app.py

import streamlit as st
import torch
from PIL import Image
from peft import PeftModel
import gc
import os

# Page config
st.set_page_config(
    page_title="Deepfake Image Analyzer",
    page_icon="🔍",
    layout="wide"
)

# App title and description
st.title("Deepfake Image Analyzer")
st.markdown("Upload an image to analyze it for possible deepfake manipulation")

# Function to free up memory
def free_memory():
    gc.collect()
    if torch.cuda.is_available():
        torch.cuda.empty_cache()

# Helper function to check CUDA
def init_device():
    """Set the appropriate device and return it"""
    if torch.cuda.is_available():
        st.sidebar.success("✓ GPU available: Using CUDA")
        return "cuda"
    else:
        st.sidebar.warning("⚠️ No GPU detected: Using CPU (analysis will be slow)")
        return "cpu"

# Set device
device = init_device()

@st.cache_resource
def load_model():
    """Load model using Unsloth, similar to your notebook code"""
    try:
        # Import Unsloth here to ensure it's loaded when needed
        from unsloth import FastVisionModel
        
        st.info("Loading base model and tokenizer using Unsloth...")
        
        # Use the same model ID and loading approach that worked in your notebook
        base_model_id = "unsloth/llama-3.2-11b-vision-instruct-unsloth-bnb-4bit"
        model, tokenizer = FastVisionModel.from_pretrained(
            base_model_id,
            load_in_4bit=True,
            torch_dtype=torch.float16,
        )
        
        # Set to inference mode
        FastVisionModel.for_inference(model)
        
        # Load the fine-tuned adapter
        st.info("Loading adapter...")
        adapter_id = "saakshigupta/deepfake-explainer-1"
        model = PeftModel.from_pretrained(model, adapter_id)
        
        return model, tokenizer
    
    except Exception as e:
        st.error(f"Error loading model: {str(e)}")
        st.exception(e)
        return None, None

# Function to fix cross-attention masks
def fix_processor_outputs(inputs):
    """Fix cross-attention mask dimensions if needed"""
    if 'cross_attention_mask' in inputs and 0 in inputs['cross_attention_mask'].shape:
        batch_size, seq_len, _, num_tiles = inputs['cross_attention_mask'].shape
        visual_features = 6404  # The exact dimension used in training
        new_mask = torch.ones(
            (batch_size, seq_len, visual_features, num_tiles),
            device=inputs['cross_attention_mask'].device
        )
        inputs['cross_attention_mask'] = new_mask
        return True, inputs
    return False, inputs

# Create sidebar with options
with st.sidebar:
    st.header("Options")
    temperature = st.slider("Temperature", min_value=0.1, max_value=1.0, value=0.7, step=0.1,
                           help="Higher values make output more random, lower values more deterministic")
    max_length = st.slider("Maximum response length", min_value=100, max_value=1000, value=500, step=50)
    
    custom_prompt = st.text_area(
        "Custom instruction (optional)",
        value="Analyze this image and determine if it's a deepfake. Provide both technical and non-technical explanations.",
        height=100
    )
    
    st.markdown("### About")
    st.markdown("""
    This app uses a fine-tuned Llama 3.2 Vision model to detect and explain deepfakes.
    
    The analyzer looks for:
    - Inconsistencies in facial features
    - Unusual lighting or shadows
    - Unnatural blur patterns
    - Artifacts around edges
    - Texture inconsistencies
    
    Model by [saakshigupta](https://huggingface.co/saakshigupta/deepfake-explainer-1)
    """)

# Load model on startup
with st.spinner("Loading model... this may take a minute."):
    try:
        model, tokenizer = load_model()
        if model is not None and tokenizer is not None:
            st.session_state['model'] = model
            st.session_state['tokenizer'] = tokenizer
            st.success("Model loaded successfully!")
        else:
            st.error("Failed to load model.")
    except Exception as e:
        st.error(f"Error during model loading: {str(e)}")
        st.exception(e)

# Main content area - file uploader
uploaded_file = st.file_uploader("Choose an image...", type=["jpg", "jpeg", "png"])

# Check if model is loaded
model_loaded = 'model' in st.session_state and st.session_state['model'] is not None

if uploaded_file is not None and model_loaded:
    # Display the image
    image = Image.open(uploaded_file).convert('RGB')
    st.image(image, caption="Uploaded Image", use_column_width=True)
    
    # Analyze button
    if st.button("Analyze Image"):
        with st.spinner("Analyzing the image... This may take 15-30 seconds"):
            try:
                # Get components from session state
                model = st.session_state['model']
                tokenizer = st.session_state['tokenizer']
                
                # Format the message for Unsloth - same as your notebook
                messages = [
                    {"role": "user", "content": [
                        {"type": "image"},
                        {"type": "text", "text": custom_prompt}
                    ]}
                ]
                
                # Apply chat template
                input_text = tokenizer.apply_chat_template(messages, add_generation_prompt=True)
                
                # Process with image
                inputs = tokenizer(
                    image,
                    input_text,
                    add_special_tokens=False,
                    return_tensors="pt",
                ).to(model.device)
                
                # Apply the cross-attention fix
                fixed, inputs = fix_processor_outputs(inputs)
                if fixed:
                    st.info("Fixed cross-attention mask dimensions")
                
                # Generate analysis
                with torch.no_grad():
                    output_ids = model.generate(
                        **inputs,
                        max_new_tokens=max_length,
                        temperature=temperature,
                        top_p=0.9
                    )
                
                # Decode the output
                response = tokenizer.decode(output_ids[0], skip_special_tokens=True)
                
                # Extract the model's response
                # Format might be different from processor.decode, check the output 
                if "assistant" in response:
                    result = response.split("assistant")[-1].strip()
                else:
                    result = response
                
                # Display result in a nice format
                st.success("Analysis complete!")
                
                # Show technical and non-technical explanations separately if they exist
                if "Technical Explanation:" in result and "Non-Technical Explanation:" in result:
                    technical, non_technical = result.split("Non-Technical Explanation:")
                    technical = technical.replace("Technical Explanation:", "").strip()
                    
                    col1, col2 = st.columns(2)
                    with col1:
                        st.subheader("Technical Analysis")
                        st.write(technical)
                    
                    with col2:
                        st.subheader("Simple Explanation")
                        st.write(non_technical)
                else:
                    st.subheader("Analysis Result")
                    st.write(result)
                
                # Free memory after analysis
                free_memory()
                
            except Exception as e:
                st.error(f"Error analyzing image: {str(e)}")
                st.exception(e)
elif not model_loaded and uploaded_file is not None:
    st.warning("Model not loaded correctly. Try refreshing the page.")
else:
    st.info("Please upload an image to begin analysis")

# Add footer
st.markdown("---")
st.markdown("Deepfake Image Analyzer")