update

app.py

@@ -11,6 +11,7 @@ import plotly.express as px
 import plotly.graph_objects as go
 from sklearn.linear_model import LinearRegression
 import numpy as np
+from scipy.spatial import ConvexHull
 
 load_dotenv()
 webhook_url = os.environ.get("WEBHOOK_URL")
@@ -310,37 +311,70 @@ def create_scaling_plot(all_data, period):
         'Average (The lower the better)': 'Compression Rate (%)'
     }, inplace=True)
 
-    # Create scatter plot
+    new_df['Log Params(B)'] = np.log(new_df['Params(B)'])
+    new_df['Log Compression Rate (%)'] = np.log(new_df['Compression Rate (%)'])
+
     fig = px.scatter(new_df,
-                     x='Params(B)',
-                     y='Compression Rate (%)',
+                     x='Log Params(B)',
+                     y='Log Compression Rate (%)',
                      title='Compression Rate Scaling Law',
                      hover_name='Name'
                      )
 
-    # Add logarithmic trendline
-    X = new_df[['Params(B)']].values
-    y = new_df['Compression Rate (%)'].values
+    names_to_connect = ['Meta-Llama-3-8B',
+                        'stablelm-3b-4e1t',
+                        'stablelm-2-1_6b',
+                        'TinyLlama-1.1B-intermediate-step-1431k-3T']
+    connection_points = new_df[new_df['Name'].isin(names_to_connect)]
+
+    new_df['Color'] = new_df['Name'].apply(lambda name: '#39C5BB' if name in names_to_connect else '#636efa')
 
-    # Perform log transformation on X
-    X_log = np.log(X)
+    fig.update_traces(marker=dict(color=new_df['Color']))
 
-    model = LinearRegression()
-    model.fit(X_log, y)
+    X = connection_points['Log Params(B)'].values.reshape(-1, 1)
+    y = connection_points['Log Compression Rate (%)'].values
+    model = LinearRegression().fit(X, y)
 
-    # Create trendline data for plot
-    X_plot = np.linspace(X_log.min() - 1, X_log.max() + 0.1, 100)
-    y_plot = model.predict(X_plot.reshape(-1, 1))
+    x_min = connection_points['Log Params(B)'].min()
+    x_max = connection_points['Log Params(B)'].max()
+    extended_x = np.linspace(x_min, x_max * 1.5, 100)
+    extended_x_original = np.exp(extended_x)
+    trend_line_y = model.predict(extended_x.reshape(-1, 1))
+    trend_line_y_original = np.exp(trend_line_y)
 
-    fig.add_trace(go.Scatter(
-        x=np.exp(X_plot),
-        y=y_plot,
+    trend_line = go.Scatter(
+        x=extended_x,
+        y=trend_line_y,
         mode='lines',
-        name='Trend',
-        line=dict(color='#39C5BB')
-    ))
+        line=dict(color='skyblue', dash='dash'),
+        name='Trend Line',
+        hovertemplate='<b>Params(B):</b> %{customdata[0]:.2f}<br>' +
+                      '<b>Compression Rate (%):</b> %{customdata[1]:.2f}<extra></extra>',
+        customdata=np.stack((extended_x_original, trend_line_y_original), axis=-1)
+    )
+
+    fig.add_trace(trend_line)
+
+    x_min = new_df['Params(B)'].min()
+    x_max = new_df['Params(B)'].max()
+    x_tick_vals = np.geomspace(x_min, x_max, num=5)
+    x_tick_text = [f"{val:.1f}" for val in x_tick_vals]
+
+    y_min = new_df['Compression Rate (%)'].min()
+    y_max = new_df['Compression Rate (%)'].max()
+    y_tick_vals = np.geomspace(y_min, y_max, num=5)
+    y_tick_text = [f"{val:.1f}" for val in y_tick_vals]
+
+    fig.update_xaxes(tickvals=np.log(x_tick_vals), ticktext=x_tick_text, title='Params(B)')
+    fig.update_yaxes(tickvals=np.log(y_tick_vals), ticktext=y_tick_text, title='Compression Rate (%)', autorange='reversed')
+
+    fig.update_layout(
+        xaxis=dict(showgrid=True, zeroline=False),
+        yaxis=dict(showgrid=True, zeroline=False)
+    )
 
     fig.update_traces(marker=dict(size=12))
+
     return fig