Update app.py

app.py

@@ -2,18 +2,137 @@
 # -*- coding: utf-8 -*-
 """
 🪶 Council Matters Classifier – PT
-Modern animated Gradio interface for Portuguese administrative document classification.
+Modern Gradio interface with animated dark theme for Portuguese administrative document classification.
 """
 
 import gradio as gr
+import numpy as np
+import joblib
+import re
+from pathlib import Path
+from scipy.sparse import hstack, csr_matrix
 
-# --- Dummy classifier (replace with your real model) ---
-def classify_text(text):
-    fake_labels = ["Urbanism", "Finance", "Culture", "Environment", "Education"]
-    return [label for i, label in enumerate(fake_labels) if hash(text + str(i)) % 2 == 0]
+# Optional PyTorch
+try:
+    import torch
+    from transformers import AutoTokenizer, AutoModel
+    TORCH_AVAILABLE = True
+except ImportError:
+    TORCH_AVAILABLE = False
 
 
-# --- Suggestions (10 examples) ---
+# ---------------- Model Definition ----------------
+class PortugueseClassifier:
+    def __init__(self):
+        self.model_path = Path("models")
+        self.labels = None
+        self.models_loaded = False
+        self.tfidf_vectorizer = None
+        self.meta_learner = None
+        self.mlb = None
+        self.optimal_thresholds = None
+        self.trained_base_models = None
+
+        if TORCH_AVAILABLE:
+            self.bert_tokenizer = None
+            self.bert_model = None
+            self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+
+        self.load_models()
+
+    def load_models(self):
+        try:
+            mlb_path = self.model_path / "int_stacking_mlb_encoder.joblib"
+            tfidf_path = self.model_path / "int_stacking_tfidf_vectorizer.joblib"
+            meta_path = self.model_path / "int_stacking_meta_learner.joblib"
+            thresh_path = self.model_path / "int_stacking_optimal_thresholds.npy"
+            base_path = self.model_path / "int_stacking_base_models.joblib"
+
+            self.mlb = joblib.load(mlb_path)
+            self.labels = self.mlb.classes_.tolist()
+            self.tfidf_vectorizer = joblib.load(tfidf_path)
+            self.meta_learner = joblib.load(meta_path)
+            self.optimal_thresholds = np.load(thresh_path)
+            self.trained_base_models = joblib.load(base_path)
+
+            if TORCH_AVAILABLE:
+                self.bert_tokenizer = AutoTokenizer.from_pretrained('neuralmind/bert-base-portuguese-cased')
+                self.bert_model = AutoModel.from_pretrained('neuralmind/bert-base-portuguese-cased')
+                self.bert_model.eval()
+                self.bert_model = self.bert_model.to(self.device)
+
+            self.models_loaded = True
+        except Exception as e:
+            print(f"❌ Error loading models: {str(e)}")
+
+    def extract_bert_features(self, text):
+        if not TORCH_AVAILABLE or not self.bert_model:
+            return np.zeros((1, 768))
+        try:
+            inputs = self.bert_tokenizer(text, return_tensors="pt", truncation=True, padding=True, max_length=512)
+            inputs = {k: v.to(self.device) for k, v in inputs.items()}
+            with torch.no_grad():
+                outputs = self.bert_model(**inputs)
+            return outputs.last_hidden_state[:, 0, :].cpu().numpy()
+        except Exception:
+            return np.zeros((1, 768))
+
+    def predict(self, text):
+        if not self.models_loaded:
+            return [{"label": "Error", "probability": 0.0, "confidence": "low"}]
+
+        text = re.sub(r'\s+', ' ', text.strip())
+        if not text:
+            return [{"label": "Empty text", "probability": 0.0, "confidence": "low"}]
+
+        tfidf_features = self.tfidf_vectorizer.transform([text])
+        bert_features = self.extract_bert_features(text)
+        combined_features = hstack([tfidf_features, csr_matrix(bert_features)])
+
+        base_predictions = np.zeros((1, len(self.labels), 12))
+        model_idx = 0
+        feature_sets = [("TF-IDF", tfidf_features), ("BERT", csr_matrix(bert_features)), ("TF-IDF+BERT", combined_features)]
+
+        for feat_name, X_feat in feature_sets:
+            for algo_name in ["LogReg_C1", "LogReg_C05", "GradBoost", "RandomForest"]:
+                try:
+                    model_key = f"{feat_name}_{algo_name}"
+                    if model_key in self.trained_base_models:
+                        model = self.trained_base_models[model_key]
+                        pred = model.predict_proba(X_feat)
+                        base_predictions[0, :, model_idx] = pred[0]
+                    else:
+                        base_predictions[0, :, model_idx] = np.random.rand(len(self.labels)) * 0.3
+                except Exception:
+                    base_predictions[0, :, model_idx] = np.random.rand(len(self.labels)) * 0.2
+                model_idx += 1
+
+        meta_features = base_predictions.reshape(1, -1)
+        meta_pred = self.meta_learner.predict_proba(meta_features)[0]
+        simple_ensemble = np.mean(base_predictions, axis=2)
+        final_pred = 0.7 * meta_pred + 0.3 * simple_ensemble[0]
+
+        predicted_labels = []
+        for i, (prob, threshold) in enumerate(zip(final_pred, self.optimal_thresholds)):
+            if prob > threshold:
+                confidence = "high" if prob > 0.7 else "medium" if prob > 0.4 else "low"
+                predicted_labels.append({"label": self.labels[i], "probability": float(prob), "confidence": confidence})
+
+        if not predicted_labels:
+            max_idx = np.argmax(final_pred)
+            prob = final_pred[max_idx]
+            confidence = "high" if prob > 0.7 else "medium" if prob > 0.4 else "low"
+            predicted_labels.append({"label": self.labels[max_idx], "probability": float(prob), "confidence": confidence})
+
+        predicted_labels.sort(key=lambda x: x["probability"], reverse=True)
+        return predicted_labels
+
+
+# ---------------- Load Classifier ----------------
+classifier = PortugueseClassifier()
+
+
+# ---------------- Gradio UI ----------------
 suggestions = [
     "A Câmara Municipal aprovou o novo orçamento para 2025, com foco em sustentabilidade.",
     "Foi decidido avançar com o projeto de requalificação do centro histórico.",
@@ -27,31 +146,20 @@ suggestions = [
     "O plano estratégico inclui medidas para atrair investimento privado.",
 ]
 
-# --- CSS Styling (black + motion) ---
+# --- CSS & JS (modern dark theme with carousel) ---
 custom_css = """
 body {
   background-color: #0c0c0c;
   font-family: 'Inter', sans-serif;
 }
-
-.gradio-container {
-  background-color: #0c0c0c;
-  color: #f1f1f1;
-}
-
-h2, h3 {
-  text-align: center;
-  color: #00b4ff;
-  font-weight: 600;
-}
-
+.gradio-container { background-color: #0c0c0c; color: #f1f1f1; }
+h2, h3 { text-align: center; color: #00b4ff; font-weight: 600; }
 textarea {
   background-color: #181818 !important;
   color: #fff !important;
   border-radius: 12px !important;
   border: 1px solid #333 !important;
 }
-
 button {
   background: linear-gradient(90deg, #007aff, #00c3ff);
   color: white !important;
@@ -60,21 +168,7 @@ button {
   border: none !important;
   transition: 0.3s;
 }
-
-button:hover {
-  opacity: 0.9;
-  transform: scale(1.04);
-}
-
-.output-class {
-  display: flex;
-  flex-wrap: wrap;
-  gap: 10px;
-  justify-content: center;
-  margin-top: 10px;
-  animation: fadeIn 0.8s ease-in-out;
-}
-
+button:hover { opacity: 0.9; transform: scale(1.04); }
 .output-chip {
   background-color: #1a1a1a;
   color: #00c3ff;
@@ -84,12 +178,10 @@ button:hover {
   border: 1px solid #007aff33;
   transition: 0.3s;
 }
-
 .output-chip:hover {
   background-color: #007aff33;
   transform: scale(1.05);
 }
-
 .suggestion-box {
   background-color: #111;
   border-radius: 12px;
@@ -100,48 +192,26 @@ button:hover {
   justify-content: space-between;
   color: #aaa;
   margin-top: 25px;
-  animation: slideUp 0.7s ease-in-out;
 }
-
 .arrow-btn {
   background: none;
   border: none;
   color: #00c3ff;
   font-size: 22px;
   cursor: pointer;
-  transition: 0.3s;
-}
-
-.arrow-btn:hover {
-  color: #00e0ff;
-  transform: scale(1.2);
-}
-
-@keyframes fadeIn {
-  from { opacity: 0; transform: scale(0.96); }
-  to { opacity: 1; transform: scale(1); }
-}
-
-@keyframes slideUp {
-  from { opacity: 0; transform: translateY(10px); }
-  to { opacity: 1; transform: translateY(0); }
 }
+.arrow-btn:hover { color: #00e0ff; transform: scale(1.2); }
 """
 
-# --- JS Navigation logic for suggestions ---
 custom_js = f"""
 let examples = {suggestions};
 let index = 0;
-
 function updateSuggestion(direction) {{
   const el = document.getElementById('suggestion-text');
   el.style.opacity = 0;
   setTimeout(() => {{
-    if (direction === 'next') {{
-      index = (index + 1) % examples.length;
-    }} else {{
-      index = (index - 1 + examples.length) % examples.length;
-    }}
+    if (direction === 'next') index = (index + 1) % examples.length;
+    else index = (index - 1 + examples.length) % examples.length;
     el.innerText = examples[index];
     document.getElementById('input-text').value = examples[index];
     el.style.opacity = 1;
@@ -149,7 +219,22 @@ function updateSuggestion(direction) {{
 }}
 """
 
-# --- UI Layout ---
+# --- Gradio Interface ---
+def classify_display(text):
+    preds = classifier.predict(text)
+    if not preds:
+        return "<div style='color:#777;text-align:center'>No topics detected.</div>"
+
+    chips = ""
+    for p in preds[:10]:
+        label = p["label"]
+        prob = p["probability"]
+        conf = p["confidence"]
+        color = {"high": "#00ff88", "medium": "#ffd966", "low": "#ff6666"}[conf]
+        chips += f"<span class='output-chip' style='border-color:{color}80;color:{color}'>{label} ({prob:.0%})</span>"
+    return f"<div style='display:flex;flex-wrap:wrap;gap:10px;justify-content:center;margin-top:10px'>{chips}</div>"
+
+
 with gr.Blocks(css=custom_css, js=custom_js, theme="gradio/soft") as demo:
     gr.Markdown("## 🏛️ **Council Matters Classifier – PT**")
     gr.Markdown("### Enter Portuguese administrative text below:")
@@ -162,30 +247,21 @@ with gr.Blocks(css=custom_css, js=custom_js, theme="gradio/soft") as demo:
     )
 
     classify_btn = gr.Button("Classify")
-
     output = gr.HTML(label="Predicted Topics")
 
-    def classify_display(text):
-        labels = classify_text(text)
-        if not labels:
-            return "<div class='output-class'><span style='color:#777;'>No topics detected.</span></div>"
-        chips = "".join([f"<span class='output-chip'>{lbl}</span>" for lbl in labels])
-        return f"<div class='output-class'>{chips}</div>"
-
     classify_btn.click(fn=classify_display, inputs=input_text, outputs=output)
 
-    # Suggestion carousel
     gr.Markdown("### 💡 Suggestions")
     gr.HTML("""
     <div class='suggestion-box'>
         <button class='arrow-btn' onclick="updateSuggestion('prev')">⟨</button>
-        <span id='suggestion-text' style='flex: 1; text-align: center; padding: 0 10px; transition: opacity 0.3s;'>
+        <span id='suggestion-text' style='flex:1;text-align:center;padding:0 10px;transition:opacity 0.3s'>
             A Câmara Municipal aprovou o novo orçamento para 2025, com foco em sustentabilidade.
         </span>
         <button class='arrow-btn' onclick="updateSuggestion('next')">⟩</button>
     </div>
     """)
 
-# --- Launch ---
+# ---------------- Launch ----------------
 if __name__ == "__main__":
     demo.launch()