Update app.py

app.py

@@ -1,67 +1,84 @@
 import streamlit as st
 import numpy as np
-try:
-    from transformers import pipeline
-except ImportError:
-    st.error("Transformers pipeline not available. Installing required version...")
-    import subprocess
-    import sys
-    subprocess.check_call([sys.executable, "-m", "pip", "install", "transformers==4.30.0"])
-    from transformers import pipeline
+from transformers import pipeline
 
-# Initialize pipelines with error handling
+# 环境稳定化配置
 @st.cache_resource
 def load_pipelines():
     try:
         sentiment_pipe = pipeline(
-            "text-classification", 
+            "text-classification",
             model="EmmaL1/CustomModel_amazon",
-            return_all_scores=True
+            return_all_scores=True,
+            device="cpu",
+            torch_dtype="float32"
         )
         qa_pipe = pipeline(
-            "question-answering", 
-            model="distilbert/distilbert-base-cased-distilled-squad"
+            "question-answering",
+            model="distilbert/distilbert-base-cased-distilled-squad",
+            device="cpu"
         )
         return sentiment_pipe, qa_pipe
     except Exception as e:
-        st.error(f"Pipeline initialization failed: {str(e)}")
+        st.error(f"Initialization Error: {str(e)}")
         st.stop()
 
 sentiment_pipeline, qa_pipeline = load_pipelines()
 
-def get_rating(prediction):
-    """Get rating (1-5 stars) from model prediction"""
-    scores = [score['score'] for score in prediction[0]]
-    return np.argmax(scores) + 1
+def get_balanced_rating(prediction):    probs = np.array([score['score'] for score in prediction[0][:5]])
+    probs = probs / (probs.sum() + 1e-8)  # 安全标准化
+    
+    # 优化后的阈值设置（可调整）
+    rating_rules = [
+        (0.95, 1),   # 概率>95%才判为1星
+        (0.85, 2),
+        (0.70, 3), 
+        (0.50, 4),
+        (0.00, 5)     # 默认5星
+    ]
+    
+    for threshold, rating in rating_rules:
+        if probs[rating-1] > threshold:
+            return rating, probs[rating-1]
+    
+    return 5, probs[4]  # 默认返回5星
 
 def main():
-    st.title("Amazon Review Analysis")
-    st.write("Analyze review sentiment and predict star rating")
+    st.title("Amazon Review Rating System")
     
-    user_input = st.text_area("Enter review text:", height=150)
-        
-    if st.button("Analyze") and user_input:
+    user_input = st.text_area("Input Review Text:", height=150)
+    
+    if st.button("Analyze"):
         with st.spinner("Processing..."):
             try:
-                # Get prediction
-                sentiment_pred = sentiment_pipeline(user_input)
-                rating = get_rating(sentiment_pred)
+                # 获取预测
+                prediction = sentiment_pipeline(user_input)
+                
+                # 计算优化后的评分
+                rating, confidence = get_balanced_rating(prediction)
                 
-                # Display results
-                st.subheader("Results")
-                st.metric("Predicted Rating", f"{rating}/5 stars")
-                st.metric("Sentiment", "Positive" if rating >= 3 else "Negative")
+                # 显示结果
+                st.subheader("Analysis Result")
+                col1, col2 = st.columns(2)
+                with col1:
+                    st.metric("Predicted Rating", f"{rating} stars")
+                with col2:
+                    st.metric("Confidence", f"{confidence:.1%}")
                 
-                # Generate justification
+                # 生成解释
                 qa_result = qa_pipeline({
-                    'question': f'Why would this get {rating} stars?',
+                    'question': f'What justifies this {rating}-star rating?',
                     'context': user_input
                 })
-                st.subheader("Key Reasons")
+                st.subheader("Key Factors")
                 st.write(qa_result['answer'])
                 
+                # 调试信息（可选）
+                if st.checkbox("Show debug info"):
+                    st.write("Raw probabilities:", [f"{p:.4f}" for p in prediction[0]])
+                
             except Exception as e:
-                st.error(f"Analysis failed: {str(e)}")
+                st.error(f"Error: {str(e)}")
 
 if __name__ == "__main__":
     main()