refine vis

app.py

@@ -1,9 +1,95 @@
 import gradio as gr
 import pandas as pd
-import pandas as pd
 import json
 import plotly.express as px
 
+from textblob import TextBlob
+from textblob.download_corpora import download_all
+
+# 下载TextBlob所需数据（只需运行一次）
+download_all()
+
+# 定义颜色映射
+ADJECTIVE_COLORS = {
+    "positive": "#4CAF50",  # 绿色
+    "negative": "#F44336",  # 红色
+    "neutral": "#FFC107"    # 黄色
+}
+
+# 自定义短语情感覆盖规则
+PHRASE_SENTIMENT_OVERRIDES = {
+    "significant drop": "negative",
+    "significant drops": "negative",
+    "sharp decline": "negative",
+    "strong performance": "positive",
+    "Poor performance": "negative"
+    # 可以继续添加更多短语规则...
+}
+
+# 负面触发词集合
+NEGATIVE_TRIGGERS = {"drop", "decline", "failure", "loss", "down", "worse", "weak", "poor"}
+
+def get_phrase_sentiment(phrase):
+    """增强的短语情感分析逻辑"""
+    # 1. 优先检查自定义规则
+    lower_phrase = phrase.lower()
+    if lower_phrase in PHRASE_SENTIMENT_OVERRIDES:
+        return PHRASE_SENTIMENT_OVERRIDES[lower_phrase]
+    
+    # 2. 检查负面触发词
+    words = TextBlob(phrase).words
+    if any(w.lower() in NEGATIVE_TRIGGERS for w in words):
+        return "negative"
+    
+    # 3. 默认情感分析
+    sentiment = TextBlob(phrase).sentiment.polarity
+    if sentiment > 0.1:
+        return "positive"
+    elif sentiment < -0.1:
+        return "negative"
+    else:
+        return "neutral"
+
+def highlight_adjectives(text):
+    """高亮形容词短语并根据情感着色"""
+    if not isinstance(text, str) or not text.strip():
+        return text
+    
+    try:
+        blob = TextBlob(text)
+        highlighted = []
+        i = 0
+        tags = blob.tags
+        
+        while i < len(tags):
+            word, tag = tags[i]
+            
+            # 检查形容词短语模式 (形容词+名词)
+            if tag.startswith('JJ') and i+1 < len(tags) and tags[i+1][1].startswith('NN'):
+                phrase = f"{word} {tags[i+1][0]}"
+                # 使用增强的情感分析
+                sentiment = get_phrase_sentiment(phrase)
+                
+                color = ADJECTIVE_COLORS.get(sentiment, "#000000")
+                highlighted.append(f'<span style="color: {color}; font-weight: bold">{phrase}</span>')
+                i += 2  # 跳过下一个词，因为已经处理了
+            elif tag.startswith('JJ'):  # 单独形容词
+                sentiment = get_phrase_sentiment(word)  # 也能处理单个词
+                color = ADJECTIVE_COLORS.get(sentiment, "#000000")
+                highlighted.append(f'<span style="color: {color}; font-weight: bold">{word}</span>')
+                i += 1
+            else:
+                highlighted.append(word)
+                i += 1
+        
+        # 保留原始空格和标点
+        return " ".join(highlighted).replace(" ,", ",").replace(" .", ".").replace(" '", "'")
+    
+    except Exception as e:
+        print(f"Error processing text: {e}")
+        return text
+
+
 def on_confirm(dataset_radio, num_parts_dropdown, perspective_radio, division_method_radio):
     # 根据用户选择的参数构建文件路径
     num_parts = num_parts_dropdown
@@ -28,7 +114,10 @@ def on_confirm(dataset_radio, num_parts_dropdown, perspective_radio, division_me
     # 加载分析报告
     analysis_result,_ = load_analysis_report(dataset_radio, num_parts_dropdown, perspective_radio, division_method_radio)
     # AI分析列
-    df["Analysis"] = df["Model"].map(lambda m: analysis_result.get(m, "No analysis provided."))
+    # df["Analysis"] = df["Model"].map(lambda m: analysis_result.get(m, "No analysis provided."))
+    df["Analysis"] = df["Model"].map(
+        lambda m: highlight_adjectives(analysis_result.get(m, "No analysis provided."))
+    )
     return df
 
 # 生成 CSS 样式
@@ -36,7 +125,6 @@ def generate_css(line_counts, token_counts, cyclomatic_complexity, problem_type,
     css = """
     #dataframe th {
         background-color: #f2f2f2
-        
     }
     """
     colors = ["#e6f7ff", "#ffeecc", "#e6ffe6", "#ffe6e6"]
@@ -261,70 +349,104 @@ def plot_visualization(dataset_radio, perspective_radio, num_parts, plot_type):
 
     return fig
 
-# 旭日图
-def plot_recommendation_sunburst(dataset_radio, num_parts_dropdown, perspective_radio, division_method_radio):
+# 桑基图展示推荐模型
+def plot_recommendation_sankey(dataset_radio, num_parts_dropdown, perspective_radio, division_method_radio):
     import plotly.graph_objects as go
+    from plotly.colors import sample_colorscale
     _, recommendation_result = load_analysis_report(dataset_radio, num_parts_dropdown, perspective_radio, division_method_radio)
-    labels = ['Model Recommendation']  # 根节点
-    parents = ['']
-    values = []
-    customdata = ['Choose your preference model']
-
-    # 统计每个场景下模型数量
-    scenario_model_count = {}
-    total_model_count = 0
-
-    for scenario, model_list in recommendation_result.items():
-        # 处理模型
-        model_items = []
-        if isinstance(model_list, dict):
-            model_items = model_list.items()
-        elif isinstance(model_list, list):
-            for d in model_list:
-                if isinstance(d, dict):
-                    for k, v in d.items():
-                        model_items.append((k, v))
-
-        scenario_model_count[scenario] = len(model_items)
-        total_model_count += len(model_items)
-
-    # 根节点 value
-    values.append(total_model_count)
-
-    # 再次遍历，填充 labels/parents/values/customdata
-    for scenario, model_list in recommendation_result.items():
-        scenario_words = scenario.split()
-        short_label = " ".join(scenario_words[:3]) + "..." if len(scenario_words) > 3 else scenario
-        labels.append(short_label)
-        parents.append('Model Recommendation')
-        values.append(scenario_model_count[scenario])
+    
+    # 定义节点层级和颜色方案
+    levels = ['Model Recommendation', 'Scenario', 'Model Family', 'Specific Model']
+    color_scale = "RdYlBu_r"
+    
+    # 节点和连接数据
+    node_labels = [levels[0]]  # 根节点
+    customdata = ["Root node"]
+    sources, targets, values = [], [], []
+    
+    # 节点索引跟踪
+    node_indices = {levels[0]: 0}
+    current_idx = 1
+    
+    # 处理推荐列表结构 {"场景1": [ {模型1:原因1}, {模型2:原因2} ], ...}
+    for scenario, model_dicts in recommendation_result.items():
+        # 添加场景节点
+        scenario_label = " ".join(scenario.split()[:3]) + ("..." if len(scenario.split()) > 3 else "")
+        node_labels.append(scenario_label)
         customdata.append(scenario)
-
-        # 处理模型
-        model_items = []
-        if isinstance(model_list, dict):
-            model_items = model_list.items()
-        elif isinstance(model_list, list):
-            for d in model_list:
-                if isinstance(d, dict):
-                    for k, v in d.items():
-                        model_items.append((k, v))
-
-        for model, reason in model_items:
-            labels.append(model)
-            parents.append(short_label)
-            values.append(1)
-            customdata.append(reason)
-
-    fig = go.Figure(go.Sunburst(
-        labels=labels,
-        parents=parents,
-        values=values,
-        branchvalues="total",
-        hovertemplate='%{customdata}<extra></extra>',
-        customdata=customdata
+        node_indices[f"scenario_{scenario}"] = current_idx
+        current_idx += 1
+        
+        # 根节点 -> 场景节点连接
+        sources.append(0)
+        targets.append(node_indices[f"scenario_{scenario}"])
+        values.append(10)
+        
+        # 处理模型列表 [ {模型1:原因1}, {模型2:原因2} ]
+        for model_dict in model_dicts:
+            for model, reason in model_dict.items():
+                # 提取模型系列 (如"GPT-4" -> "GPT")
+                family = model.split('-')[0].split('_')[0]
+                
+                # 添加模型系列节点 (如果不存在)
+                if f"family_{family}" not in node_indices:
+                    node_labels.append(family)
+                    customdata.append(f"Model family: {family}")
+                    node_indices[f"family_{family}"] = current_idx
+                    current_idx += 1
+                
+                # 场景 -> 模型系列连接
+                sources.append(node_indices[f"scenario_{scenario}"])
+                targets.append(node_indices[f"family_{family}"])
+                values.append(8)
+                
+                # 添加具体模型节点 (如果不存在)
+                if f"model_{model}" not in node_indices:
+                    node_labels.append(model)
+                    customdata.append(f"<b>{model}</b><br>{reason}")
+                    node_indices[f"model_{model}"] = current_idx
+                    current_idx += 1
+                
+                # 模型系列 -> 具体模型连接
+                sources.append(node_indices[f"family_{family}"])
+                targets.append(node_indices[f"model_{model}"])
+                values.append(5)
+    
+    # 生成颜色 (确保颜色数量匹配节点数量)
+    node_colors = ["#2c7bb6"]  # 根节点颜色
+    node_colors += sample_colorscale(color_scale, [n/(len(node_labels)-1) for n in range(1, len(node_labels))])
+    
+    # 创建桑基图
+    fig = go.Figure(go.Sankey(
+        arrangement="perpendicular",
+        node=dict(
+            pad=20,
+            thickness=15,
+            line=dict(color="rgba(0,0,0,0.3)", width=0.2),
+            label=node_labels,
+            color=node_colors,
+            hovertemplate='%{label}<extra></extra>',
+            x=[0] + [0.33]*len([n for n in node_indices if n.startswith('scenario_')]) 
+                 + [0.66]*len([n for n in node_indices if n.startswith('family_')])
+                 + [1.0]*len([n for n in node_indices if n.startswith('model_')]),
+        ),
+        link=dict(
+            source=sources,
+            target=targets,
+            value=values,
+            color="rgba(180,180,180,0.4)",
+            customdata=[customdata[t] for t in targets],
+            hovertemplate='%{customdata}<extra></extra>'
+        )
     ))
-    fig.update_layout(margin=dict(t=10, l=10, r=10, b=10), height=500)
+    
+    fig.update_layout(
+        title_text="<b>Model Recommendation Flow</b>",
+        font_size=11,
+        height=700,
+        margin=dict(t=80, l=20, r=20, b=20)
+    )
+    
     return fig
 
 ### Gradio代码部分 ### 
@@ -429,7 +551,11 @@ with gr.Blocks(css=custom_css) as iface:
             with gr.Tabs():
                 # 表格
                 with gr.TabItem("Ranking Table"):
-                    data_table = gr.Dataframe(headers=["Model", "Score","Analysis"],interactive=True)
+                    data_table = gr.Dataframe(headers=["Model", "Score","Analysis"],
+                                              interactive=True,
+                                              datatype="html",  # 指定第三列为HTML
+                                              render=True, # 启用HTML渲染
+                                              )
                 # 可视化
                 with gr.TabItem("Visualization"):
                     plot_type = gr.Radio(
@@ -441,9 +567,9 @@ with gr.Blocks(css=custom_css) as iface:
                 # AI分析
                 with gr.TabItem("Model selection suggestions"):
                     with gr.Column():
-                        gr.Markdown("<h2 class='markdown-title'>🎯 Model Recommendation</h2>")
+                        # gr.Markdown("<h2 class='markdown-title'>🎯 Model Recommendation</h2>")
                         recommendation_plot = gr.Plot()
-                        scenario_legend = gr.Markdown(value="")  # 新增图例
+                        # scenario_legend = gr.Markdown(value="")  # 新增图例
     
     def update_perspective_options(dataset):
         if dataset == "MBPP":
@@ -480,8 +606,8 @@ with gr.Blocks(css=custom_css) as iface:
         fn=plot_visualization,
         inputs=[dataset_radio, perspective_radio, num_parts_slider, plot_type],
         outputs=chart
-    ).then(
-        fn=plot_recommendation_sunburst,
+    ).then( 
+        fn=plot_recommendation_sankey,
         inputs=[dataset_radio, num_parts_slider, perspective_radio, division_method_radio],
         outputs=[recommendation_plot]  # 注意这里是列表
     )

requirements.txt

@@ -2,4 +2,4 @@ huggingface-hub==0.24.2
 
 pip==24.0
 plotly==5.23.0
-
+textblob