Update app.py

app.py

@@ -19,41 +19,40 @@ st.write("automative reply")
 # Text input for user to enter the comment
 text = st.text_area("Enter your comment", "")
 
+def main():
+  user_review = "I love the fast delivery, but the product quality could be better."
+  response = response_gen(user_review)
+  print(response)
+  sound_gen(response)
+  return
+
+
+if st.button("Comment"):
 
 ##########################################
 # Step 1：情感分析 - 分析用户评论的情感倾向
 ##########################################
-# Perform tasks when the user clicks the "Comment" button
-if st.button("Comment"):
 
-    pipe = pipeline("text-classification", model="j-hartmann/emotion-english-distilroberta-base")
 
-# 使用 "j-hartmann/emotion-english-distilroberta-base" 模型进行多维度情感分类
-emotion_classifier = pipeline(
-    "text-classification",
-    model="j-hartmann/emotion-english-distilroberta-base",
-    return_all_scores=True
-)
+def analyze_dominant_emotion(user_review):
+    # Initialize the emotion classifier pipeline
+    emotion_classifier = pipeline(
+         "text-classification",
+         model="Thea231/jhartmann_emotion_finetuning",
+        return_all_scores=True
+     )
 
-# 示例用户评论（可替换为实际评论）
-user_review = "I love the fast delivery, but the product quality could be better."
+    # Get emotion predictions for the input review
+    emotion_results = emotion_classifier(user_review)[0]  # Get first result (single input case)
 
-# 对评论进行情感分析
-emotion_results = emotion_classifier(user_review)[0]  # 返回列表中的第一个结果（单条输入）
+    # Extract the emotion with highest confidence score
+    dominant_emotion = max(emotion_results, key=lambda x: x['score'])
 
-# 打印所有情感维度及其分数
+    return dominant_emotion
 
-print("情感分析结果（多维度）：")
-for emotion in emotion_results:
-    print(f"{emotion['label']}: {emotion['score']:.4f}")
-    
-    st.write("Text:", text)
-    st.write("Label:", max_label)
-    st.write("Score:", max_score)
-    
 # 提取置信度最高的情感标签（可选）
-dominant_emotion = max(emotion_results, key=lambda x: x['score'])
-print("\n主导情感：", dominant_emotion['label'], f"(置信度: {dominant_emotion['score']:.2f})")
+# dominant_emotion = analyze_dominant_emotion(user_review)
+# print("\n主导情感：", dominant_emotion['label'], f"(置信度: {dominant_emotion['score']:.2f})")
 
 
 
@@ -61,139 +60,146 @@ print("\n主导情感：", dominant_emotion['label'], f"(置信度: {dominant_em
 # Step 2：回复生成 - 根据情感生成自动回复
 ##########################################
 
-emotion_strategies = {
-    "anger": {
-        "prompt": (
-            "Customer complaint: '{review}'\n\n"
-            "As a customer service representative, craft a professional response that:\n"
-            "- Begins with sincere apology and acknowledgment\n"
-            "- Clearly explains solution process with concrete steps\n"
-            "- Offers appropriate compensation/redemption\n"
-            "- Keeps tone humble and solution-focused (3-4 sentences)\n\n"
-            "Response:"
-        )
-    },
-    "disgust": {
-        "prompt": (
-            "Customer quality concern: '{review}'\n\n"
-            "As a customer service representative, craft a response that:\n"
-            "- Immediately acknowledges the product issue\n"
-            "- Explains quality control measures being taken\n"
-            "- Provides clear return/replacement instructions\n"
-            "- Offers goodwill gesture (3-4 sentences)\n\n"
-            "Response:"
-        )
-    },
-    "fear": {
-        "prompt": (
-            "Customer safety concern: '{review}'\n\n"
-            "As a customer service representative, craft a reassuring response that:\n"
-            "- Directly addresses the safety worries\n"
-            "- References relevant certifications/standards\n"
-            "- Offers dedicated support contact\n"
-            "- Provides satisfaction guarantee (3-4 sentences)\n\n"
-            "Response:"
-        )
-    },
-    "joy": {
-        "prompt": (
-            "Customer review: '{review}'\n\n"
-            "As a customer service representative, craft a concise response that:\n"
-            "- Specifically acknowledges both positive and constructive feedback\n"
-            "- Briefly mentions loyalty/referral programs\n"
-            "- Ends with shopping invitation (3-4 sentences)\n\n"
-            "Response:"
-        )
-    },
-    "neutral": {
-        "prompt": (
-            "Customer feedback: '{review}'\n\n"
-            "As a customer service representative, craft a balanced response that:\n"
-            "- Provides additional relevant product information\n"
-            "- Highlights key service features\n"
-            "- Politely requests more detailed feedback\n"
-            "- Maintains professional tone (3-4 sentences)\n\n"
-            "Response:"
-        )
-    },
-    "sadness": {
-        "prompt": (
-            "Customer disappointment: '{review}'\n\n"
-            "As a customer service representative, craft an empathetic response that:\n"
-            "- Shows genuine understanding of the issue\n"
-            "- Proposes personalized recovery solution\n"
-            "- Offers extended support options\n"
-            "- Maintains positive outlook (3-4 sentences)\n\n"
-            "Response:"
-        )
-    },
-    "surprise": {
-        "prompt": (
-            "Customer enthusiastic feedback: '{review}'\n\n"
-            "As a customer service representative, craft a response that:\n"
-            "- Matches customer's positive energy appropriately\n"
-            "- Highlights unexpected product benefits\n"
-            "- Invites to user community/events\n"
-            "- Maintains brand voice (3-4 sentences)\n\n"
-            "Response:"
-        )
-    }
-}
-
-
-
-# 生成回复Prompt
-template = emotion_strategies[dominant_emotion['label'].lower()]["prompt"]
-prompt = template.format(review=user_review)
-print(prompt)
-
-
-
-
-# 加载Llama-3作为text generation模型
-tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-3.2-1B")
-model = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-3.2-1B")
-
-
-
-
-inputs = tokenizer(prompt, return_tensors="pt")
-outputs = model.generate(**inputs, max_new_tokens=100)
-
-input_length = inputs.input_ids.shape[1]
-response = tokenizer.decode(outputs[0][input_length:], skip_special_tokens=True)
-print(response)
 
+def prompt_gen(user_review):
+  dominant_emotion = analyze_dominant_emotion(user_review)
+  emotion_strategies = {
+      "anger": {
+          "prompt": (
+              "Customer complaint: '{review}'\n\n"
+              "As a customer service representative, craft a professional response that:\n"
+              "- Begins with sincere apology and acknowledgment\n"
+              "- Clearly explains solution process with concrete steps\n"
+              "- Offers appropriate compensation/redemption\n"
+              "- Keeps tone humble and solution-focused (3-4 sentences)\n\n"
+              "Response:"
+          )
+      },
+      "disgust": {
+          "prompt": (
+              "Customer quality concern: '{review}'\n\n"
+              "As a customer service representative, craft a response that:\n"
+              "- Immediately acknowledges the product issue\n"
+              "- Explains quality control measures being taken\n"
+              "- Provides clear return/replacement instructions\n"
+              "- Offers goodwill gesture (3-4 sentences)\n\n"
+              "Response:"
+          )
+      },
+      "fear": {
+          "prompt": (
+              "Customer safety concern: '{review}'\n\n"
+              "As a customer service representative, craft a reassuring response that:\n"
+              "- Directly addresses the safety worries\n"
+              "- References relevant certifications/standards\n"
+              "- Offers dedicated support contact\n"
+              "- Provides satisfaction guarantee (3-4 sentences)\n\n"
+              "Response:"
+          )
+      },
+      "joy": {
+          "prompt": (
+              "Customer review: '{review}'\n\n"
+              "As a customer service representative, craft a concise response that:\n"
+              "- Specifically acknowledges both positive and constructive feedback\n"
+              "- Briefly mentions loyalty/referral programs\n"
+              "- Ends with shopping invitation (3-4 sentences)\n\n"
+              "Response:"
+          )
+      },
+      "neutral": {
+          "prompt": (
+              "Customer feedback: '{review}'\n\n"
+              "As a customer service representative, craft a balanced response that:\n"
+              "- Provides additional relevant product information\n"
+              "- Highlights key service features\n"
+              "- Politely requests more detailed feedback\n"
+              "- Maintains professional tone (3-4 sentences)\n\n"
+              "Response:"
+          )
+      },
+      "sadness": {
+          "prompt": (
+              "Customer disappointment: '{review}'\n\n"
+              "As a customer service representative, craft an empathetic response that:\n"
+              "- Shows genuine understanding of the issue\n"
+              "- Proposes personalized recovery solution\n"
+              "- Offers extended support options\n"
+              "- Maintains positive outlook (3-4 sentences)\n\n"
+              "Response:"
+          )
+      },
+      "surprise": {
+          "prompt": (
+              "Customer enthusiastic feedback: '{review}'\n\n"
+              "As a customer service representative, craft a response that:\n"
+              "- Matches customer's positive energy appropriately\n"
+              "- Highlights unexpected product benefits\n"
+              "- Invites to user community/events\n"
+              "- Maintains brand voice (3-4 sentences)\n\n"
+              "Response:"
+          )
+      }
+  }
+  # 生成回复Prompt
+  template = emotion_strategies[dominant_emotion['label'].lower()]["prompt"]
+  prompt = template.format(review=user_review)
+  print(prompt)
+  return prompt
+
+
+
+def response_gen(user_review):
+  prompt = prompt_gen(user_review)
+  # 加载Llama-3作为text generation模型
+  tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-3.2-1B")
+  model = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-3.2-1B")
+  inputs = tokenizer(prompt, return_tensors="pt")
+  outputs = model.generate(**inputs, max_new_tokens=100)
+
+  input_length = inputs.input_ids.shape[1]
+  response = tokenizer.decode(outputs[0][input_length:], skip_special_tokens=True)
+  # print(response)
+  return response
 
 
 
 ##########################################
 # Step 3：语音生成 - 根据回复合成语音
 ##########################################
+def sound_gen(response):
+  # 加载模型和处理器
+  #processor = SpeechT5Processor.from_pretrained("microsoft/speecht5_tts")
+  #speech_model = SpeechT5ForTextToSpeech.from_pretrained("microsoft/speecht5_tts")
+  #vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan")
+
+  processor = SpeechT5Processor.from_pretrained("microsoft/speecht5_tts")
+  model = SpeechT5ForTextToSpeech.from_pretrained("microsoft/speecht5_tts")
+  vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan")
+
+  # 创建默认的说话人嵌入
+  embeddings_dataset = load_dataset("Matthijs/cmu-arctic-xvectors", split="validation")
+  speaker_embeddings = torch.tensor(embeddings_dataset[7306]["xvector"]).unsqueeze(0)  # 女性中性语音
+
+  # 文本预处理和语音合成
+  inputs = processor(text=response, return_tensors="pt")
+  spectrogram = model.generate_speech(inputs["input_ids"], speaker_embeddings)
 
-# 加载模型和处理器
-#processor = SpeechT5Processor.from_pretrained("microsoft/speecht5_tts")
-#speech_model = SpeechT5ForTextToSpeech.from_pretrained("microsoft/speecht5_tts")
-#vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan")
+  # 使用声码器生成波形音频
+  with torch.no_grad():
+      speech = vocoder(spectrogram)
 
-processor = SpeechT5Processor.from_pretrained("microsoft/speecht5_tts")
-model = SpeechT5ForTextToSpeech.from_pretrained("microsoft/speecht5_tts")
-vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan")
+  # 保存为WAV文件(16kHz采样率)
+  sf.write("customer_service_response.wav", speech.numpy(), samplerate=16000)
 
-# 创建默认的说话人嵌入
-embeddings_dataset = load_dataset("Matthijs/cmu-arctic-xvectors", split="validation")
-speaker_embeddings = torch.tensor(embeddings_dataset[7306]["xvector"]).unsqueeze(0)  # 女性中性语音
+  print("语音生成完成，已保存为 customer_service_response.wav")
+  return
 
-# 文本预处理和语音合成
-inputs = processor(text=response, return_tensors="pt")
-spectrogram = model.generate_speech(inputs["input_ids"], speaker_embeddings)
 
-# 使用声码器生成波形音频
-with torch.no_grad():
-    speech = vocoder(spectrogram)
+    sf.write("customer_service_response.wav", speech.numpy(), samplerate=16000)
+    st.text("I wanna tell you that")
+    st.audio("customer_service_response.wav")
 
-# 保存为WAV文件(16kHz采样率)
-sf.write("customer_service_response.wav", speech.numpy(), samplerate=16000)
-st.text("I wanna tell you that")
-st.audio("customer_service_response.wav")
 
+if __name__ == "__main__":
+    main()