""" Gradio App for BERT Traditional Chinese Classifier 用於測試繁體中文分類模型（大陸繁體 vs 台灣繁體）的互動式介面 """ import gradio as gr import torch import torch.nn.functional as F from transformers import AutoTokenizer, AutoModelForSequenceClassification from collections import Counter # ========== 配置 ========== REPO_ID = "renhehuang/bert-traditional-chinese-classifier" LABELS = {0: "大陸繁體", 1: "台灣繁體"} MAX_LEN, STRIDE = 384, 128 # ========== 初始化模型 ========== print("🔄 載入模型與 tokenizer...") device = ( "mps" if torch.backends.mps.is_available() else ("cuda" if torch.cuda.is_available() else "cpu") ) tokenizer = AutoTokenizer.from_pretrained(REPO_ID, cache_dir=".cache") model = AutoModelForSequenceClassification.from_pretrained(REPO_ID, cache_dir=".cache") model.to(device).eval() print(f"✅ 模型已載入至 {device}") # ========== 工具函數 ========== def chunk_encode(text, max_len=MAX_LEN, stride=STRIDE): """長文本分塊編碼""" ids = tokenizer(text, add_special_tokens=False, return_attention_mask=False)["input_ids"] if len(ids) <= max_len - 2: enc = tokenizer(text, truncation=True, max_length=max_len, return_attention_mask=True, return_tensors="pt") return [enc] enc = tokenizer(text, truncation=True, max_length=max_len, stride=stride, return_overflowing_tokens=True, return_attention_mask=True, return_tensors="pt") return [{"input_ids": enc["input_ids"][i:i+1], "attention_mask": enc["attention_mask"][i:i+1]} for i in range(len(enc["input_ids"]))] @torch.inference_mode() def predict_single(text: str): """單次推論""" if not text or not text.strip(): return "⚠️ 請輸入文本", "", {} chunks = chunk_encode(text) probs_all = [] for ch in chunks: logits = model( input_ids=ch["input_ids"].to(device), attention_mask=ch["attention_mask"].to(device) ).logits probs_all.append(F.softmax(logits, dim=-1).cpu()) avg = torch.cat(probs_all, 0).mean(0) label_id = int(avg.argmax()) confidence = float(avg[label_id]) # 格式化輸出 result_text = f"🏷️ **{LABELS[label_id]}**" confidence_text = f"📊 信心度: **{confidence:.2%}**" probabilities = { "大陸繁體": float(avg[0]), "台灣繁體": float(avg[1]) } return result_text, confidence_text, probabilities @torch.inference_mode() def predict_voting(text: str, n_runs: int = 3): """多次投票推論（MC Dropout）""" if not text or not text.strip(): return "⚠️ 請輸入文本", "", {}, "" chunks = chunk_encode(text) prev_training = model.training run_prob_list = [] try: model.train() # 啟用 dropout for _ in range(n_runs): probs_all = [] for ch in chunks: logits = model( input_ids=ch["input_ids"].to(device), attention_mask=ch["attention_mask"].to(device) ).logits probs_all.append(F.softmax(logits, dim=-1).cpu()) run_prob_list.append(torch.cat(probs_all, 0).mean(0)) finally: model.train() if prev_training else model.eval() probs_stack = torch.stack(run_prob_list, 0) per_run_ids = probs_stack.argmax(-1).tolist() vote_counts = Counter(per_run_ids) mean_probs = probs_stack.mean(0) voted_id = max(vote_counts.items(), key=lambda kv: (kv[1], mean_probs[kv[0]].item()))[0] confidence = float(mean_probs[voted_id]) # 格式化輸出 result_text = f"🏷️ **{LABELS[voted_id]}**" confidence_text = f"📊 平均信心度: **{confidence:.2%}**" probabilities = { "大陸繁體": float(mean_probs[0]), "台灣繁體": float(mean_probs[1]) } vote_info = f"🗳️ 投票結果: {vote_counts[voted_id]}/{n_runs} 次" return result_text, confidence_text, probabilities, vote_info # ========== Gradio 介面 ========== examples = [ ["這個軟件的界面設計得很好。"], ["這個軟體的介面設計得很好。"], ["我需要下載這個程序到計算機上。"], ["我需要下載這個程式到電腦上。"], ] with gr.Blocks(title="BERT 繁體中文分類器", theme=gr.themes.Soft()) as demo: gr.Markdown(""" # 🔍 BERT 繁體中文分類器 區分「大陸繁體」與「台灣繁體」的 BERT 分類模型 - 支援長文本自動分塊處理（max_len=384） - 提供單次推論與多次投票（MC Dropout）模式 """) with gr.Tab("📝 單次推論"): with gr.Row(): with gr.Column(): input_single = gr.Textbox( label="輸入文本", placeholder="請輸入繁體中文文本...", lines=5 ) btn_single = gr.Button("🚀 開始分類", variant="primary") with gr.Column(): output_label_single = gr.Markdown(label="預測結果") output_conf_single = gr.Markdown(label="信心度") output_probs_single = gr.Label(label="機率分布", num_top_classes=2) gr.Examples( examples=examples, inputs=input_single, label="範例文本" ) btn_single.click( fn=predict_single, inputs=input_single, outputs=[output_label_single, output_conf_single, output_probs_single] ) with gr.Tab("🗳️ 投票推論（MC Dropout）"): with gr.Row(): with gr.Column(): input_voting = gr.Textbox( label="輸入文本", placeholder="請輸入繁體中文文本...", lines=5 ) n_runs = gr.Slider( minimum=3, maximum=10, value=3, step=1, label="投票次數", info="推論次數越多，結果越穩定但速度較慢" ) btn_voting = gr.Button("🚀 開始投票分類", variant="primary") with gr.Column(): output_label_voting = gr.Markdown(label="預測結果") output_conf_voting = gr.Markdown(label="平均信心度") output_probs_voting = gr.Label(label="平均機率分布", num_top_classes=2) output_vote_info = gr.Markdown(label="投票統計") gr.Examples( examples=examples, inputs=input_voting, label="範例文本" ) btn_voting.click( fn=predict_voting, inputs=[input_voting, n_runs], outputs=[output_label_voting, output_conf_voting, output_probs_voting, output_vote_info] ) with gr.Tab("ℹ️ 關於模型"): gr.Markdown(""" ## 模型資訊 - **模型**: ckiplab/bert-base-chinese - **任務**: 繁體中文文本分類（大陸繁體 vs 台灣繁體） - **準確率**: 87.71% - **訓練樣本**: 156,824 ## 標籤定義 - **大陸繁體（中國繁體）**: 使用「软件、视频、程序、计算机」等詞彙 - **台灣繁體**: 使用「軟體、影片、程式、電腦」等詞彙 ## 功能特色 - ✅ 長文本自動分塊處理（384 tokens，stride 128） - ✅ Focal Loss 處理類別不平衡 - ✅ Multi-Sample Dropout 提升泛化 - ✅ MC Dropout 投票提升穩健性 ## 使用建議 - 對於重要決策，建議使用「投票推論」模式並設定 5-10 次投票 - 信心度 ≥ 85% 的預測較為可靠 - 混用詞彙、專業術語或極短文本可能影響準確度 --- 📦 **模型倉庫**: [renhehuang/bert-traditional-chinese-classifier](https://huggingface.co/renhehuang/bert-traditional-chinese-classifier) 📄 **授權**: Apache 2.0 """) if __name__ == "__main__": demo.launch( share=False, server_name="0.0.0.0", server_port=7860, show_error=True )