Create app.py

app.py

@@ -0,0 +1,284 @@
+import gradio as gr
+import pandas as pd
+from huggingface_hub import hf_hub_download
+from PIL import Image
+import zipfile
+import os
+import random
+
+# Global variables
+df = None
+images_dir = None
+
+def setup_dataset():
+    """Download and setup the dataset (called once on startup)"""
+    global df, images_dir
+    
+    print("Loading metadata...")
+    # Load metadata
+    csv_path = hf_hub_download(
+        repo_id="Deva8/Generative-VQA-V2-Curated",
+        filename="main_metadata.csv",
+        repo_type="dataset"
+    )
+    df = pd.read_csv(csv_path)
+    
+    print("Downloading images zip (this may take a few minutes)...")
+    # Download zip file
+    zip_path = hf_hub_download(
+        repo_id="Deva8/Generative-VQA-V2-Curated",
+        filename="gen_vqa_v2-images.zip",
+        repo_type="dataset"
+    )
+    
+    # Extract images
+    images_dir = "./extracted_images"
+    if not os.path.exists(images_dir):
+        print("Extracting images...")
+        os.makedirs(images_dir, exist_ok=True)
+        with zipfile.ZipFile(zip_path, 'r') as zip_ref:
+            zip_ref.extractall(images_dir)
+    
+    print(f"✓ Dataset ready! {len(df)} examples loaded.")
+    return f"Dataset loaded successfully! {len(df):,} examples available."
+
+def get_random_sample():
+    """Get a random sample from the dataset"""
+    if df is None:
+        return None, "Please wait, dataset is loading...", "", ""
+    
+    # Get random row
+    sample = df.sample(1).iloc[0]
+    
+    # Load image
+    img_path = os.path.join(images_dir, sample['file_name'])
+    img = Image.open(img_path)
+    
+    question = sample['question']
+    answer = sample['answer']
+    metadata = f"Image ID: {sample['image_id']} | Question ID: {sample['question_id']}"
+    
+    return img, question, answer, metadata
+
+def search_by_question(query):
+    """Search for questions containing the query"""
+    if df is None:
+        return None, "Dataset not loaded yet", "", ""
+    
+    if not query or len(query.strip()) < 3:
+        return None, "Please enter at least 3 characters to search", "", ""
+    
+    # Search for matching questions
+    matches = df[df['question'].str.contains(query, case=False, na=False)]
+    
+    if len(matches) == 0:
+        return None, f"No questions found containing '{query}'", "", ""
+    
+    # Get random match
+    sample = matches.sample(1).iloc[0]
+    
+    # Load image
+    img_path = os.path.join(images_dir, sample['file_name'])
+    img = Image.open(img_path)
+    
+    question = sample['question']
+    answer = sample['answer']
+    metadata = f"Image ID: {sample['image_id']} | Question ID: {sample['question_id']} | Found {len(matches)} matches"
+    
+    return img, question, answer, metadata
+
+def search_by_answer(query):
+    """Search for specific answers"""
+    if df is None:
+        return None, "Dataset not loaded yet", "", ""
+    
+    if not query or len(query.strip()) < 1:
+        return None, "Please enter an answer to search", "", ""
+    
+    # Search for matching answers
+    matches = df[df['answer'].str.lower() == query.lower().strip()]
+    
+    if len(matches) == 0:
+        return None, f"No examples found with answer '{query}'", "", ""
+    
+    # Get random match
+    sample = matches.sample(1).iloc[0]
+    
+    # Load image
+    img_path = os.path.join(images_dir, sample['file_name'])
+    img = Image.open(img_path)
+    
+    question = sample['question']
+    answer = sample['answer']
+    metadata = f"Image ID: {sample['image_id']} | Question ID: {sample['question_id']} | Found {len(matches)} examples with this answer"
+    
+    return img, question, answer, metadata
+
+def get_statistics():
+    """Get dataset statistics"""
+    if df is None:
+        return "Dataset not loaded yet"
+    
+    stats = f"""
+    # 📊 Dataset Statistics
+    
+    - **Total Examples**: {len(df):,}
+    - **Unique Images**: {df['image_id'].nunique():,}
+    - **Unique Answers**: {df['answer'].nunique():,}
+    
+    ## Top 10 Most Common Answers:
+    
+    """
+    
+    top_answers = df['answer'].value_counts().head(10)
+    for i, (answer, count) in enumerate(top_answers.items(), 1):
+        stats += f"{i}. **{answer}** - {count} examples\n"
+    
+    stats += f"\n## Question Length Distribution:\n"
+    stats += f"- Average: {df['question'].str.split().str.len().mean():.1f} words\n"
+    stats += f"- Min: {df['question'].str.split().str.len().min()} words\n"
+    stats += f"- Max: {df['question'].str.split().str.len().max()} words\n"
+    
+    stats += f"\n## Answer Length Distribution:\n"
+    stats += f"- Average: {df['answer'].str.split().str.len().mean():.2f} words\n"
+    stats += f"- Single word answers: {(df['answer'].str.split().str.len() == 1).sum():,} ({(df['answer'].str.split().str.len() == 1).sum() / len(df) * 100:.1f}%)\n"
+    
+    return stats
+
+# Initialize dataset on startup
+print("Starting dataset setup...")
+setup_status = setup_dataset()
+print(setup_status)
+
+# Create Gradio interface
+with gr.Blocks(title="Generative VQA v2 Dataset Explorer", theme=gr.themes.Soft()) as demo:
+    gr.Markdown("""
+    # 🎯 Generative VQA-V2-Curated Dataset Explorer
+    
+    Explore the **Generative VQA v2 Curated** dataset - a balanced, cleaned version of VQA v2 
+    optimized for generative visual question answering.
+    
+    **Dataset**: [Deva8/Generative-VQA-V2-Curated](https://huggingface.co/datasets/Deva8/Generative-VQA-V2-Curated)
+    
+    ---
+    """)
+    
+    with gr.Tabs():
+        # Tab 1: Random Samples
+        with gr.Tab("🎲 Random Samples"):
+            gr.Markdown("### Click the button to see random examples from the dataset")
+            
+            with gr.Row():
+                random_btn = gr.Button("🔄 Get Random Sample", variant="primary", size="lg")
+            
+            with gr.Row():
+                with gr.Column(scale=1):
+                    random_image = gr.Image(label="Image", type="pil")
+                
+                with gr.Column(scale=1):
+                    random_question = gr.Textbox(label="❓ Question", lines=2)
+                    random_answer = gr.Textbox(label="✅ Answer", lines=1)
+                    random_metadata = gr.Textbox(label="ℹ️ Metadata", lines=1)
+            
+            random_btn.click(
+                fn=get_random_sample,
+                outputs=[random_image, random_question, random_answer, random_metadata]
+            )
+        
+        # Tab 2: Search by Question
+        with gr.Tab("🔍 Search Questions"):
+            gr.Markdown("### Search for questions containing specific keywords")
+            
+            with gr.Row():
+                question_query = gr.Textbox(
+                    label="Search Query",
+                    placeholder="e.g., 'color', 'many', 'wearing', 'holding'",
+                    lines=1
+                )
+                question_search_btn = gr.Button("🔎 Search", variant="primary")
+            
+            with gr.Row():
+                with gr.Column(scale=1):
+                    question_image = gr.Image(label="Image", type="pil")
+                
+                with gr.Column(scale=1):
+                    question_text = gr.Textbox(label="❓ Question", lines=2)
+                    question_answer = gr.Textbox(label="✅ Answer", lines=1)
+                    question_metadata = gr.Textbox(label="ℹ️ Metadata", lines=1)
+            
+            question_search_btn.click(
+                fn=search_by_question,
+                inputs=[question_query],
+                outputs=[question_image, question_text, question_answer, question_metadata]
+            )
+        
+        # Tab 3: Search by Answer
+        with gr.Tab("🎯 Search Answers"):
+            gr.Markdown("### Find examples with specific answers")
+            
+            with gr.Row():
+                answer_query = gr.Textbox(
+                    label="Answer to Search",
+                    placeholder="e.g., 'red', 'cat', '2', 'eating'",
+                    lines=1
+                )
+                answer_search_btn = gr.Button("🔎 Search", variant="primary")
+            
+            gr.Markdown("**Popular answers**: white, black, blue, red, 2, 3, brown, green, pizza, dog")
+            
+            with gr.Row():
+                with gr.Column(scale=1):
+                    answer_image = gr.Image(label="Image", type="pil")
+                
+                with gr.Column(scale=1):
+                    answer_question = gr.Textbox(label="❓ Question", lines=2)
+                    answer_text = gr.Textbox(label="✅ Answer", lines=1)
+                    answer_metadata = gr.Textbox(label="ℹ️ Metadata", lines=1)
+            
+            answer_search_btn.click(
+                fn=search_by_answer,
+                inputs=[answer_query],
+                outputs=[answer_image, answer_question, answer_text, answer_metadata]
+            )
+        
+        # Tab 4: Statistics
+        with gr.Tab("📊 Statistics"):
+            gr.Markdown("### Dataset Statistics and Analysis")
+            
+            stats_btn = gr.Button("📈 Show Statistics", variant="primary")
+            stats_output = gr.Markdown()
+            
+            stats_btn.click(
+                fn=get_statistics,
+                outputs=[stats_output]
+            )
+    
+    gr.Markdown("""
+    ---
+    
+    ## About This Dataset
+    
+    **Generative VQA-V2-Curated** is a cleaned and balanced version of VQA v2:
+    
+    - ✅ Removed yes/no questions
+    - ✅ Balanced answer distribution (max 600 per answer)
+    - ✅ Filtered ambiguous questions
+    - ✅ 135,268 high-quality QA pairs
+    - ✅ 1,251 unique answer classes
+    
+    **License**: CC BY 4.0 (COCO + VQA v2)
+    
+    **Citation**:
+    ```bibtex
+    @misc{devarajan_genvqa_2026,
+      author = {Devarajan},
+      title = {Generative-VQA-V2-Curated},
+      year = {2026},
+      publisher = {Hugging Face},
+    }
+    ```
+    """)
+
+# Launch the app
+if __name__ == "__main__":
+    demo.launch()