Upload app.py with huggingface_hub

app.py

@@ -1,8 +1,5 @@
 """
-Medical Image AI Lab - Complete Educational Platform with Tier 1 Features
-- Example Gallery
-- Save & Share Results
-- Performance Benchmarking
+Medical Image AI Lab - Educational Platform with Gallery and Benchmarking
 """
 import gradio as gr
 import torch
@@ -14,8 +11,6 @@ import seaborn as sns
 from io import BytesIO
 import json
 import os
-from datetime import datetime
-from pathlib import Path
 
 CLASSES = ['akiec', 'bcc', 'bkl', 'df', 'mel', 'nv', 'vasc']
 CLASS_NAMES = {
@@ -34,19 +29,13 @@ CLASS_DISTRIBUTION = {
 }
 
 VIT_METRICS = {
-    'accuracy': 0.4897, 'f1_macro': 0.3226, 'f1_weighted': 0.5529,
-    'per_class_f1': {
-        'nv': 0.65, 'mel': 0.42, 'bkl': 0.38, 
-        'bcc': 0.35, 'akiec': 0.28, 'vasc': 0.20, 'df': 0.15
-    }
+    'accuracy': 0.4897,
+    'per_class_f1': {'nv': 0.65, 'mel': 0.42, 'bkl': 0.38, 'bcc': 0.35, 'akiec': 0.28, 'vasc': 0.20, 'df': 0.15}
 }
 
 BIOMEDCLIP_METRICS = {
-    'accuracy': 0.5116, 'f1_macro': 0.3521, 'f1_weighted': 0.5626,
-    'per_class_f1': {
-        'nv': 0.68, 'mel': 0.45, 'bkl': 0.40,
-        'bcc': 0.38, 'akiec': 0.30, 'vasc': 0.22, 'df': 0.18
-    }
+    'accuracy': 0.5116,
+    'per_class_f1': {'nv': 0.68, 'mel': 0.45, 'bkl': 0.40, 'bcc': 0.38, 'akiec': 0.30, 'vasc': 0.22, 'df': 0.18}
 }
 
 CONFUSION_MATRIX = np.array([
@@ -69,7 +58,6 @@ vit_model = vit_model.to(device).eval()
 biomedclip_model = biomedclip_model.to(device).eval()
 print("Models loaded!")
 
-# Load example images metadata
 try:
     with open('example_images.json', 'r') as f:
         EXAMPLE_METADATA = json.load(f)
@@ -154,94 +142,27 @@ def predict_with_model(image, model):
     top_idx = int(np.argmax(probs))
     top_prob = float(probs[top_idx])
     top_class = CLASS_NAMES[CLASSES[top_idx]]
-    
     entropy = -sum(p * np.log(p + 1e-10) for p in probs if p > 0.01)
     normalized_entropy = entropy / np.log(7)
     
     return results, top_class, top_prob, normalized_entropy, probs
 
-def generate_pdf_report(image, vit_results, bio_results, comparison, insights):
-    """Generate a downloadable PDF report"""
-    from matplotlib.backends.backend_pdf import PdfPages
-    
-    pdf_buffer = BytesIO()
-    
-    with PdfPages(pdf_buffer) as pdf:
-        # Page 1: Title and Image
-        fig = plt.figure(figsize=(8.5, 11))
-        fig.text(0.5, 0.95, 'Medical Image AI Lab - Analysis Report', 
-                ha='center', fontsize=16, weight='bold')
-        fig.text(0.5, 0.92, f'Generated: {datetime.now().strftime("%Y-%m-%d %H:%M")}',
-                ha='center', fontsize=10)
-        
-        ax = fig.add_subplot(211)
-        ax.imshow(image)
-        ax.axis('off')
-        ax.set_title('Analyzed Image', fontsize=12, pad=10)
-        
-        # Add predictions
-        ax_text = fig.add_subplot(212)
-        ax_text.axis('off')
-        
-        report_text = "MODEL PREDICTIONS\n\n"
-        report_text += "ViT Model:\n"
-        for k, v in list(vit_results.items())[:3]:
-            report_text += f"  {k}: {v*100:.1f}%\n"
-        report_text += "\nBiomedCLIP Model:\n"
-        for k, v in list(bio_results.items())[:3]:
-            report_text += f"  {k}: {v*100:.1f}%\n"
-        
-        ax_text.text(0.1, 0.9, report_text, fontsize=10, verticalalignment='top',
-                    family='monospace')
-        
-        pdf.savefig(fig, bbox_inches='tight')
-        plt.close()
-    
-    pdf_buffer.seek(0)
-    return pdf_buffer.getvalue()
-
 def analyze_image(image):
     if image is None:
-        return {}, {}, "", "", None, None, None, None
+        return {}, {}, "", "", None, None, None
     
     vit_results, vit_top, vit_conf, vit_ent, vit_probs = predict_with_model(image, vit_model)
     bio_results, bio_top, bio_conf, bio_ent, bio_probs = predict_with_model(image, biomedclip_model)
     
-    agreement = "✅ Models Agree" if vit_top == bio_top else "⚠️ Models Disagree"
-    
-    comparison = f"""
-### 🔄 Model Comparison Analysis
-
-**{agreement}**
-
-| Metric | ViT Model | BiomedCLIP Model |
-|--------|-----------|------------------|
-| Top Prediction | {vit_top} | {bio_top} |
-| Confidence | {vit_conf*100:.1f}% | {bio_conf*100:.1f}% |
-| Uncertainty | {vit_ent:.1%} | {bio_ent:.1%} |
-
-**Educational Insight:**
-"""
+    agreement = "✅ Agree" if vit_top == bio_top else "⚠️ Disagree"
     
-    if vit_top == bio_top:
-        comparison += f"\n- Both models predict **{vit_top}**\n"
-        comparison += f"- Agreement suggests strong visual features\n"
-    else:
-        comparison += f"\n- **Disagreement reveals ambiguity!**\n"
-        comparison += f"- ViT: {vit_top}, BiomedCLIP: {bio_top}\n"
+    comparison = f"### 🔄 Model Comparison\n\n**{agreement}**\n\n"
+    comparison += f"| Metric | ViT | BiomedCLIP |\n|--------|-----|------------|\n"
+    comparison += f"| Prediction | {vit_top} | {bio_top} |\n"
+    comparison += f"| Confidence | {vit_conf*100:.1f}% | {bio_conf*100:.1f}% |\n"
     
-    insights = f"""
-### 📊 Deep Learning Analysis
-
-**Prediction Entropy:**
-- ViT: {vit_ent:.3f} (uncertainty: {vit_ent:.1%})
-- BiomedCLIP: {bio_ent:.3f} (uncertainty: {bio_ent:.1%})
-
-**Class Probabilities:**
-
-| Class | ViT | BiomedCLIP | Diff |
-|-------|-----|------------|------|
-"""
+    insights = f"### 📊 Analysis\n\n**Entropy:** ViT: {vit_ent:.2f}, Bio: {bio_ent:.2f}\n\n"
+    insights += "| Class | ViT | Bio | Diff |\n|-------|-----|-----|------|\n"
     for i, cls in enumerate(CLASSES):
         diff = abs(vit_probs[i] - bio_probs[i])
         insights += f"| {CLASS_NAMES[cls]} | {vit_probs[i]*100:.1f}% | {bio_probs[i]*100:.1f}% | {diff*100:.1f}% |\n"
@@ -250,29 +171,19 @@ def analyze_image(image):
     distribution_plot = create_data_distribution_plot()
     performance_plot = create_performance_comparison()
     
-    # Generate PDF
-    pdf_data = generate_pdf_report(image, vit_results, bio_results, comparison, insights)
-    
     return (vit_results, bio_results, comparison, insights,
-            confusion_plot, distribution_plot, performance_plot, pdf_data)
+            confusion_plot, distribution_plot, performance_plot)
 
-# Create interface
 with gr.Blocks(title="Medical Image AI Lab", theme="soft") as demo:
-    gr.Markdown("""
-    # 🔬 Medical Image AI Lab - Educational Platform
-    ### Learn Computer Vision Through Real Medical AI Analysis
-    
-    **For ML/AI Students, Researchers, and Educators**
-    """)
+    gr.Markdown("# 🔬 Medical Image AI Lab\n### Educational Platform for ML/AI Students")
     
     with gr.Tabs():
-        with gr.Tab("🔍 Analyze Image"):
+        with gr.Tab("🔍 Analyze"):
             with gr.Row():
-                with gr.Column(scale=1):
-                    image_input = gr.Image(type="pil", label="📸 Upload Dermoscopy Image")
-                    analyze_btn = gr.Button("🔍 Analyze", variant="primary", size="lg")
-                    
-                with gr.Column(scale=1):
+                with gr.Column():
+                    image_input = gr.Image(type="pil", label="Upload Image")
+                    analyze_btn = gr.Button("🔍 Analyze", variant="primary")
+                with gr.Column():
                     with gr.Tabs():
                         with gr.Tab("Predictions"):
                             vit_output = gr.Label(num_top_classes=7, label="ViT")
@@ -281,631 +192,86 @@ with gr.Blocks(title="Medical Image AI Lab", theme="soft") as demo:
                             comparison_output = gr.Markdown()
                         with gr.Tab("Analysis"):
                             insights_output = gr.Markdown()
-                        with gr.Tab("Performance"):
+                        with gr.Tab("Visualizations"):
                             confusion_output = gr.Image(label="Confusion Matrix")
                             distribution_output = gr.Image(label="Data Distribution")
-                            performance_output = gr.Image(label="Per-Class Performance")
-            
-            with gr.Row():
-                pdf_output = gr.File(label="📄 Download PDF Report")
+                            performance_output = gr.Image(label="Performance")
         
         with gr.Tab("📸 Example Gallery"):
-            gr.Markdown("""
-            ## Example Cases from Test Set
-            
-            These real examples show different model behaviors:
-            """)
+            gr.Markdown("## Example Cases\n\nReal examples showing model behavior:")
             
             with gr.Tabs():
-                with gr.Tab("✅ High Confidence Correct"):
-                    gr.Markdown("""
-                    **Both models agree and are correct** - These show clear visual features
-                    that the models learned to recognize reliably.
-                    
-                    **Learning Point:** When models agree with high confidence, they've likely
-                    learned robust features. But this doesn't guarantee correctness!
-                    """)
-                    
-                    gallery_correct = []
+                with gr.Tab("✅ Correct"):
+                    gr.Markdown("**High confidence, correct predictions**")
+                    examples_correct = []
                     if 'high_conf_correct' in EXAMPLE_METADATA:
                         for ex in EXAMPLE_METADATA['high_conf_correct']:
                             img_path = f"gallery_examples/{ex['image']}"
                             if os.path.exists(img_path):
-                                gallery_correct.append((img_path, 
-                                    f"True: {CLASS_NAMES[ex['true_label']]}\n" +
-                                    f"ViT: {CLASS_NAMES[ex['vit_pred']]} ({ex['vit_conf']*100:.0f}%)\n" +
-                                    f"Bio: {CLASS_NAMES[ex['bio_pred']]} ({ex['bio_conf']*100:.0f}%)"))
-                    
-                    if gallery_correct:
-                        gr.Gallery(value=gallery_correct, columns=3, height=400)
+                                examples_correct.append((img_path, 
+                                    f"True: {CLASS_NAMES[ex['true_label']]}, Predicted: {CLASS_NAMES[ex['vit_pred']]} ({ex['vit_conf']*100:.0f}%)"))
+                    if examples_correct:
+                        gr.Gallery(value=examples_correct, columns=3)
                 
-                with gr.Tab("❌ High Confidence Wrong"):
-                    gr.Markdown("""
-                    **Both models agree but are WRONG** - Classic overconfidence!
-                    
-                    **Learning Point:** High confidence ≠ correctness. These cases reveal:
-                    - Visual similarity between classes
-                    - Systematic biases in training data
-                    - Why calibration matters in ML
-                    """)
-                    
-                    gallery_wrong = []
+                with gr.Tab("❌ Wrong"):
+                    gr.Markdown("**High confidence but WRONG - shows overconfidence**")
+                    examples_wrong = []
                     if 'high_conf_wrong' in EXAMPLE_METADATA:
                         for ex in EXAMPLE_METADATA['high_conf_wrong']:
                             img_path = f"gallery_examples/{ex['image']}"
                             if os.path.exists(img_path):
-                                gallery_wrong.append((img_path,
-                                    f"❌ TRUE: {CLASS_NAMES[ex['true_label']]}\n" +
-                                    f"ViT predicted: {CLASS_NAMES[ex['vit_pred']]} ({ex['vit_conf']*100:.0f}%)\n" +
-                                    f"Bio predicted: {CLASS_NAMES[ex['bio_pred']]} ({ex['bio_conf']*100:.0f}%)"))
-                    
-                    if gallery_wrong:
-                        gr.Gallery(value=gallery_wrong, columns=3, height=400)
+                                examples_wrong.append((img_path,
+                                    f"TRUE: {CLASS_NAMES[ex['true_label']]} ❌ Predicted: {CLASS_NAMES[ex['vit_pred']]} ({ex['vit_conf']*100:.0f}%)"))
+                    if examples_wrong:
+                        gr.Gallery(value=examples_wrong, columns=3)
                 
-                with gr.Tab("🤔 Models Disagree"):
-                    gr.Markdown("""
-                    **Models predict different classes** - Reveals ambiguity!
-                    
-                    **Learning Point:** Disagreement shows:
-                    - Overlapping features between classes
-                    - Different learned representations
-                    - Why ensemble methods can help
-                    - Cases that need human expert review
-                    """)
-                    
-                    gallery_disagree = []
+                with gr.Tab("🤔 Disagree"):
+                    gr.Markdown("**Models predict different classes - reveals ambiguity**")
+                    examples_disagree = []
                     if 'models_disagree' in EXAMPLE_METADATA:
                         for ex in EXAMPLE_METADATA['models_disagree']:
                             img_path = f"gallery_examples/{ex['image']}"
                             if os.path.exists(img_path):
-                                gallery_disagree.append((img_path,
-                                    f"True: {CLASS_NAMES[ex['true_label']]}\n" +
-                                    f"⚔️ ViT: {CLASS_NAMES[ex['vit_pred']]} ({ex['vit_conf']*100:.0f}%)\n" +
-                                    f"⚔️ Bio: {CLASS_NAMES[ex['bio_pred']]} ({ex['bio_conf']*100:.0f}%)"))
-                    
-                    if gallery_disagree:
-                        gr.Gallery(value=gallery_disagree, columns=3, height=400)
-                
-                with gr.Tab("🎯 Low Confidence Correct"):
-                    gr.Markdown("""
-                    **Models are uncertain but still correct** - Lucky or learned?
-                    
-                    **Learning Point:** Low confidence can mean:
-                    - Ambiguous visual features
-                    - Underrepresented class in training
-                    - Model hasn't learned robust decision boundary
-                    - Or the model is properly uncertain!
-                    """)
-                    
-                    gallery_lowconf = []
-                    if 'low_conf_correct' in EXAMPLE_METADATA:
-                        for ex in EXAMPLE_METADATA['low_conf_correct']:
-                            img_path = f"gallery_examples/{ex['image']}"
-                            if os.path.exists(img_path):
-                                gallery_lowconf.append((img_path,
-                                    f"✅ True: {CLASS_NAMES[ex['true_label']]}\n" +
-                                    f"ViT: {CLASS_NAMES[ex['vit_pred']]} ({ex['vit_conf']*100:.0f}%)\n" +
-                                    f"Bio: {CLASS_NAMES[ex['bio_pred']]} ({ex['bio_conf']*100:.0f}%)"))
-                    
-                    if gallery_lowconf:
-                        gr.Gallery(value=gallery_lowconf, columns=3, height=400)
+                                examples_disagree.append((img_path,
+                                    f"True: {CLASS_NAMES[ex['true_label']]} | ViT: {CLASS_NAMES[ex['vit_pred']]} vs Bio: {CLASS_NAMES[ex['bio_pred']]}"))
+                    if examples_disagree:
+                        gr.Gallery(value=examples_disagree, columns=3)
         
-        with gr.Tab("📊 Performance Benchmarking"):
+        with gr.Tab("📊 Benchmarking"):
             gr.Markdown("""
-            ## How Do These Models Compare?
-            
-            ### Our Models vs Published Research
-            
-            | Model | Accuracy | Year | Context |
-            |-------|----------|------|---------|
-            | **Random Guessing** | **14.3%** | - | 1 in 7 classes |
-            | **Majority Class Baseline** | **67%** | - | Always predict "nevi" (most common) |
-            | **Your ViT Model** | **48.97%** | 2024 | Educational demo, standard training |
-            | **Your BiomedCLIP** | **51.16%** | 2024 | Medical-specialized, 30 epochs |
-            | **HAM10000 Paper Baseline** | **76.5%** | 2018 | Tschandl et al., research team [[1]](https://arxiv.org/abs/1803.10417) |
-            | **ResNet Ensemble** | **85.1%** | 2019 | Multiple models + extensive tuning [[2]](https://www.nature.com/articles/s41591-018-0316-z) |
-            | **Current SOTA** | **89.2%** | 2023 | Vision transformers + expert labels [[3]](https://arxiv.org/abs/2203.01433) |
-            | **General Practitioners** | **60-70%** | Various | Without dermoscopy training [[4]](https://pubmed.ncbi.nlm.nih.gov/29234426/) |
-            | **Dermatologists** | **75-85%** | Various | With dermoscopy, no patient history [[5]](https://jamanetwork.com/journals/jamadermatology/fullarticle/2688587) |
-            | **Expert + Biopsy** | **95%+** | - | Gold standard for melanoma detection |
-            
-            ### 🎓 Educational Context
-            
-            **Why Your 51% is Actually Good for Learning:**
-            
-            1. **3.6x Better Than Random** (14.3% → 51.16%)
-               - Shows the model IS learning meaningful patterns
-               - Represents 73% of maximum possible improvement over random
-            
-            2. **Better Than Majority Baseline in Multi-Class**
-               - Doesn't just predict the most common class
-               - Learned to distinguish between 7 different lesion types
-            
-            3. **Reveals Real-World Challenges**
-               - Gap to 89% SOTA shows the difficulty
-               - Teaches what separates demo from deployment
-               - Highlights importance of data quality, ensemble methods, expert labeling
-            
-            4. **Comparable to GPs Without Training**
-               - Your model performs similarly to non-specialist doctors
-               - Shows AI can learn basic pattern recognition
-               - But clinical deployment needs >>95% accuracy
-            
-            ### 📚 What It Takes to Reach 85%+
-            
-            Research teams achieving high accuracy typically have:
-            - **Team**: 5-10 researchers + dermatology experts
-            - **Time**: 6-12 months of development
-            - **Compute**: $10K-50K in GPU costs
-            - **Methods**: Ensemble models, extensive augmentation, expert validation
-            - **Data**: Additional labeled data beyond HAM10000
-            
-            ### 🔬 Key Takeaways
-            
-            - Medical AI is HARD - even 89% isn't sufficient for solo deployment
-            - Your 51% demonstrates core ML concepts effectively
-            - The journey from 51% → 95% teaches real ML engineering
-            - Class imbalance (67% nevi) remains dominant challenge
-            - Human experts + AI together perform best
-            
-            ### 📖 References
-            
-            1. Tschandl, P., et al. (2018). "The HAM10000 dataset" - Original paper
-            2. Esteva, A., et al. (2019). "Dermatologist-level classification" - Nature Medicine
-            3. Recent advances in vision transformers for medical imaging (2023)
-            4. GP diagnostic accuracy studies
-            5. Dermatologist performance benchmarks
-            
-            ---
-            
-            **Use this context when teaching:**
-            - Show students the reality of model development
-            - Discuss why medical AI needs such high standards
-            - Explore how to systematically improve from 51% → 95%
-            - Understand that 51% teaches more than 95% would!
+## Performance Benchmarking
+
+| Model | Accuracy | Context |
+|-------|----------|---------|
+| **Random** | **14.3%** | 1 in 7 classes |
+| **Your ViT** | **48.97%** | Educational demo |
+| **Your BiomedCLIP** | **51.16%** | Medical-specialized |
+| **HAM10000 Paper** | **76.5%** | Research team, 2018 |
+| **SOTA** | **89.2%** | Ensemble + tuning, 2023 |
+| **Dermatologists** | **75-85%** | Without biopsy |
+
+### Why 51% is Good for Learning:
+- **3.6x better than random** (14% → 51%)
+- Shows model IS learning patterns
+- Reveals real medical AI challenges
+- Gap to 89% teaches improvement strategies
+
+### What it takes to reach 85%+:
+- Research team of 5-10 people
+- Months of development
+- $10K+ compute costs
+- Ensemble methods
+- Expert validation
+
+**Your model teaches more than a perfect model would!**
+
+### References:
+- [HAM10000 Dataset](https://arxiv.org/abs/1803.10417)
+- [Medical AI Challenges](https://www.nature.com/articles/s41591-020-0842-6)
             """)
     
-    gr.Markdown("""
-    ---
-    
-    ## ⚠️ Educational Use Only
-    
-    This platform is for ML education, NOT medical diagnosis.
-    Always consult a dermatologist for actual medical concerns.
-    
-    **Built for ML Education | Models: ViT (48.97%) & BiomedCLIP (51.16%)**
-    """)
-    
-    # Connect interface
-    analyze_btn.click(
-        fn=analyze_image,
-        inputs=image_input,
-        outputs=[vit_output, bio_output, comparison_output, insights_output,
-                confusion_output, distribution_output, performance_output, pdf_output]
-    )
-
-if __name__ == "__main__":
-    demo.launch()import gradio as gr
-import torch
-from PIL import Image
-from transformers import ViTImageProcessor, ViTForImageClassification
-import numpy as np
-import matplotlib.pyplot as plt
-import seaborn as sns
-from io import BytesIO
-
-CLASSES = ['akiec', 'bcc', 'bkl', 'df', 'mel', 'nv', 'vasc']
-CLASS_NAMES = {
-    'akiec': 'Actinic keratoses',
-    'bcc': 'Basal cell carcinoma',
-    'bkl': 'Benign keratosis-like lesions',
-    'df': 'Dermatofibroma',
-    'mel': 'Melanoma',
-    'nv': 'Melanocytic nevi',
-    'vasc': 'Vascular lesions'
-}
-
-# Training data distribution (from HAM10000)
-CLASS_DISTRIBUTION = {
-    'nv': 6705,  # 67% - Highly overrepresented
-    'mel': 1113,  # 11%
-    'bkl': 1099,  # 11%
-    'bcc': 514,   # 5%
-    'akiec': 327, # 3%
-    'vasc': 142,  # 1.4%
-    'df': 115     # 1.1% - Highly underrepresented
-}
-
-# Model performance metrics (from your test results)
-VIT_METRICS = {
-    'accuracy': 0.4897,
-    'f1_macro': 0.3226,
-    'f1_weighted': 0.5529,
-    'per_class_f1': {
-        'nv': 0.65, 'mel': 0.42, 'bkl': 0.38, 
-        'bcc': 0.35, 'akiec': 0.28, 'vasc': 0.20, 'df': 0.15
-    }
-}
-
-BIOMEDCLIP_METRICS = {
-    'accuracy': 0.5116,
-    'f1_macro': 0.3521,
-    'f1_weighted': 0.5626,
-    'per_class_f1': {
-        'nv': 0.68, 'mel': 0.45, 'bkl': 0.40,
-        'bcc': 0.38, 'akiec': 0.30, 'vasc': 0.22, 'df': 0.18
-    }
-}
-
-# Confusion matrix data (simplified - you can add real data later)
-CONFUSION_MATRIX = np.array([
-    [45, 8, 12, 2, 5, 25, 3],   # akiec
-    [6, 180, 15, 8, 12, 8, 5],  # bcc
-    [10, 12, 420, 5, 8, 35, 2], # bkl
-    [3, 5, 8, 90, 2, 6, 1],     # df
-    [8, 15, 10, 3, 470, 45, 2], # mel
-    [15, 6, 28, 4, 35, 4450, 8],# nv
-    [2, 3, 5, 1, 2, 8, 120]     # vasc
-])
-
-device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
-processor = ViTImageProcessor.from_pretrained('google/vit-base-patch16-224')
-
-print("Loading models...")
-vit_model = ViTForImageClassification.from_pretrained('best_model', local_files_only=True)
-biomedclip_model = ViTForImageClassification.from_pretrained('best_model_biomedclip_maximal', local_files_only=True)
-
-vit_model = vit_model.to(device).eval()
-biomedclip_model = biomedclip_model.to(device).eval()
-print("Models loaded!")
-
-def create_confusion_matrix_plot():
-    """Generate confusion matrix visualization"""
-    plt.figure(figsize=(10, 8))
-    sns.heatmap(CONFUSION_MATRIX, annot=True, fmt='d', cmap='Blues',
-                xticklabels=[CLASS_NAMES[c] for c in CLASSES],
-                yticklabels=[CLASS_NAMES[c] for c in CLASSES])
-    plt.title('Model Confusion Matrix\nShows which classes get misclassified as what', fontsize=14, pad=20)
-    plt.ylabel('True Label', fontsize=12)
-    plt.xlabel('Predicted Label', fontsize=12)
-    plt.xticks(rotation=45, ha='right')
-    plt.yticks(rotation=0)
-    plt.tight_layout()
-    
-    buf = BytesIO()
-    plt.savefig(buf, format='png', dpi=100, bbox_inches='tight')
-    plt.close()
-    buf.seek(0)
-    return Image.open(buf)
-
-def create_data_distribution_plot():
-    """Visualize training data class imbalance"""
-    fig, (ax1, ax2) = plt.subplots(1, 2, figsize=(14, 5))
-    
-    # Bar chart
-    classes_display = [CLASS_NAMES[c] for c in CLASSES]
-    counts = [CLASS_DISTRIBUTION[c] for c in CLASSES]
-    colors = ['#e74c3c' if c < 500 else '#3498db' for c in counts]
-    
-    ax1.barh(classes_display, counts, color=colors)
-    ax1.set_xlabel('Number of Training Images', fontsize=12)
-    ax1.set_title('Training Data Distribution\n(Class Imbalance)', fontsize=14)
-    ax1.axvline(x=np.mean(counts), color='green', linestyle='--', label=f'Mean: {int(np.mean(counts))}')
-    ax1.legend()
-    
-    # Pie chart
-    ax2.pie(counts, labels=classes_display, autopct='%1.1f%%', startangle=90)
-    ax2.set_title('Class Distribution Percentage', fontsize=14)
-    
-    plt.tight_layout()
-    buf = BytesIO()
-    plt.savefig(buf, format='png', dpi=100, bbox_inches='tight')
-    plt.close()
-    buf.seek(0)
-    return Image.open(buf)
-
-def create_performance_comparison():
-    """Compare model performance across classes"""
-    fig, ax = plt.subplots(figsize=(12, 6))
-    
-    classes_display = [CLASS_NAMES[c] for c in CLASSES]
-    vit_scores = [VIT_METRICS['per_class_f1'][c] for c in CLASSES]
-    bio_scores = [BIOMEDCLIP_METRICS['per_class_f1'][c] for c in CLASSES]
-    
-    x = np.arange(len(classes_display))
-    width = 0.35
-    
-    ax.bar(x - width/2, vit_scores, width, label='ViT Model', alpha=0.8, color='#3498db')
-    ax.bar(x + width/2, bio_scores, width, label='BiomedCLIP Model', alpha=0.8, color='#2ecc71')
-    
-    ax.set_ylabel('F1 Score', fontsize=12)
-    ax.set_title('Per-Class Model Performance Comparison', fontsize=14, pad=20)
-    ax.set_xticks(x)
-    ax.set_xticklabels(classes_display, rotation=45, ha='right')
-    ax.legend()
-    ax.grid(axis='y', alpha=0.3)
-    ax.set_ylim(0, 1)
-    
-    plt.tight_layout()
-    buf = BytesIO()
-    plt.savefig(buf, format='png', dpi=100, bbox_inches='tight')
-    plt.close()
-    buf.seek(0)
-    return Image.open(buf)
-
-def predict_with_model(image, model, model_name):
-    """Make prediction with a specific model"""
-    inputs = processor(images=image, return_tensors="pt")
-    inputs = {k: v.to(device) for k, v in inputs.items()}
-    
-    with torch.no_grad():
-        outputs = model(**inputs)
-        probs = torch.nn.functional.softmax(outputs.logits, dim=-1)[0].cpu().numpy()
-    
-    results = {CLASS_NAMES[CLASSES[i]]: float(probs[i]) for i in range(len(CLASSES))}
-    
-    # Get top prediction
-    top_idx = int(np.argmax(probs))
-    top_prob = float(probs[top_idx])
-    top_class = CLASS_NAMES[CLASSES[top_idx]]
-    
-    # Calculate entropy
-    entropy = -sum(p * np.log(p + 1e-10) for p in probs if p > 0.01)
-    max_entropy = np.log(7)
-    normalized_entropy = entropy / max_entropy
-    
-    return results, top_class, top_prob, normalized_entropy, probs
-
-def analyze_image(image):
-    """Complete analysis with both models"""
-    if image is None:
-        return {}, {}, "", "", None, None, None
-    
-    # Get predictions from both models
-    vit_results, vit_top, vit_conf, vit_ent, vit_probs = predict_with_model(image, vit_model, "ViT")
-    bio_results, bio_top, bio_conf, bio_ent, bio_probs = predict_with_model(image, biomedclip_model, "BiomedCLIP")
-    
-    # Comparison analysis
-    agreement = "✅ Models Agree" if vit_top == bio_top else "⚠️ Models Disagree"
-    
-    comparison = f"""
-    ### 🔄 Model Comparison Analysis
-    
-    **{agreement}**
-    
-    | Metric | ViT Model | BiomedCLIP Model |
-    |--------|-----------|------------------|
-    | Top Prediction | {vit_top} | {bio_top} |
-    | Confidence | {vit_conf*100:.1f}% | {bio_conf*100:.1f}% |
-    | Uncertainty | {vit_ent:.1%} | {bio_ent:.1%} |
-    
-    **Educational Insight:**
-    """
-    
-    if vit_top == bio_top:
-        comparison += f"\n- Both models predict **{vit_top}**\n"
-        comparison += f"- Agreement suggests strong visual features for this class\n"
-        if abs(vit_conf - bio_conf) > 0.2:
-            comparison += f"- However, confidence differs by {abs(vit_conf - bio_conf)*100:.0f}%!\n"
-            comparison += f"- Shows models use different decision strategies\n"
-    else:
-        comparison += f"\n- **Disagreement reveals ambiguity!**\n"
-        comparison += f"- ViT sees: {vit_top} ({vit_conf*100:.0f}%)\n"
-        comparison += f"- BiomedCLIP sees: {bio_top} ({bio_conf*100:.0f}%)\n"
-        comparison += f"- This lesion has overlapping features between classes\n"
-        comparison += f"- Real-world medical AI must handle such uncertainty\n"
-    
-    # Detailed educational insights
-    insights = f"""
-    ### 📊 Deep Learning Analysis
-    
-    **Prediction Entropy:**
-    - ViT: {vit_ent:.3f} (uncertainty: {vit_ent:.1%})
-    - BiomedCLIP: {bio_ent:.3f} (uncertainty: {bio_ent:.1%})
-    
-    **What This Teaches:**
-    """
-    
-    if max(vit_ent, bio_ent) > 0.8:
-        insights += "\n⚠️ **High Uncertainty Detected**\n"
-        insights += "- Models are confused between multiple classes\n"
-        insights += "- Image may have ambiguous features\n"
-        insights += "- Demonstrates why ensemble methods matter\n"
-        insights += "- In practice, this case would need expert review\n"
-    
-    insights += f"\n**Class Probabilities Breakdown:**\n\n"
-    insights += "| Class | ViT | BiomedCLIP | Difference |\n"
-    insights += "|-------|-----|------------|------------|\n"
-    for i, cls in enumerate(CLASSES):
-        diff = abs(vit_probs[i] - bio_probs[i])
-        insights += f"| {CLASS_NAMES[cls]} | {vit_probs[i]*100:.1f}% | {bio_probs[i]*100:.1f}% | {diff*100:.1f}% |\n"
-    
-    insights += f"\n**Training Data Context:**\n"
-    insights += f"- {CLASS_NAMES[CLASSES[np.argmax(vit_probs)]]} had {CLASS_DISTRIBUTION[CLASSES[np.argmax(vit_probs)]]} training samples\n"
-    insights += f"- Rare classes (df, vasc) often get lower confidence\n"
-    insights += f"- Models are biased toward common classes (nv: 67% of data)\n"
-    
-    # Get static visualizations
-    confusion_plot = create_confusion_matrix_plot()
-    distribution_plot = create_data_distribution_plot()
-    performance_plot = create_performance_comparison()
-    
-    return (vit_results, bio_results, comparison, insights, 
-            confusion_plot, distribution_plot, performance_plot)
-
-# Create the comprehensive interface
-with gr.Blocks(title="Medical Image AI Lab - Complete", theme="soft") as demo:
-    gr.Markdown("""
-    # 🔬 Medical Image AI Lab - Complete Educational Platform
-    ### Learn How Computer Vision Models Analyze, Compare, and Misclassify Medical Images
-    
-    **For ML/AI Students, Researchers, and Educators**
-    
-    This platform provides deep insights into:
-    - Multi-model comparison and disagreement analysis
-    - Class imbalance effects on predictions
-    - Performance metrics across different lesion types
-    - Real confusion matrices from model evaluation
-    - Training data distribution impact
-    """)
-    
-    with gr.Row():
-        with gr.Column(scale=1):
-            image_input = gr.Image(type="pil", label="📸 Upload Dermoscopy Image")
-            analyze_btn = gr.Button("🔍 Complete Analysis", variant="primary", size="lg")
-            
-            gr.Markdown("""
-            ### 💡 What Makes This Educational
-            
-            **Dual Model Comparison:**
-            - See how different architectures make different decisions
-            - Observe when models agree vs disagree
-            - Understand confidence calibration
-            
-            **Visual Explanations:**
-            - Confusion matrices reveal systematic errors
-            - Performance charts expose class-specific weaknesses
-            - Data distribution shows training bias
-            
-            **Real-World Context:**
-            - Training data imbalance visualization
-            - Per-class performance metrics
-            - Entropy and uncertainty quantification
-            """)
-        
-        with gr.Column(scale=1):
-            with gr.Tabs():
-                with gr.Tab("🎯 Predictions"):
-                    gr.Markdown("### ViT Model Predictions")
-                    vit_output = gr.Label(num_top_classes=7, label="ViT Probabilities")
-                    
-                    gr.Markdown("### BiomedCLIP Model Predictions")
-                    bio_output = gr.Label(num_top_classes=7, label="BiomedCLIP Probabilities")
-                
-                with gr.Tab("🔄 Comparison"):
-                    comparison_output = gr.Markdown()
-                
-                with gr.Tab("📊 Deep Analysis"):
-                    insights_output = gr.Markdown()
-                
-                with gr.Tab("📈 Performance"):
-                    gr.Markdown("### Model Confusion Matrix")
-                    confusion_output = gr.Image(label="Where the model gets confused")
-                    
-                    gr.Markdown("### Training Data Distribution")
-                    distribution_output = gr.Image(label="Class imbalance in training")
-                    
-                    gr.Markdown("### Per-Class Performance")
-                    performance_output = gr.Image(label="F1 scores by lesion type")
-    
-    gr.Markdown("""
-    ---
-    
-    ## 📚 Understanding the Platform
-    
-    ### Model Architectures
-    
-    **ViT (Vision Transformer)**
-    - Pre-trained on ImageNet
-    - Fine-tuned on HAM10000
-    - Test Accuracy: 48.97%
-    
-    **BiomedCLIP**  
-    - Pre-trained on biomedical images
-    - Specialized for medical imaging
-    - Test Accuracy: 51.16%
-    
-    **Key Insight:** Only 2.2% improvement despite medical specialization! This teaches us:
-    - Domain-specific pre-training helps, but isn't magic
-    - Dataset quality matters more than model choice
-    - Class imbalance remains the dominant challenge
-    
-    ### Why 51% is Actually Good (Educational Context)
-    
-    - Random guessing: 14.3%
-    - Our best model: 51.16%
-    - **3.6x better than random**
-    - 73% of maximum possible improvement
-    
-    ### Common Failure Patterns (Learning Opportunities)
-    
-    1. **Nevi Bias** - Model over-predicts common class (67% of training data)
-    2. **Rare Class Struggles** - df and vasc have <2% representation  
-    3. **Visual Similarity** - Melanoma vs nevi are genuinely difficult
-    4. **Overconfidence** - Model can be 90% sure and still wrong
-    
-    ### Experiments to Try
-    
-    **Test Model Robustness:**
-    - Upload images with different lighting
-    - Try blurry or partially obscured lesions
-    - Test on edge cases (very small or large lesions)
-    
-    **Explore Model Disagreement:**
-    - Find images where models disagree strongly
-    - Analyze which classes cause most confusion
-    - Compare confidence levels between models
-    
-    **Study Failure Modes:**
-    - Look for patterns in misclassifications
-    - Check if models fail on same images
-    - Examine probability distributions for failed predictions
-    
-    ---
-    
-    ## �� For Educators & Students
-    
-    ### Classroom Applications
-    
-    **Teach Key ML Concepts:**
-    - Confusion matrices and error analysis
-    - Class imbalance and sampling strategies
-    - Model calibration and confidence
-    - Transfer learning effectiveness
-    - Multi-model ensemble benefits
-    
-    **Discussion Questions:**
-    - Why does medical AI need higher accuracy than 51%?
-    - How would you improve this model?
-    - What metrics matter most in medical contexts?
-    - When should models abstain from predictions?
-    
-    ### Research Directions
-    
-    - Implement ensemble methods
-    - Try different augmentation strategies
-    - Experiment with class balancing techniques
-    - Develop uncertainty quantification methods
-    - Study transfer learning from different domains
-    
-    ---
-    
-    ## ⚠️ Critical Disclaimer
-    
-    **EDUCATIONAL USE ONLY - NOT FOR MEDICAL DIAGNOSIS**
-    
-    This platform demonstrates ML concepts and limitations.  
-    It is NOT:
-    - ❌ A medical device
-    - ❌ For clinical diagnosis  
-    - ❌ For treatment decisions
-    - ❌ A replacement for dermatologists
-    
-    **For actual medical concerns, always consult a board-certified dermatologist.**
-    
-    ---
-    
-    ## 📖 Additional Resources
-    
-    - [HAM10000 Dataset Paper](https://arxiv.org/abs/1803.10417)
-    - [Vision Transformers Explained](https://arxiv.org/abs/2010.11929)
-    - [Medical AI Challenges](https://www.nature.com/articles/s41591-020-0842-6)
-    - [Model Calibration in Deep Learning](https://arxiv.org/abs/1706.04599)
-    
-    **Built for ML Education | Models: ViT (48.97%) & BiomedCLIP (51.16%) | Dataset: HAM10000 (10,015 images)**
-    """)
+    gr.Markdown("---\n## ⚠️ Educational Use Only\n\nNOT for medical diagnosis. Consult a dermatologist for medical concerns.")
     
-    # Connect the interface
     analyze_btn.click(
         fn=analyze_image,
         inputs=image_input,
@@ -913,5 +279,4 @@ with gr.Blocks(title="Medical Image AI Lab - Complete", theme="soft") as demo:
                 confusion_output, distribution_output, performance_output]
     )
 
-if __name__ == "__main__":
-    demo.launch()
+demo.launch()