Update app.py

app.py

@@ -6,10 +6,14 @@ import joblib
 import numpy as np
 import librosa
 import gradio as gr
+import time
+import re
+from transformers import pipeline
 from huggingface_hub import hf_hub_download
 from deepface import DeepFace
 from transformers import AutoTokenizer, AutoModelForSequenceClassification, pipeline
 
+
 # --- 1. 下載並載入 SVM 模型 ---
 # 這裡 repo_id 填你的模型倉庫路徑，例如 "GCLing/emotion-svm-model"
 # filename 填上傳到該倉庫的檔案名，例如 "svm_emotion_model.joblib"
@@ -20,12 +24,21 @@ svm_model = joblib.load(model_path)
 print("SVM model loaded.")
 
 # --- 2. 載入文字情緒分析模型 ---
-# 以 uer/roberta-base-finetuned-chinanews-chinese 為例；可替換成其他合適的中文情感分類模型
-print("Loading text sentiment model...")
-tokenizer = AutoTokenizer.from_pretrained("uer/roberta-base-finetuned-chinanews-chinese")
-model_txt = AutoModelForSequenceClassification.from_pretrained("uer/roberta-base-finetuned-chinanews-chinese")
-text_emotion = pipeline("sentiment-analysis", model=model_txt, tokenizer=tokenizer)
-print("Text sentiment model loaded.")
+zero_shot = pipeline("zero-shot-classification", model="joeddav/xlm-roberta-large-xnli")
+candidate_labels = ["joy", "sadness", "anger", "fear", "surprise", "disgust"]
+label_map_en2cn = {
+    "joy": "高興", "sadness": "悲傷", "anger": "憤怒",
+    "fear": "恐懼", "surprise": "驚訝", "disgust": "厭惡"
+}
+emo_keywords = {
+    "happy": ["開心","快樂","愉快","喜悦","喜悅","歡喜","興奮","高興"],
+    "angry": ["生氣","憤怒","不爽","發火","火大","氣憤"],
+    "sad": ["傷心","難過","哭","難受","心酸","憂","悲","哀","痛苦","慘","愁"],
+    "surprise": ["驚訝","意外","嚇","驚詫","詫異","訝異","好奇"],
+    "fear": ["怕","恐懼","緊張","懼","膽怯","畏"]
+}
+# 简单否定词列表
+negations = ["不","沒","沒有","別","勿","非"]
 
 # --- 3. 聲音特徵擷取函式 ---
 def extract_feature(signal: np.ndarray, sr: int) -> np.ndarray:
@@ -89,22 +102,44 @@ def predict_voice(audio_path: str):
 
 
     
-def predict_text(text: str):
-    # 这里至少要有一行缩进代码，不能留空
+def predict_text_mixed(text: str):
+    """
+    先用 keyword_emotion 规则；若未命中再用 zero-shot 分类，
+    返回 {中文标签: float_score} 的 dict，供 gr.Label 显示。
+    """
     if not text or text.strip() == "":
         return {}
-    # 其余逻辑...
-
-    if not text or text.strip()=="":
-        return {}
+    # 规则优先
+    res = keyword_emotion(text)
+    if res:
+        # 只返回最高那一项及其比例，也可返回完整分布
+        top_emo = max(res, key=res.get)
+        # 可将英文 key 转成中文，若需要
+        # mapping: happy->高兴, angry->愤怒, etc.
+        mapping = {
+            "happy": "高兴",
+            "angry": "愤怒",
+            "sad": "悲伤",
+            "surprise": "惊讶",
+            "fear": "恐惧"
+        }
+        cn = mapping.get(top_emo, top_emo)
+        return {cn: res[top_emo]}
+    # 规则未命中，zero-shot fallback
     try:
-        pred = text_emotion(text)[0]
-        result = {pred["label"]: float(pred["score"])}
-        print("Text sentiment result:", result)
+        out = zero_shot(text, candidate_labels=candidate_labels,
+                        hypothesis_template="这句话表达了{}情绪")
+        result = {}
+        for lab, sc in zip(out["labels"], out["scores"]):
+            cn = label_map_en2cn.get(lab.lower(), lab)
+            result[cn] = float(sc)
         return result
     except Exception as e:
-        print("predict_text error:", e)
-        return {}
+        print("zero-shot error:", e)
+        return {"中性": 1.0}
+
+
+
 
 
 # --- 5. 建立 Gradio 介面 ---
@@ -128,10 +163,11 @@ with gr.Blocks() as demo:
             audio.change(fn=predict_voice, inputs=audio, outputs=audio_output)
 
         with gr.TabItem("文字情緒"):
-            text = gr.Textbox(lines=3, placeholder="請輸入中文文字…")
-            text_output = gr.Label(label="文字情緒結果")
-            btn = gr.Button("分析文字")
-            btn.click(fn=predict_text, inputs=text, outputs=text_output)
+            gr.Markdown("### 文字情緒 分析 (规则+zero-shot)")
+            with gr.Row():
+                text = gr.Textbox(lines=3, placeholder="請輸入中文文字…")
+                text_out = gr.Label(label="文字情緒結果")
+            text.submit(fn=predict_text_mixed, inputs=text, outputs=text_out)