Update app.py

app.py

@@ -7,6 +7,7 @@ from langchain_openai import ChatOpenAI
 from langchain_experimental.agents.agent_toolkits import create_pandas_dataframe_agent
 from langchain.agents.agent_types import AgentType
 from langchain_google_genai import ChatGoogleGenerativeAI
+import plotly.graph_objects as go
 
 def explain_df(query, df):
     agent = create_pandas_dataframe_agent(
@@ -82,6 +83,7 @@ numeric_columns = df.select_dtypes(include=[np.number]).columns
 numeric_columns = numeric_columns.drop('model_physical_size')
 df[numeric_columns] = (df[numeric_columns]*100).round(2)
 df['model_physical_size'] = df['model_physical_size'].round(2)
+
 full_df = df.merge(perf_df, left_on='hf_name', right_on='hf_name', how='left')
 
 with gr.Blocks() as demo:
@@ -102,21 +104,45 @@ with gr.Blocks() as demo:
                 latency_line_plot = gr.Plot(label="Latency vs Average Accuracy")
             with gr.Row():
                 data_table = gr.Dataframe(value=df, label="Result Table")
+
             def update_outputs(selected_tasks):
                 if not selected_tasks:
                     return df[['model', 'precision']], None, None
                 filtered_df = df[['model', 'precision', 'model_physical_size','hf_name'] + selected_tasks]
                 # average accuracy of selected tasks
                 filtered_df['avg_accuracy'] = filtered_df[selected_tasks].mean(axis=1)
+                
                 bar_fig = px.bar(filtered_df, x='model', y='avg_accuracy', color='precision', barmode='group')
                 line_fig = px.line(filtered_df, x='model_physical_size', y='avg_accuracy', color='model', symbol='precision')
+                pareto_df = filtered_df.sort_values('model_physical_size')
+                pareto_df = pareto_df.loc[pareto_df['avg_accuracy'].cummax().drop_duplicates().index]
+                # Add Pareto frontier to line_plot
+                line_fig.add_trace(go.Scatter(
+                    x=pareto_df['model_physical_size'],
+                    y=pareto_df['avg_accuracy'],
+                    mode='lines+markers',
+                    name='Pareto Frontier'
+                ))
+                
                 # set title of bar_fig
                 bar_fig.update_layout(title=f'tasks: {", ".join(selected_tasks)}')
                 line_fig.update_layout(title=f'tasks: {", ".join(selected_tasks)}')
                 with_perf_df = filtered_df.merge(perf_df, left_on='hf_name', right_on='hf_name', how='left')
                 throughput_line_fig = px.line(with_perf_df, x='output_throughput', y='avg_accuracy', color='model', symbol='precision')
                 latency_line_fig = px.line(with_perf_df, x="avg_e2e_latency", y='avg_accuracy', color='model', symbol='precision')
+                
+                pareto_df = with_perf_df.sort_values('avg_e2e_latency')
+                pareto_df = pareto_df.loc[pareto_df['avg_accuracy'].cummax().drop_duplicates().index]
+                
+                latency_line_fig.add_trace(go.Scatter(
+                    x=pareto_df['avg_e2e_latency'],
+                    y=pareto_df['avg_accuracy'],
+                    mode='lines+markers',
+                    name='Pareto Frontier'
+                ))
+                print(with_perf_df)
                 return with_perf_df, bar_fig, line_fig, throughput_line_fig, latency_line_fig
+            
             selected_tasks.change(
                 fn=update_outputs,
                 inputs=selected_tasks,