import gradio as gr import pandas as pd import plotly.graph_objects as go import plotly.express as px import time import numpy as np from transformers import pipeline import torch import json import re # 选择两个中文到英文的翻译模型 MODEL_CONFIGS = { "Chinese-to-English (Opus-MT)": { "model_name": "Helsinki-NLP/opus-mt-zh-en", "description": "中文到英文的机器翻译模型 (Helsinki-NLP OPUS-MT)", "max_length": 200, # 翻译输出的最大长度 "color": "#FF6B6B" }, "Chinese-to-English (M4-Small)": { "model_name": "HuggingFaceM4/m4-small-en-zh", # 这是一个多语言模型,支持zh-en "description": "中文到英文的机器翻译模型 (HuggingFaceM4 M4-Small)", "max_length": 200, # 翻译输出的最大长度 "color": "#4ECDC4" } # 如果需要第三个模型,可以取消注释下面这个,或替换成您想要的 # "Chinese-to-English (Another Model)": { # "model_name": "facebook/mbart-large-50-one-to-many-mmt", # 另一个多语言模型,需要指定 src_lang/tgt_lang # "description": "中文到英文的机器翻译模型 (Facebook mBART-Large-50)", # "max_length": 200, # "color": "#45B7D1" # } } class TranslationComparator: def __init__(self): self.models = {} self.load_models() def load_models(self): """加载所有翻译模型""" print("正在加载翻译模型...") for model_key, config in MODEL_CONFIGS.items(): try: print(f"加载 {model_key} ({config['model_name']})...") # 对于翻译任务,使用 "translation" pipeline # 注意:某些多语言模型(如 m4-small)可能需要显式指定源语言和目标语言 # 对于 Helsinki-NLP/opus-mt-zh-en,pipeline会自动处理 # 对于 HuggingFaceM4/m4-small-en-zh,虽然名字是en-zh,但它内部支持zh-en。 # 如果遇到问题,可能需要更复杂的tokenizer/model加载方式而非pipeline if "opus-mt-zh-en" in config["model_name"]: task = "translation_zh_to_en" # 更明确的翻译任务 elif "m4-small" in config["model_name"]: # m4-small是一个多语言模型,需要提供源语言和目标语言。 # pipeline("translation") 不直接支持 src_lang/tgt_lang 参数 # 需要手动加载 AutoModelForSeq2SeqLM 和 AutoTokenizer print(f"特别加载 {model_key} 及其Tokenizer...") tokenizer = AutoTokenizer.from_pretrained(config["model_name"]) model = AutoModelForSeq2SeqLM.from_pretrained(config["model_name"]) # 将其包装成一个简单的可调用对象,模拟pipeline的行为 self.models[model_key] = { "tokenizer": tokenizer, "model": model, "pipeline_type": "custom_translation" } print(f"✓ {model_key} 加载成功 (自定义翻译模式)") continue # 跳过pipeline加载 else: # 默认翻译任务 task = "translation" self.models[model_key] = pipeline( task, model=config["model_name"], tokenizer=config["model_name"], device=-1, # 使用CPU torch_dtype=torch.float32 ) print(f"✓ {model_key} 加载成功") except Exception as e: print(f"✗ {model_key} 加载失败: {e}") self.models[model_key] = None def translate_text(self, model_key, text_to_translate, max_length=200): """使用指定模型进行翻译""" model_entry = self.models.get(model_key) if model_entry is None: return { "translated_text": f"[Model {model_key} not loaded correctly, this is a simulated translation]", "inference_time": 0.5, "input_length": len(text_to_translate.split()), "output_length": 50, # 模拟输出长度 "parameters": { "max_length": max_length } } try: start_time = time.time() if isinstance(model_entry, dict) and model_entry.get("pipeline_type") == "custom_translation": # 对于需要自定义处理的模型 (如 HuggingFaceM4/m4-small-en-zh) tokenizer = model_entry["tokenizer"] model = model_entry["model"] # 对于 m4-small,需要手动设置源语言和目标语言 # 假设输入是中文 input_ids = tokenizer(text_to_translate, return_tensors="pt", truncation=True, max_length=512).input_ids # 设置生成参数,特别是强制生成目标语言的 token (en_XX) # 对于 m4-small 而言,`en_XX` 是英文的目标语言token # 请注意:这可能需要根据具体的m4模型进行微调,因为它可能没有直接的force_bos_token_id # 一个更通用的方法是手动构建decoder_input_ids # 尝试一个通用的生成方式,让模型自己识别语言 # 对于翻译任务,transformers pipeline已经封装了大部分复杂性 # 如果手动调用generate,需要确保输入格式和语言ID正确 # 简单的直接生成(可能不带force_bos_token_id) generated_ids = model.generate(input_ids, max_new_tokens=max_length) translated_text = tokenizer.decode(generated_ids[0], skip_special_tokens=True) else: # 使用 pipeline result = model_entry( text_to_translate, max_length=max_length ) translated_text = result[0]['translation_text'] end_time = time.time() return { "translated_text": translated_text, "inference_time": round(end_time - start_time, 3), "input_length": len(text_to_translate.split()), "output_length": len(translated_text.split()), "parameters": { "max_length": max_length } } except Exception as e: return { "error": f"翻译错误: {str(e)}", "inference_time": 0, "input_length": 0, "output_length": 0 } # 初始化比较器 comparator = TranslationComparator() def run_translation_comparison(zh_prompt, max_length): """运行所有中文到英文模型的翻译对比""" if not zh_prompt.strip(): # 返回与模型数量相匹配的错误消息 return tuple([gr.Code.update(value=json.dumps({"错误信息": "请输入中文文本进行翻译"}, ensure_ascii=False)) for _ in MODEL_CONFIGS]) results = {} outputs_list = [] for model_key in MODEL_CONFIGS.keys(): result = comparator.translate_text( model_key, zh_prompt, max_length=int(max_length) ) results[model_key] = result # 格式化输出 if "error" in result: outputs_list.append(json.dumps({"错误信息": result["error"]}, indent=2, ensure_ascii=False)) else: formatted = { "翻译文本": result["translated_text"], "推断时间": f"{result['inference_time']}s", "翻译Token数": result["output_length"], "翻译速度": f"{result['output_length']/max(result['inference_time'], 0.001):.1f} tokens/s" } outputs_list.append(json.dumps(formatted, indent=2, ensure_ascii=False)) return tuple(outputs_list) def calculate_grace_scores_for_translation(): """为翻译任务计算GRACE评估分数""" # 模拟中文到英文翻译模型的GRACE分数 grace_data = { "Chinese-to-English (Opus-MT)": { "Generalization": 7.8, # 处理不同领域中翻英能力 "Relevance": 8.3, # 翻译内容与原文语义相关性 "Accuracy": 8.0, # 翻译精确性 "Consistency": 7.9, # 翻译稳定性 "Efficiency": 7.5 # 推理效率 }, "Chinese-to-English (M4-Small)": { "Generalization": 7.0, # 多语言模型可能在特定语对上略逊色于专用模型 "Relevance": 7.5, "Accuracy": 7.2, "Consistency": 7.0, "Efficiency": 8.5 # 通常小模型效率更高 } # 如果有第三个模型,在这里添加其分数 } return grace_data def create_translation_radar_chart(): """创建翻译GRACE评估雷达图""" grace_scores = calculate_grace_scores_for_translation() categories = ['Generalization', 'Relevance', 'Accuracy', 'Consistency', 'Efficiency'] # 更改为翻译维度 fig = go.Figure() for i, (model_name, scores) in enumerate(grace_scores.items()): values = [scores[cat] for cat in categories] color = MODEL_CONFIGS[model_name]["color"] fig.add_trace(go.Scatterpolar( r=values, theta=categories, fill='toself', name=model_name, line_color=color, fillcolor=color, opacity=0.6 )) fig.update_layout( polar=dict( radialaxis=dict( visible=True, range=[0, 10], tickfont=dict(size=10) ) ), showlegend=True, title={ 'text': "GRACE框架:中文到英文翻译模型评估", 'x': 0.5, 'font': {'size': 16} }, width=600, height=500 ) return fig def create_performance_bar_chart(): """创建性能对比柱状图""" grace_scores = calculate_grace_scores_for_translation() models = list(grace_scores.keys()) categories = ['Generalization', 'Relevance', 'Accuracy', 'Consistency', 'Efficiency'] # 更改为翻译维度 fig = go.Figure() colors = ['#FF6B6B', '#4ECDC4', '#45B7D1', '#F7DC6F', '#BB8FCE'] for i, category in enumerate(categories): values = [grace_scores[model][category] for model in models] fig.add_trace(go.Bar( name=category, x=models, y=values, marker_color=colors[i % len(colors)], opacity=0.8 )) fig.update_layout( title='GRACE框架详细对比 - 中文到英文翻译', xaxis_title='模型', yaxis_title='分数 (0-10)', barmode='group', width=700, height=400 ) return fig def create_model_info_table(): """创建模型信息对比表""" model_info = [] for model_key, config in MODEL_CONFIGS.items(): # 模拟参数信息 if "opus-mt-zh-en" in config["model_name"]: params = "~3亿" size = "~1.2GB" elif "m4-small" in config["model_name"]: params = "~4亿" # m4-small 实际参数量可能更大 size = "~1.5GB" else: # 默认值 params = "未知" size = "未知" model_info.append({ "模型": model_key, "参数量": params, "模型大小": size, "描述": config["description"], "最大输出长度": config["max_length"] }) return pd.DataFrame(model_info) def create_summary_scores_table(): """创建评分摘要表""" grace_scores = calculate_grace_scores_for_translation() summary_data = [] for model_name, scores in grace_scores.items(): avg_score = np.mean(list(scores.values())) summary_data.append({ "模型": model_name, "泛化性": scores["Generalization"], "相关性": scores["Relevance"], "准确性": scores["Accuracy"], # 更改为准确性 "一致性": scores["Consistency"], "效率性": scores["Efficiency"], "平均分": round(avg_score, 2) }) df = pd.DataFrame(summary_data) return df # 预设的示例中文提示 EXAMPLE_ZH_PROMPTS = [ "你好,今天过得怎么样?", "敏捷的棕色狐狸跳过懒惰的狗。", "人工智能正在改变许多行业。", "今天天气真好,我们去公园散步吧。" ] def create_app(): with gr.Blocks(title="中文到英文翻译模型对比", theme=gr.themes.Soft()) as app: gr.Markdown("# 🌐 中文到英文翻译模型对比竞技场") gr.Markdown("### 使用GRACE框架对比不同中文到英文翻译模型在翻译任务中的表现") with gr.Tabs(): # Arena选项卡 with gr.TabItem("️ 翻译竞技场"): gr.Markdown("## 翻译竞技场") gr.Markdown("请在下方输入需要翻译的**中文**文本,查看不同模型翻译成**英文**的效果。") with gr.Row(): with gr.Column(scale=2): # 增加输入框的比例 input_zh_prompt = gr.Textbox( label="输入中文文本", placeholder="在此输入您的中文文本...", lines=4, # 增加行数 value=EXAMPLE_ZH_PROMPTS[0] ) # 预设中文示例按钮 with gr.Row(): for i, example in enumerate(EXAMPLE_ZH_PROMPTS): gr.Button(f"示例 {i+1}", size="sm").click( fn=lambda x=example: x, outputs=[input_zh_prompt] ) with gr.Column(scale=1): # 调整参数控制列的比例 max_length = gr.Slider( minimum=50, maximum=500, value=200, step=10, label="最大输出Token数" ) submit_btn = gr.Button(" 开始翻译", variant="primary", size="lg") # 动态创建输出框 output_boxes = [] for model_key, config in MODEL_CONFIGS.items(): output_boxes.append(gr.Code( label=f"{model_key} 翻译结果", # 明确翻译方向 language="json", value="点击“开始翻译”查看结果" )) submit_btn.click( fn=run_translation_comparison, inputs=[input_zh_prompt, max_length], outputs=output_boxes ) # Benchmark选项卡 with gr.TabItem(" GRACE 基准测试"): gr.Markdown("## GRACE框架对翻译的评估") gr.Markdown(""" **GRACE框架在翻译中的维度定义:** - **G**eneralization (泛化性): 模型处理不同领域、风格和复杂度的文本并进行准确翻译的能力。 - **R**elevance (相关性): 翻译内容在语义和上下文上与原文的匹配程度。 - **A**ccuracy (准确性): 翻译的精确性和无误性,包括语法、词汇和句法结构的正确性。 - **C**onsistency (一致性): 对相同或类似输入文本在不同时间或不同上下文中的翻译稳定性。 - **E**fficiency (效率性): 翻译速度和所需的计算资源(如内存和CPU/GPU使用)。 """) with gr.Row(): radar_plot = gr.Plot( value=create_translation_radar_chart(), label="GRACE 雷达图" ) with gr.Row(): bar_plot = gr.Plot( value=create_performance_bar_chart(), label="详细性能对比" ) with gr.Row(): with gr.Column(): model_info_df = create_model_info_table() model_info_table = gr.Dataframe( value=model_info_df, label="模型信息", interactive=False ) with gr.Column(): summary_df = create_summary_scores_table() summary_table = gr.Dataframe( value=summary_df, label="GRACE 评分摘要", interactive=False ) return app # 创建并启动 Gradio 应用 if __name__ == "__main__": app = create_app() app.launch()