import gradio as gr from openai import OpenAI from huggingface_hub import InferenceClient import os import time import requests import urllib.parse import pandas as pd from langchain_community.document_loaders import DataFrameLoader from langchain_text_splitters import RecursiveCharacterTextSplitter from langchain_huggingface import HuggingFaceEmbeddings from langchain_community.vectorstores import FAISS from PIL import Image, ImageFont, ImageDraw import warnings import cv2 import numpy as np from deepface import DeepFace # 忽略 pandas 的一些警告 warnings.filterwarnings("ignore") # ========================================== # 0. 環境變數與 RAG 初始化 (維持不變) # ========================================== GROQ_API_KEY = os.getenv("GROQ_API_KEY") HF_TOKEN = os.getenv("HF_TOKEN") # --- 資料讀取 --- df = pd.DataFrame() try: df = pd.read_csv('restaurants.csv', engine='python') df.columns = df.columns.str.strip() def classify_mood(row): name_str = str(row.get('Name', '')) rag_str = str(row.get('RAG_Content', '')) text = (name_str + " " + rag_str).lower() tags = [] rules = { "開心/慶祝": ["牛排", "steak", "pizza", "炸", "甜點", "蛋糕", "buffet", "bistro"], "傷心/疲憊": ["粥", "湯", "warm", "麵", "小吃", "豆花", "soup"], "生氣/發洩": ["辣", "spicy", "麻辣", "鍋", "curry", "燒肉", "bbq", "fry"], "平靜/放鬆": ["cafe", "coffee", "tea", "素食", "brunch", "sushi", "居酒屋"] } for mood, keywords in rules.items(): for kw in keywords: if kw in text: tags.append(mood) if not tags: tags.append("隨意/探索") return ", ".join(list(set(tags))) if 'Mood_Tags' not in df.columns: df['Mood_Tags'] = df.apply(classify_mood, axis=1) print("✅ 成功讀取 restaurants.csv") except Exception as e: print(f"⚠️ 讀取 restaurants.csv 失敗: {e}") # --- Prompt 資料庫 --- df_prompts = pd.DataFrame() PROMPT_COL_NAME = 'Visual_prompt' try: if not df.empty: df_prompts = df.copy() if 'Name' in df_prompts.columns: df_prompts.set_index('Name', inplace=True) except: pass # --- RAG --- retriever = None if not df.empty and 'RAG_Content' in df.columns: try: loader = DataFrameLoader(df, page_content_column="RAG_Content") documents = loader.load() text_splitter = RecursiveCharacterTextSplitter(chunk_size=500, chunk_overlap=50) docs = text_splitter.split_documents(documents) embeddings = HuggingFaceEmbeddings(model_name="sentence-transformers/all-MiniLM-L6-v2") vectorstore = FAISS.from_documents(docs, embeddings) retriever = vectorstore.as_retriever(search_kwargs={"k": 3}) except: pass # ========================================== # 1. 核心功能函式 (餐廳推薦 - 維持不變) # ========================================== def get_restaurant_data(mood_score_str, food_choice): if df.empty: return None, True score_map = { "1 (心情差)": ["傷心/疲憊", "生氣/發洩"], "2 (不太好)": ["傷心/疲憊", "生氣/發洩"], "3 (普通)": ["平靜/放鬆", "隨意/探索"], "4 (不錯)": ["開心/慶祝", "平靜/放鬆"], "5 (超棒)": ["開心/慶祝"] } target_moods = score_map.get(mood_score_str, []) food_keyword = "" if food_choice == "吃飯": food_keyword = "飯" elif food_choice == "吃麵": food_keyword = "麵" candidates = df.copy() if target_moods: pattern = "|".join(target_moods) if 'Mood_Tags' in candidates.columns: candidates = candidates[candidates['Mood_Tags'].str.contains(pattern, regex=True, na=False)] if food_keyword: candidates = candidates[ candidates['Name'].str.contains(food_keyword, case=False, na=False) | candidates['RAG_Content'].str.contains(food_keyword, case=False, na=False) ] if candidates.empty: return df.sample(1).iloc[0], True return candidates.sample(1).iloc[0], False def generate_content_with_groq(restaurant_name, restaurant_detail, user_diary, mood_score): if not GROQ_API_KEY: return "⚠️ 請設定 GROQ_API_KEY" client = OpenAI(api_key=GROQ_API_KEY, base_url="https://api.groq.com/openai/v1") system_prompt = "你是一個幽默、懂吃且善解人意的 AI 朋友。請根據使用者的日記和心情來推薦餐廳。" user_msg = f"心情：{mood_score}\n日記：{user_diary}\n餐廳：{restaurant_name}\n資料：{restaurant_detail}\n請推薦這家餐廳並給予安慰或慶祝。" try: response = client.chat.completions.create( model="llama-3.3-70b-versatile", messages=[{"role": "system", "content": system_prompt}, {"role": "user", "content": user_msg}] ) return response.choices[0].message.content except Exception as e: return f"Groq Error: {str(e)}" def generate_image_huggingface(prompt): if not HF_TOKEN: return None try: hf_client = InferenceClient(token=HF_TOKEN) return hf_client.text_to_image(prompt=prompt, model="stabilityai/stable-diffusion-xl-base-1.0") except: return None def mood_agent_logic(score_input, food_input, diary_input, debug_mode): if not score_input: yield "⚠️ 請先選擇心情分數！", None, "" return restaurant, is_random = get_restaurant_data(score_input, food_input) if restaurant is None: yield "資料庫讀取錯誤", None, "" return name = str(restaurant['Name']) url = str(restaurant.get('URL', f'https://www.google.com/maps/search/?api=1&query={urllib.parse.quote(name)}')) rag_info = str(restaurant.get('RAG_Content', '')) if retriever: try: docs = retriever.invoke(name) if docs: rag_info = "\n".join([d.page_content for d in docs]) except: pass img_prompt = f"Delicious food from {name}, cinematic lighting, 8k" if not df_prompts.empty and name in df_prompts.index: try: val = df_prompts.loc[name, PROMPT_COL_NAME] if isinstance(val, pd.Series): val = val.iloc[0] if pd.notna(val): img_prompt = str(val) except: pass ai_text = generate_content_with_groq(name, rag_info, diary_input, score_input) debug_text = f"\n\nPrompt: `{img_prompt}`" if debug_mode else "" final_response = f"### 🍽️ 推薦：{name}\n\n{ai_text}{debug_text}" map_html = f'
🗺️ Google Map 導航
' yield final_response, None, map_html image_output = generate_image_huggingface(img_prompt) yield final_response, image_output, map_html # ========================================== # 2. 情緒辨識整合模組 # ========================================== emotion_text_obj = {'angry': '生氣', 'disgust': '噁心', 'fear': '害怕', 'happy': '開心', 'sad': '難過', 'surprise': '驚訝', 'neutral': '正常'} def putText(img, x, y, text, size=50, color=(255, 255, 255)): try: fontpath = 'NotoSansTC-VariableFont_wght.ttf' if not os.path.exists(fontpath): return img font = ImageFont.truetype(fontpath, size) imgPil = Image.fromarray(img) draw = ImageDraw.Draw(imgPil) displayText = emotion_text_obj.get(text, text) draw.text((x, y), displayText, fill=color, font=font) return np.array(imgPil) except: return img # 核心辨識函式 (重構以供重複使用) def process_emotion(frame): detected_emotion = "neutral" mapped_score = "3 (普通)" try: analyze = DeepFace.analyze(frame, actions=['emotion'], enforce_detection=False) if isinstance(analyze, list): analyze = analyze[0] detected_emotion = analyze['dominant_emotion'] frame = putText(frame, 20, 40, detected_emotion) if detected_emotion == 'happy': mapped_score = "5 (超棒)" elif detected_emotion == 'surprise': mapped_score = "4 (不錯)" elif detected_emotion == 'neutral': mapped_score = "3 (普通)" elif detected_emotion in ['sad', 'fear']: mapped_score = "2 (不太好)" elif detected_emotion in ['angry', 'disgust']: mapped_score = "1 (心情差)" except: pass return frame, mapped_score # 單次手動截圖 def manual_detect(frame): if frame is None: return None, None return process_emotion(frame) # ★ 新增：切換錄製狀態按鈕邏輯 def toggle_recording_state(is_recording): # 切換 True/False new_state = not is_recording # 根據新狀態改變按鈕文字 if new_state: return new_state, "⏹️ 停止錄製 (即時偵測中)" else: return new_state, "🔴 開始錄製 (即時偵測)" # ★ 新增：自動串流偵測 (包含限速邏輯) def auto_stream_detect(frame, is_recording, last_time): # 如果沒開錄製，或畫面空的，就跳過不做事 (節省資源) if not is_recording or frame is None: return gr.skip(), gr.skip(), last_time # 限速：距離上次執行不到 1.0 秒，就跳過 current_time = time.time() if current_time - last_time < 1.0: return gr.skip(), gr.skip(), last_time # 執行辨識 processed_frame, score = process_emotion(frame) # 回傳：更新圖片, 更新分數, 更新最後執行時間 return processed_frame, score, current_time # ========================================== # 3. Gradio 介面建構 # ========================================== with gr.Blocks(title="AI 心情食堂") as demo: # 定義狀態變數 (看不見的變數) is_recording = gr.State(False) # 是否正在錄製 last_process_time = gr.State(0) # 上次偵測的時間 with gr.Column(visible=True) as main_app_col: gr.Markdown(f"## 🍱 AI 心情食堂導航") gr.Markdown("請看著鏡頭，讓 AI 幫你判斷今天的心情分數！") with gr.Row(): # 左側：輸入區 with gr.Column(scale=1): gr.Markdown("### 📸 步驟 1：心情偵測") with gr.Row(): # 左邊：即時鏡頭 webcam_input = gr.Image(sources=["webcam"], label="即時鏡頭", streaming=True) # 右邊：偵測結果 captured_image = gr.Image(label="偵測結果", interactive=False) with gr.Row(): # 按鈕 1：錄製開關 (會變色變字) record_btn = gr.Button("🔴 開始錄製 (即時偵測)", variant="stop") # 按鈕 2：單次截圖 (維持不變) snapshot_btn = gr.Button("📸 單次截圖", variant="secondary") gr.Markdown("### 📝 步驟 2：確認與補充") score_input = gr.Radio( ["1 (心情差)", "2 (不太好)", "3 (普通)", "4 (不錯)", "5 (超棒)"], label="1. 心情分數 (AI 自動填入)", value="3 (普通)" ) food_input = gr.Radio(["吃飯", "吃麵", "隨便"], label="2. 想吃什麼", value="隨便") diary_input = gr.Textbox(lines=3, label="3. 心情日記", placeholder="選填...") debug_mode_btn = gr.Checkbox(label="🔧 除錯資訊", value=False) submit_btn = gr.Button("🚀 送出給 Agent", variant="primary") # 右側：輸出區 with gr.Column(scale=1): agent_output = gr.Markdown(label="AI 回應") image_output = gr.Image(label="AI 推薦美食圖", type="pil", width=400) map_output = gr.HTML(label="地圖導航") # --- 事件綁定 --- # 1. 錄製按鈕邏輯：點擊 -> 切換狀態 -> 改變按鈕文字 record_btn.click( fn=toggle_recording_state, inputs=[is_recording], outputs=[is_recording, record_btn] ) # 2. 串流偵測邏輯：鏡頭畫面改變 -> 檢查狀態 -> (如果錄製中) 執行偵測 -> 更新右邊圖片 webcam_input.stream( fn=auto_stream_detect, inputs=[webcam_input, is_recording, last_process_time], outputs=[captured_image, score_input, last_process_time], show_progress=False # 隱藏進度條，讓畫面更順暢 ) # 3. 單次截圖邏輯 (維持不變) snapshot_btn.click( fn=manual_detect, inputs=[webcam_input], outputs=[captured_image, score_input] ) # 4. 送出 Agent submit_btn.click( fn=mood_agent_logic, inputs=[score_input, food_input, diary_input, debug_mode_btn], outputs=[agent_output, image_output, map_output] ) if __name__ == "__main__": demo.launch()