update

app.py

@@ -10,11 +10,12 @@ observed_points = []
 generated_curves = []
 generated_configurations = []
 history = {}
+rank = []
 surrogate_model = FTPFN()
 
 # Function to generate curves
 def generate_curves(num_curves, max_length):
-    global generated_curves, generated_configurations, history
+    global generated_curves, generated_configurations, history, rank
     reset_optimization()
 
     configurations, curves = sample_curves(num_hyperparameters=num_curves, 
@@ -35,6 +36,8 @@ def generate_curves(num_curves, max_length):
 
     generated_curves = torch.FloatTensor(curves)
     generated_configurations = torch.FloatTensor(configurations)
+    rank = list(range(num_curves))
+
     for i in range(num_curves):
         history[i] = 0
     return fig, None, hyperparam_display
@@ -106,6 +109,9 @@ def bayesian_optimization_step(num_curves, max_length):
         next_point = np.random.randint(0, num_curves)
     else:
         next_point = next_observed_point(history, generated_curves, generated_configurations)
+    
+    rank.remove(next_point)
+    rank.append(next_point)
 
     if next_point in history:
         history[next_point] += 1
@@ -120,22 +126,29 @@ def bayesian_optimization_step(num_curves, max_length):
     # Plot the updated curves with uncertainty and observed points
     fig, ax = plt.subplots()
 
-    for i, curve in enumerate(curves):
-        color = generated_configurations[i].numpy().tolist()
-        ax.plot(np.arange(1, len(curves[0]) + 1), curve, alpha=0.1, color=color)
-        x = np.arange(max(history[i], 1), len(q05[i]) + max(history[i], 1))
-        ax.fill_between(x, q05[i], q95[i], alpha=0.3, color=color)
-    
-    for curve_id, epoch in history.items():
+    for r in rank:
+        curve_id = r
+        epoch = history[curve_id]
+        curve = curves[curve_id]
         color = generated_configurations[curve_id].numpy().tolist()
-        ax.plot(np.arange(1, epoch+1), curves[curve_id][:epoch], 'ro', color=color)
-        ax.plot(np.arange(1, epoch+1), curves[curve_id][:epoch], color=color)
-    
+
+        # Plot the full curve with reduced opacity
+        ax.plot(np.arange(1, len(curves[0]) + 1), curve, alpha=0.1, color=color)
+
+        # Plot the uncertainty region
+        x = np.arange(max(history[curve_id], 1), len(q05[curve_id]) + max(history[curve_id], 1))
+        ax.fill_between(x, q05[curve_id], q95[curve_id], alpha=0.3, color=color)
+
+        # Plot observed points and lines for the fully observed part of the curve
+        ax.plot(np.arange(1, epoch + 1), curve[:epoch], 'ro', color=color)
+        ax.plot(np.arange(1, epoch + 1), curve[:epoch], color=color)
+
     plt.xlim(0, len(curves[0]) + 1)
     plt.ylim(0, 1)
-    plt.title(f"Step {len(observed_points)}")
+    plt.title(f"Iteration {len(observed_points)}: Observing curve {next_point+1} at step {history[next_point]}")
     plt.xlabel("t")
     plt.ylabel("y")
+    
     return fig
 
 # Reset function for Bayesian optimization
@@ -143,6 +156,7 @@ def reset_optimization():
     global observed_points, generated_curves, generated_configurations, history
     observed_points = []
     history = {}
+    rank = []
     for i in range(len(generated_curves)):
         history[i] = 0
     return None
@@ -160,11 +174,12 @@ with gr.Blocks() as demo:
     gr.HTML(scroll_script)
 
     # Add a title
-    gr.Markdown("# ifBO: In-context Freeze-Thaw Bayesian Optimization")
+    gr.Markdown("<h1 style=\"text-align: center;\">ifBO: In-context Freeze-Thaw Bayesian Optimization</h1>")
     gr.Markdown("Paper: [https://arxiv.org/pdf/2404.16795](https://arxiv.org/pdf/2404.16795)")
+    gr.Markdown("Code: [http://github.com/automl/ifBO](http://github.com/automl/ifBO)")
      
     # First section for curve generation
-    gr.Markdown("### Curve Generation")
+    gr.Markdown("## Curve Generation")
     gr.Markdown("""
     To generate' a set of synthetic curves (according to the proposed curve prior, see section 4.1), please specify the number of curves and the maximum length of each curve, then click 'Generate'. 
     """)
@@ -180,7 +195,7 @@ with gr.Blocks() as demo:
             curve_plot = gr.Plot()
 
     # Separate section for Bayesian optimization
-    gr.Markdown("### Bayesian Optimization")
+    gr.Markdown("## Bayesian Optimization")
     gr.Markdown("""
         After generating the curves, click 'One Step with ifBO' to take an optimization step.
         During this step, **ifBO** will select the next point to evaluate based on the previously observed points, using the MFPI-random acquisition function (see section 4.2).