Update app.py

app.py

@@ -1,7 +1,8 @@
 # ================================
 # 🪞 MoodMirror+ — Conversational Emotional Self-Care
-# Uses ONLY the GoEmotions dataset (no pretrained model)
-# Trains TF-IDF + OneVsRest Logistic Regression on first run, caches to /data
+# Dataset-only: trains a TF-IDF + OneVsRest Logistic Regression on GoEmotions
+# Cache du modèle + DB dans /data quand dispo (HF Spaces: activer Persistent storage)
+# Toujours donner au moins 1 conseil + parfois une citation
 # ================================
 import os
 import re
@@ -36,8 +37,14 @@ print(f"[MM] SQLite path:   {DB_PATH}")
 print(f"[MM] Model path:    {MODEL_PATH}")
 
 # ---------------- Crisis & regex ----------------
-CRISIS_RE = re.compile(r"\b(self[- ]?harm|suicid|kill myself|end my life|overdose|cutting|i don.?t want to live|can.?t go on)\b", re.I)
-CLOSING_RE = re.compile(r"\b(thanks?|thank you|that'?s all|bye|goodbye|see you|take care|ok bye|no thanks?)\b", re.I)
+CRISIS_RE = re.compile(
+    r"\b(self[- ]?harm|suicid|kill myself|end my life|overdose|cutting|i don.?t want to live|can.?t go on)\b",
+    re.I,
+)
+CLOSING_RE = re.compile(
+    r"\b(thanks?|thank you|that'?s all|bye|goodbye|see you|take care|ok bye|no thanks?)\b",
+    re.I,
+)
 
 CRISIS_NUMBERS = {
     "United States": "Call or text **988** (24/7 Suicide & Crisis Lifeline). If in immediate danger, call **911**.",
@@ -284,16 +291,22 @@ def train_or_load_model():
     Y_val = mlb.transform(y_val)
 
     clf = Pipeline([
-        ("tfidf", TfidfVectorizer(lowercase=True, ngram_range=(1, 2), min_df=2, max_df=0.9, strip_accents="unicode")),
-        ("ovr", OneVsRestClassifier(LogisticRegression(solver="saga", max_iter=1000, n_jobs=-1, class_weight="balanced"), n_jobs=-1))
+        ("tfidf", TfidfVectorizer(
+            lowercase=True, ngram_range=(1, 2), min_df=2, max_df=0.9, strip_accents="unicode"
+        )),
+        ("ovr", OneVsRestClassifier(
+            LogisticRegression(solver="saga", max_iter=1000, n_jobs=-1, class_weight="balanced"),
+            n_jobs=-1
+        ))
     ])
 
     print("[MM] Training classifier...")
     clf.fit(X_train, Y_train)
-
-    print(f"[MM] Validation macro F1: {f1_score(Y_val, clf.predict(X_val), average='macro', zero_division=0):.3f}")
+    macro_f1 = f1_score(Y_val, clf.predict(X_val), average="macro", zero_division=0)
+    print(f"[MM] Validation macro F1: {macro_f1:.3f}")
 
     joblib.dump({"version": MODEL_VERSION, "pipeline": clf, "mlb": mlb, "label_names": label_names}, MODEL_PATH)
+    print(f"[MM] Saved model -> {MODEL_PATH}")
     return clf, mlb, label_names
 
 try:
@@ -304,7 +317,8 @@ except Exception as e:
 
 # ---------------- Emotion detection ----------------
 def classify_text(text: str):
-    if not CLASSIFIER: return []
+    if not CLASSIFIER:
+        return []
     try:
         proba = CLASSIFIER.predict_proba([text])[0]
     except AttributeError:
@@ -316,7 +330,8 @@ def classify_text(text: str):
 
 def detect_emotions(text: str):
     chosen = classify_text(text)
-    if not chosen: return "neutral"
+    if not chosen:
+        return "neutral"
     bucket = {}
     for label, p in chosen:
         app = GOEMO_TO_APP.get(label.lower(), "neutral")
@@ -324,9 +339,16 @@ def detect_emotions(text: str):
     return max(bucket, key=bucket.get)
 
 # ---------------- Reply composer ----------------
-def compose_support_legacy(main_emotion: str, is_first_msg: bool) -> str:
-    tip = random.choice(SUGGESTIONS.get(main_emotion, ["Take a slow breath. One small act of kindness can shift your day."]))
-    quote = random.choice(QUOTES.get(main_emotion, ["“No matter what you feel right now, this moment will pass.”"]))
+def compose_support(main_emotion: str, is_first_msg: bool) -> str:
+    # Always include an advice tip; 50% chance to add a quote
+    tip = random.choice(SUGGESTIONS.get(
+        main_emotion,
+        ["Take a slow breath. One small act of kindness can shift your day."]
+    ))
+    quote = random.choice(QUOTES.get(
+        main_emotion,
+        ["“No matter what you feel right now, this moment will pass.”"]
+    ))
     include_quote = random.random() < 0.5
 
     reply = tip
@@ -351,4 +373,76 @@ def chat_step(message, history, country, save_session):
     if CLOSING_RE.search(message):
         return ("Thank you 💛 Take care of yourself. Small steps matter. 🌿", "#FFFFFF")
 
-    emotion = detect_emotions(" ".join
+    recent = " ".join(message.split()[-100:])
+    emotion = detect_emotions(recent)
+    color = COLOR_MAP.get(emotion, "#FFFFFF")
+
+    if save_session:
+        log_session(country, message, emotion)
+
+    reply = compose_support(emotion, is_first_msg=not bool(history))
+    return reply, color
+
+# ---------------- Gradio UI ----------------
+init_db()
+
+custom_css = """
+:root, body, .gradio-container { transition: background-color 0.8s ease !important; }
+.typing { font-style: italic; opacity: 0.8; animation: blink 1s infinite; }
+@keyframes blink { 50% {opacity: 0.4;} }
+"""
+
+with gr.Blocks(css=custom_css, title="🪞 MoodMirror+ (Dataset-only Edition)") as demo:
+    style_injector = gr.HTML("")
+    gr.Markdown(
+        "### 🪞 MoodMirror+ — Emotional Support & Inspiration 🌸\n"
+        "Powered only by the **GoEmotions dataset** (trained locally on startup).\n\n"
+        "_Not medical advice. If you feel unsafe, please reach out for help immediately._"
+    )
+
+    with gr.Row():
+        country = gr.Dropdown(choices=list(CRISIS_NUMBERS.keys()),
+                              value="Other / Not listed", label="Country")
+        save_ok = gr.Checkbox(value=False, label="Save anonymized session (no personal data)")
+
+    chat = gr.Chatbot(height=360)
+    msg = gr.Textbox(placeholder="Type how you feel...", label="Your message")
+    send = gr.Button("Send")
+    typing = gr.Markdown("", elem_classes="typing")
+
+    # Optional: dataset preview
+    with gr.Accordion("🔎 Preview GoEmotions samples", open=False):
+        with gr.Row():
+            n_examples = gr.Slider(1, 10, value=5, step=1, label="Number of examples")
+            split = gr.Dropdown(["train", "validation", "test"], value="train", label="Split")
+            refresh = gr.Button("Show samples")
+        table = gr.Dataframe(headers=["text", "labels"], row_count=5, wrap=True)
+
+        def refresh_samples(n, split_name):
+            try:
+                ds = load_dataset("google-research-datasets/go_emotions", "simplified")
+                names = ds["train"].features["labels"].feature.names
+                rows = ds[split_name].shuffle(seed=42).select(range(min(int(n), len(ds[split_name]))))
+                return [[t, ", ".join([names[i] for i in labs])] for t, labs in zip(rows["text"], rows["labels"])]
+            except Exception as e:
+                return [[f"Dataset load error: {e}", ""]]
+
+        refresh.click(refresh_samples, inputs=[n_examples, split], outputs=[table])
+
+    def respond(user_msg, chat_hist, country_choice, save_flag):
+        if not user_msg or not user_msg.strip():
+            yield chat_hist + [[user_msg, "Please share a short sentence about how you feel 🙂"]], "", "", ""
+            return
+        yield chat_hist, "💭 MoodMirror is thinking...", "", ""
+        reply, color = chat_step(user_msg, chat_hist, country_choice, bool(save_flag))
+        style_tag = f"<style>:root,body,.gradio-container{{background:{color}!important;}}</style>"
+        yield chat_hist + [[user_msg, reply]], "", style_tag, ""
+
+    send.click(respond, inputs=[msg, chat, country, save_ok],
+               outputs=[chat, typing, style_injector, msg], queue=True)
+    msg.submit(respond, inputs=[msg, chat, country, save_ok],
+               outputs=[chat, typing, style_injector, msg], queue=True)
+
+if __name__ == "__main__":
+    demo.queue()
+    demo.launch()