import gradio as gr print("Gradio version:", gr.__version__) import os, time, re import numpy as np import joblib import librosa from huggingface_hub import hf_hub_download from deepface import DeepFace from transformers import pipeline # 如果不手动用 AutoTokenizer/AutoModel，就不必 import AutoTokenizer, AutoModelForSequenceClassification # --- 1. 加载 SVM 语音模型 --- print("Downloading SVM model from Hugging Face Hub...") model_path = hf_hub_download(repo_id="GCLing/emotion-svm-model", filename="svm_emotion_model.joblib") print(f"SVM model downloaded to: {model_path}") svm_model = joblib.load(model_path) print("SVM model loaded.") # --- 2. 文本情绪分析：规则+zero-shot --- try: zero_shot = pipeline("zero-shot-classification", model="joeddav/xlm-roberta-large-xnli") except Exception as e: print("加载 zero-shot pipeline 失败：", e) zero_shot = None 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" } # 关键词列表：注意繁简体一致，或可添加两种形式 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"], "disgust": ["噁心","厭惡","反感"] # 如需“厭惡”等 } # 否定词列表 negations = ["不","沒","沒有","別","勿","非"] def keyword_emotion(text: str): """ 规则方法：扫描 emo_keywords，处理前置否定词。 返回 None 或 {} 表示规则未命中；否则返回非空 dict，例如 {'angry': 2, 'sad':1} 或归一化 {'angry':0.67,'sad':0.33}。 """ if not text or text.strip() == "": return None text_proc = text.strip() # 中文不需要 lower counts = {emo: 0 for emo in emo_keywords} for emo, kws in emo_keywords.items(): for w in kws: idx = text_proc.find(w) if idx != -1: # 检查前一到两字符是否否定词 neg = False for neg_word in negations: plen = len(neg_word) if idx - plen >= 0 and text_proc[idx-plen:idx] == neg_word: neg = True break if not neg: counts[emo] += 1 else: # 若否定，可选择减分或忽略；这里忽略 pass total = sum(counts.values()) if total > 0: # 归一化 return {emo: counts[emo] / total for emo in counts if counts[emo] > 0} else: return None def predict_text_mixed(text: str): """ 文本情绪分析：先规则，若规则命中返回最高情绪及其比例；否则fallback zero-shot返回多类别分布。 返回 dict[str, float]，供 Gradio Label 显示。 """ print("predict_text_mixed called, text:", repr(text)) if not text or text.strip() == "": print("輸入為空，返回空") return {} # 规则优先 res = keyword_emotion(text) print("keyword_emotion result:", res) if res: # 只返回最高项：也可返回完整分布 res top_emo = max(res, key=res.get) # 例如 "angry" mapping = { "happy": "高興", "angry": "憤怒", "sad": "悲傷", "surprise": "驚訝", "fear": "恐懼", "disgust": "厭惡" } cn = mapping.get(top_emo, top_emo) prob = float(res[top_emo]) print(f"使用規則方法，返回: {{'{cn}': {prob}}}") return {cn: prob} # 规则未命中，zero-shot fallback if zero_shot is None: print("zero_shot pipeline 未加载，返回中性") return {"中性": 1.0} try: out = zero_shot(text, candidate_labels=candidate_labels, hypothesis_template="这句話表達了{}情緒") print("zero-shot 返回:", out) result = {} for lab, sc in zip(out["labels"], out["scores"]): cn = label_map_en2cn.get(lab.lower(), lab) result[cn] = float(sc) print("zero-shot 结果映射中文:", result) return result except Exception as e: print("zero-shot error:", e) return {"中性": 1.0} # --- 3. 语音情绪预测函数 --- def extract_feature(signal: np.ndarray, sr: int) -> np.ndarray: mfcc = librosa.feature.mfcc(y=signal, sr=sr, n_mfcc=13) return np.concatenate([np.mean(mfcc, axis=1), np.var(mfcc, axis=1)]) def predict_voice(audio_path: str): if not audio_path: print("predict_voice: 无 audio_path，跳过") return {} try: signal, sr = librosa.load(audio_path, sr=None) feat = extract_feature(signal, sr) probs = svm_model.predict_proba([feat])[0] labels = svm_model.classes_ return {labels[i]: float(probs[i]) for i in range(len(labels))} except Exception as e: print("predict_voice error:", e) return {} # --- 4. 人脸情绪预测函数 --- def predict_face(img: np.ndarray): print("predict_face called, img is None?", img is None) if img is None: return {} try: res = DeepFace.analyze(img, actions=["emotion"], detector_backend="opencv") if isinstance(res, list): first = res[0] if res else {} emo = first.get("emotion", {}) if isinstance(first, dict) else {} else: emo = res.get("emotion", {}) if isinstance(res, dict) else {} # 转 float，确保 JSON 可序列化 emo_fixed = {k: float(v) for k, v in emo.items()} print("predict_face result:", emo_fixed) return emo_fixed except Exception as e: print("DeepFace.analyze error:", e) return {} # --- 5. Gradio 界面 --- with gr.Blocks() as demo: gr.Markdown("## 多模態情緒分析示例") with gr.Tabs(): # 臉部情緒 Tab with gr.TabItem("臉部情緒"): gr.Markdown("### 臉部情緒 (即時 Webcam Streaming 分析)") with gr.Row(): webcam = gr.Image(sources="webcam", streaming=True, type="numpy", label="攝像頭畫面") face_out = gr.JSON(label="情緒原始結果") webcam.stream(fn=predict_face, inputs=webcam, outputs=face_out) # 語音情緒 Tab with gr.TabItem("語音情緒"): gr.Markdown("### 語音情緒 分析") with gr.Row(): audio = gr.Audio(sources="microphone", streaming=False, type="filepath", label="錄音") voice_out = gr.Label(label="語音情緒結果") audio.change(fn=predict_voice, inputs=audio, outputs=voice_out) # 文字情緒 Tab with gr.TabItem("文字情緒"): gr.Markdown("### 文字情緒 分析 (規則+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") # 可以用回車提交 btn = gr.Button("分析") btn.click(fn=predict_text_mixed, inputs=text, outputs=text_out) return demo if __name__ == "__main__": demo = build_interface() demo.launch()