, MAX_SEQ_LEN), dtype=np.float32)
             length = min(len(norm_tokens), MAX_SEQ_LEN)
             
             for i in range(length):
                 X[0, i] = vcb.get(norm_tokens[i], vcb.get(UNK_TOKEN, 0))
-                mask[0, i] = 1.0
                 
-            probs = model([X, mask], training=False).numpy()
+            probs = model(X, training=False).numpy()
             labels = [IDX2LABEL[int(np.argmax(probs[0, i]))] for i in range(length)]
             
             current_type = None