Go back to accuracy plot

app.py

@@ -61,7 +61,10 @@ MODEL_INFO = [
 ]
 
 DEMO_LEARNING_RATE = 0.05 # don't use default; use something more fun
-DEMO_ALPHA = 0.25
+DEMO_ALPHA = 0.9 # 0.25   # this is more fun if showing the confusion matrices
+
+# Toggle between confusion matrix and accuracy chart
+USE_CONFUSION_MATRIX = False  # Set to True for confusion matrices, False for accuracy bars
 
 def create_species_guide_content():
     """Create the species identification guide content"""
@@ -370,6 +373,15 @@ def create_probability_chart():
     return temp_fig
 
 def create_accuracy_chart():
+    """Create either confusion matrices or accuracy bar chart based on USE_CONFUSION_MATRIX toggle"""
+    global coda_selector, oracle, dataset, iteration_count
+
+    if USE_CONFUSION_MATRIX:
+        return create_confusion_matrix_chart()
+    else:
+        return create_accuracy_bar_chart()
+
+def create_confusion_matrix_chart():
     """Create confusion matrix estimates for each model side by side"""
     global coda_selector, iteration_count
 
@@ -438,52 +450,52 @@ def create_accuracy_chart():
     plt.close(fig)
     return temp_fig
 
-# OLD CODE - True Model Accuracies Bar Chart (kept for easy reversion)
-# def create_accuracy_chart():
-#     """Create a bar chart showing true accuracy of each model"""
-#     global oracle, dataset
-#
-#     if oracle is None or dataset is None:
-#         # Fallback for initial state
-#         model_labels = [info['name'] for info in MODEL_INFO]
-#         accuracies = np.random.random(len(MODEL_INFO))  # Random accuracies for now
-#     else:
-#         true_losses = oracle.true_losses(dataset.preds)
-#         # Convert losses to accuracies (assuming loss is 1 - accuracy)
-#         accuracies = (1 - true_losses).detach().cpu().numpy().flatten()
-#         model_labels = ["   " + info['name'] for info in MODEL_INFO[:len(accuracies)]]
-#
-#     # Find the index of the highest accuracy
-#     best_idx = np.argmax(accuracies)
-#
-#     fig, ax = plt.subplots(figsize=(8, 2.8), dpi=150)
-#
-#     # Create colors array - highlight the best model
-#     colors = ['red' if i == best_idx else 'forestgreen' for i in range(len(model_labels))]
-#     bars = ax.bar(range(len(model_labels)), accuracies, color=colors, alpha=0.7)
-#
-#     # Add text above the highest bar
-#     ax.text(best_idx, accuracies[best_idx] + 0.005, 'True best model',
-#             ha='center', va='bottom', fontsize=12, fontweight='bold')
-#
-#     ax.set_ylabel('True (oracle) \naccuracy of model', fontsize=12)
-#     ax.set_title('True Model Accuracies', fontsize=12)
-#     ax.set_ylim(np.min(accuracies) - 0.025, np.max(accuracies) + 0.05)
-#
-#     # Set x-axis labels and ticks
-#     ax.set_xticks(range(len(model_labels)))
-#     ax.set_xticklabels(model_labels, fontsize=12, ha='center')
-#
-#     # Add logos to x-axis
-#     for i, model_info in enumerate(MODEL_INFO[:len(accuracies)]):
-#         add_logo_to_x_axis(ax, i, model_info['logo'], model_info['name'])
-#     plt.yticks(fontsize=12)
-#     plt.tight_layout()
-#
-#     # Save the figure and close it to prevent memory leaks
-#     temp_fig = fig
-#     plt.close(fig)
-#     return temp_fig
+def create_accuracy_bar_chart():
+    """Create a bar chart showing true accuracy of each model (with muted colors)"""
+    global oracle, dataset
+
+    if oracle is None or dataset is None:
+        # Fallback for initial state
+        fig, ax = plt.subplots(figsize=(8, 2.8), dpi=150)
+        ax.text(0.5, 0.5, 'Start demo to see model accuracies',
+                ha='center', va='center', fontsize=12)
+        ax.axis('off')
+        plt.tight_layout()
+        temp_fig = fig
+        plt.close(fig)
+        return temp_fig
+
+    true_losses = oracle.true_losses(dataset.preds)
+    # Convert losses to accuracies (assuming loss is 1 - accuracy)
+    accuracies = (1 - true_losses).detach().cpu().numpy().flatten()
+    model_labels = [" "*(9 if info['name']=='LAION CLIP' else 4 if info['name']=='SigLIP2' else 6) + info['name'] for info in MODEL_INFO[:len(accuracies)]]
+
+    # Find the index of the highest accuracy
+    best_idx = np.argmax(accuracies)
+
+    fig, ax = plt.subplots(figsize=(8, 2.8), dpi=150)
+
+    # Use muted, uniform color for all bars (no highlighting)
+    bars = ax.bar(range(len(model_labels)), accuracies, color='#8B9DC3', alpha=0.6)
+
+    ax.set_ylabel('True (oracle) \naccuracy of model', fontsize=12)
+    ax.set_title('True Model Accuracies', fontsize=12)
+    ax.set_ylim(np.min(accuracies) - 0.025, np.max(accuracies) + 0.05)
+
+    # Set x-axis labels and ticks
+    ax.set_xticks(range(len(model_labels)))
+    ax.set_xticklabels(model_labels, fontsize=12, ha='center')
+
+    # Add logos to x-axis
+    for i, model_info in enumerate(MODEL_INFO[:len(accuracies)]):
+        add_logo_to_x_axis(ax, i, model_info['logo'], model_info['name'])
+    plt.yticks(fontsize=12)
+    plt.tight_layout()
+
+    # Save the figure and close it to prevent memory leaks
+    temp_fig = fig
+    plt.close(fig)
+    return temp_fig
 
 # Create the Gradio interface
 with gr.Blocks(title="CODA: Wildlife Photo Classification Challenge",