app.py

"""
ISL Sign Language Translation - TechMatrix Solvers Initiative
Main Streamlit Application

Developed by: TechMatrix Solvers Team
- Abhay Gupta (Team Lead)
- Kripanshu Gupta (Backend Developer) 
- Dipanshu Patel (UI/UX Designer)
- Bhumika Patel (Deployment & Female Presenter)

Institution: Shri Ram Group of Institutions
"""

import streamlit as st

# Configure Streamlit page first
st.set_page_config(
    page_title="ISL Translation - TechMatrix Solvers",
    page_icon="🤟",
    layout="wide",
    initial_sidebar_state="expanded"
)

# Show loading message
st.write("🚀 TechMatrix Solvers ISL Translator Loading...")

# Import dependencies with error handling
try:
    import os
    # Set environment variables for better compatibility
    os.environ["KERAS_BACKEND"] = "tensorflow"
    os.environ["HF_HOME"] = "/tmp/huggingface"
    os.environ["TRANSFORMERS_CACHE"] = "/tmp/transformers"
    
    # Core imports
    import numpy as np
    import pandas as pd
    import tempfile
    import time
    from PIL import Image
    import subprocess
    from typing import NamedTuple
    import json
    import shutil
    import platform
    import uuid
    
    # Try OpenCV import
    try:
        import cv2
    except Exception as cv_error:
        st.warning(f"OpenCV import issue: {cv_error}")
        cv2 = None
    
    # Try ML library imports
    try:
        import keras
        from keras.models import Sequential
        from keras.layers import LSTM, Dense, Bidirectional, Dropout, Input, BatchNormalization
    except Exception as keras_error:
        st.warning(f"Keras import issue: {keras_error}")
        keras = None
    
    # Try video processing
    try:
        import ffmpeg
    except Exception as ffmpeg_error:
        st.warning(f"FFmpeg import issue: {ffmpeg_error}")
        ffmpeg = None
    
    # Try HuggingFace Hub
    try:
        from huggingface_hub import hf_hub_download
    except Exception as hf_error:
        st.warning(f"HuggingFace Hub import issue: {hf_error}")
        hf_hub_download = None
    
    # Try custom modules with fallback - import each module separately for better error handling
    pose_models = None
    expression_mapping = None
    isl_processor = None
    utils = None
    
    # Try importing pose_models
    try:
        from pose_models import create_bodypose_model, create_handpose_model
        pose_models = True
        st.success("✅ Pose models imported successfully")
    except Exception as pose_error:
        st.warning(f"Pose models import issue: {pose_error}")
        pose_models = None
    
    # Try importing expression_mapping
    try:
        from expression_mapping import expression_mapping
        st.success("✅ Expression mapping imported successfully")
    except Exception as expr_error:
        st.warning(f"Expression mapping import issue: {expr_error}")
        # Fallback expression mapping
        expression_mapping = {
            'hello': 0, 'thank_you': 1, 'please': 2, 'sorry': 3, 'help': 4,
            'good': 5, 'bad': 6, 'yes': 7, 'no': 8, 'water': 9,
            'food': 10, 'home': 11, 'work': 12, 'school': 13, 'family': 14
        }
    
    # Try importing ISL processor
    try:
        from isl_processor import ISLTranslationModel
        isl_processor = True
        st.success("✅ ISL processor imported successfully")
    except Exception as isl_error:
        st.warning(f"ISL processor import issue: {isl_error}")
        isl_processor = None
    
    # Try importing pose_utils
    try:
        import pose_utils as utils
        st.success("✅ Pose utils imported successfully")
    except Exception as utils_error:
        st.warning(f"Pose utils import issue: {utils_error}")
        utils = None
    
    st.success("✅ Core dependencies loaded successfully!")
except ImportError as e:
    st.error(f"❌ Critical import error: {e}")
    st.error("Running in fallback mode with limited functionality.")

# Ensure we have utils available globally after the main import block
if utils is None:
    try:
        import pose_utils as utils
        st.info("ℹ️ Pose utils loaded on secondary attempt")
    except ImportError as utils_error:
        st.error(f"❌ Failed to import pose_utils: {utils_error}")
        utils = None

# Ensure expression_mapping is available and create index-to-label mapping
if expression_mapping is None:
    st.warning("⚠️ Using fallback expression mapping")
    expression_mapping = {
        0: 'hello', 1: 'thank_you', 2: 'please', 3: 'sorry', 4: 'help',
        5: 'good', 6: 'bad', 7: 'yes', 8: 'no', 9: 'water',
        10: 'food', 11: 'home', 12: 'work', 13: 'school', 14: 'family'
    }

# Create index-to-label mapping function for safe access
def get_sign_label(index):
    """Safely get sign label from prediction index"""
    if isinstance(expression_mapping, dict):
        return expression_mapping.get(int(index), f'unknown_sign_{index}')
    else:
        return f'sign_{index}'

# System information will be shown in About section


class VideoProbeResult(NamedTuple):
    """Structure for video probe results"""
    return_code: int
    json: str
    error: str


def probe_video_info(file_path) -> VideoProbeResult:
    """
    Probe video file for metadata using FFprobe
    
    Args:
        file_path: Path to video file
        
    Returns:
        VideoProbeResult containing metadata
    """
    command_array = [
        "ffprobe",
        "-v", "quiet",
        "-print_format", "json",
        "-show_format",
        "-show_streams",
        file_path
    ]
    result = subprocess.run(
        command_array, 
        stdout=subprocess.PIPE, 
        stderr=subprocess.PIPE, 
        universal_newlines=True
    )
    return VideoProbeResult(
        return_code=result.returncode,
        json=result.stdout,
        error=result.stderr
    )


# Define feature columns for time series processing
body_features = [f'bodypeaks_x_{i}' for i in range(15)] + [f'bodypeaks_y_{i}' for i in range(15)]
hand0_features = [f'hand0peaks_x_{i}' for i in range(21)] + [f'hand0peaks_y_{i}' for i in range(21)] + [f'hand0peaks_peaktxt{i}' for i in range(21)]
hand1_features = [f'hand1peaks_x_{i}' for i in range(21)] + [f'hand1peaks_y_{i}' for i in range(21)] + [f'hand1peaks_peaktxt{i}' for i in range(21)]

feature_columns_processed = body_features + hand0_features + hand1_features
label_columns = ['Expression_encoded']


@st.cache_resource
def create_time_series_sequences(isl_data, feature_columns, label_columns, window_size=20):
    """
    Creates time series sequences from DataFrame with specified window size
    
    Args:
        isl_data: Input DataFrame with ISL data
        feature_columns: List of feature column names
        label_columns: List of label column names  
        window_size: Size of temporal window for sequence creation
        
    Returns:
        tuple: (X_sequences, y_sequences) for training/inference
    """
    if isl_data.empty:
        return [], []

    X_sequences = []
    y_sequences = []
    
    for group, file_df in isl_data.groupby(['Type', 'Expression_encoded', 'FileName']):
        expr_type, expression, filename = group
        
        # Create blank frame for padding
        blank_frame = np.zeros((1, 156))
        
        for idx, window_data in enumerate([file_df[i:i+window_size] for i in range(0, file_df.shape[0], 1)]):
            if window_data.shape[0] < window_size:
                # Pad sequence with blank frames at the beginning
                padding_needed = window_size - window_data.shape[0]
                padded_sequence = np.concatenate(
                    (np.repeat(blank_frame, padding_needed, axis=0), 
                     window_data[feature_columns].values), 
                    axis=0
                )
                X_sequences.append(padded_sequence)
                y_sequences.append(expression)
                continue
            
            X_sequences.append(window_data[feature_columns].values)
            y_sequences.append(expression)

    return X_sequences, y_sequences


# Global translation model variable
translation_model = None


@st.cache_resource
def load_translation_model():
    """
    Load and configure the LSTM translation model
    
    Returns:
        Configured Keras Sequential model for ISL translation or None if failed
    """
    try:
        if keras is None or hf_hub_download is None:
            st.warning("Keras or HuggingFace Hub not available. Model loading skipped.")
            return None
            
        # Download pre-trained model file
        model_file = hf_hub_download(
            repo_id="sunilsarolkar/isl-translation-model",
            filename="isl_model_final.keras"
        )
        
        # Try to load the complete model first
        try:
            model = keras.models.load_model(model_file)
            st.success("✅ Model loaded successfully from saved file")
            return model
        except Exception as load_error:
            st.warning(f"Failed to load complete model: {load_error}")
            st.info("Attempting to build model architecture and load weights...")
            
            # Fallback: Build model architecture and load weights
            model = Sequential()
            model.add(Input(shape=((20, 156))))
            model.add(keras.layers.Masking(mask_value=0.))
            model.add(BatchNormalization())
            model.add(Bidirectional(LSTM(32, recurrent_dropout=0.2, return_sequences=True)))
            
            model.add(Dropout(0.2))
            model.add(Bidirectional(LSTM(32, recurrent_dropout=0.2)))
            
            model.add(keras.layers.Activation('elu'))
            model.add(Dense(32, use_bias=False, kernel_initializer='he_normal'))
            
            model.add(BatchNormalization())
            model.add(Dropout(0.2))
            model.add(keras.layers.Activation('elu'))
            model.add(Dense(32, kernel_initializer='he_normal', use_bias=False))
            
            model.add(BatchNormalization())
            model.add(keras.layers.Activation('elu'))
            model.add(Dropout(0.2))
            # Determine number of classes - use 167 for the full dataset or fallback size
            num_classes = len(list(expression_mapping.keys())) if expression_mapping else 167
            model.add(Dense(num_classes, activation='softmax'))
            
            # Try to load weights
            model.load_weights(model_file)
            st.success("✅ Model architecture built and weights loaded successfully")
            return model
        
    except Exception as e:
        st.error(f"Failed to load translation model: {e}")
        return None


# Load test data
@st.cache_data
def load_test_data():
    """Load test dataset and file information"""
    testing_cleaned_path = hf_hub_download(
        repo_id="sunilsarolkar/isl-test-data",
        filename="testing_cleaned.csv",
        repo_type="dataset"
    )
    
    test_files_path = hf_hub_download(
        repo_id="sunilsarolkar/isl-test-data", 
        filename="test_files.csv",
        repo_type="dataset"
    )
    
    testing_df = pd.read_csv(testing_cleaned_path)
    test_files_df = pd.read_csv(test_files_path)
    
    return testing_df, test_files_df


# Test data will be loaded when needed


class VideoWriter:
    """Custom video writer using FFmpeg for better compatibility"""
    
    def __init__(self, output_file, input_fps, input_framesize, input_pix_fmt, input_vcodec):
        self.ff_process = (
            ffmpeg
            .input('pipe:',
                   format='rawvideo',
                   pix_fmt="bgr24",
                   s=f'{input_framesize[1]}x{input_framesize[0]}',
                   r=input_fps)
            .output(output_file, pix_fmt=input_pix_fmt, vcodec=input_vcodec)
            .overwrite_output()
            .run_async(pipe_stdin=True)
        )

    def write_frame(self, frame):
        """Write a single frame to the video"""
        self.ff_process.stdin.write(frame.tobytes())

    def close(self):
        """Close the video writer"""
        self.ff_process.stdin.close()
        self.ff_process.wait()


def calculate_weighted_average(numbers, weights):
    """
    Calculate weighted average of numbers
    
    Args:
        numbers: List of numbers
        weights: List of weights
        
    Returns:
        float: Weighted average
    """
    if sum(weights) == 0:
        return 0
    return sum(x * y for x, y in zip(numbers, weights)) / sum(weights)


@st.cache_data
def resize_image(image, width=None, height=None, interpolation=cv2.INTER_AREA):
    """
    Resize image maintaining aspect ratio
    
    Args:
        image: Input image
        width: Target width
        height: Target height
        interpolation: OpenCV interpolation method
        
    Returns:
        Resized image
    """
    dimensions = None
    (h, w) = image.shape[:2]

    if width is None and height is None:
        return image

    if width is None:
        ratio = height / float(h)
        dimensions = (int(w * ratio), height)
    else:
        ratio = width / float(w)
        dimensions = (width, int(h * ratio))

    resized = cv2.resize(image, dimensions, interpolation=interpolation)
    return resized


# Page configuration already set at the top

st.title('🤟 ISL Sign Language Translation - TechMatrix Solvers Initiative')

# Add custom CSS for sidebar styling
st.markdown(
    """
    <style>
    [data-testid="stSidebar"][aria-expanded="true"] > div:first-child {
        width: 350px;
    }
    [data-testid="stSidebar"][aria-expanded="false"] > div:first-child {
        width: 350px;
        margin-left: -350px;
    }
    
    .team-info {
        background: linear-gradient(135deg, #667eea 0%, #764ba2 100%);
        color: white;
        padding: 1rem;
        border-radius: 0.5rem;
        margin: 1rem 0;
    }
    
    .tech-matrix-header {
        background: linear-gradient(90deg, #1e3a8a, #7c3aed);
        color: white;
        padding: 1rem;
        border-radius: 0.5rem;
        text-align: center;
        margin-bottom: 1rem;
    }
    </style>
    """,
    unsafe_allow_html=True,
)

# Add team branding header
st.markdown(
    """
    <div class="tech-matrix-header">
        <h2>🚀 TechMatrix Solvers</h2>
        <p>Innovating Accessible Technology Solutions</p>
    </div>
    """, 
    unsafe_allow_html=True
)

# Sidebar configuration
st.sidebar.title('🤟 ISL Translation System')
st.sidebar.subheader('Configuration')

# Team information in sidebar
st.sidebar.markdown(
    """
    <div class="team-info">
    <h3>👨‍💻 Development Team</h3>
    <ul>
    <li><strong>Abhay Gupta</strong> - Team Lead</li>
    <li><strong>Kripanshu Gupta</strong> - Backend Dev</li>
    <li><strong>Dipanshu Patel</strong> - UI/UX Designer</li>
    <li><strong>Bhumika Patel</strong> - Deployment</li>
    </ul>
    <p><em>Shri Ram Group of Institutions</em></p>
    </div>
    """, 
    unsafe_allow_html=True
)

# Initialize frame-wise outputs storage
frame_predictions = {}

# Application mode selection
app_mode = st.sidebar.selectbox(
    'Choose Application Mode',
    ['About Project', 'Test Video Translation']
)

if app_mode == 'About Project':
    st.markdown(
        """
        ## 🎯 Project Overview
        
        Welcome to the **ISL Sign Language Translation System** developed by **TechMatrix Solvers**. 
        This cutting-edge application demonstrates real-time Indian Sign Language recognition and 
        translation using advanced deep learning techniques.
        
        ### 🏗️ Technical Architecture
        
        Our system combines multiple state-of-the-art technologies:
        
        1. **Body Pose Estimation**: 25-point skeletal tracking using OpenPose
        2. **Hand Landmark Detection**: 21-point hand keypoint identification  
        3. **Temporal Modeling**: Bidirectional LSTM networks for sequence analysis
        4. **Real-time Processing**: Optimized inference pipeline for live translation
        """
    )
    
    st.markdown(
        """
        ### 📊 Dataset Information
        
        Our model is trained on the comprehensive [INCLUDE dataset](https://zenodo.org/records/4010759):
        """
    )
    
    # Dataset statistics table
    dataset_stats = {
        "Metric": [
            "Categories", "Total Words", "Training Videos", 
            "Avg Videos/Class", "Avg Video Length", "Resolution", "Frame Rate"
        ],
        "Value": [
            "15", "263", "4,257", "16.3", "2.57s", "1920x1080", "25fps"
        ]
    }
    st.table(pd.DataFrame(dataset_stats))
    
    # Display dataset processing visualization
    try:
        categories_image = np.array(Image.open('original_project/categories_processed.png'))
        st.image(categories_image, caption="📈 Processed Categories Distribution")
    except:
        st.info("📊 Dataset visualization images will be displayed when available")
    
    # Model architecture information
    st.markdown(
        """
        ### 🧠 Neural Network Architecture
        
        ```python
        # TechMatrix Solvers LSTM Translation Model
        model = Sequential([
            Input(shape=(20, 156)),  # 20-frame temporal window
            Masking(mask_value=0.),
            BatchNormalization(),
            Bidirectional(LSTM(32, recurrent_dropout=0.2, return_sequences=True)),
            Dropout(0.2),
            Bidirectional(LSTM(32, recurrent_dropout=0.2)),
            Dense(32, activation='elu'),
            BatchNormalization(), 
            Dropout(0.2),
            Dense(len(expression_mapping), activation='softmax')
        ])
        ```
        
        **Model Statistics:**
        - Total Parameters: 82,679 (322.96 KB)
        - Trainable Parameters: 82,239 (321.25 KB) 
        - Input Features: 156-dimensional vectors
        - Temporal Window: 20 frames
        """
    )
    
    # Technology stack
    col1, col2 = st.columns(2)
    
    with col1:
        st.markdown(
            """
            ### 🛠️ Technology Stack
            
            **Frontend & UI:**
            - Streamlit (Interactive Web App)
            - Custom CSS Styling
            - Responsive Design
            
            **Deep Learning:**
            - Keras/TensorFlow Backend
            - PyTorch Integration
            - LSTM Networks
            - OpenPose Models
            """
        )
    
    with col2:
        st.markdown(
            """
            ### 📱 Key Features
            
            **Real-time Processing:**
            - Live video analysis
            - Pose keypoint extraction
            - Temporal sequence modeling
            - Confidence scoring
            
            **User Experience:**
            - Intuitive interface
            - Visual feedback
            - Progress tracking
            - Result visualization
            """
        )
    
    # System Information
    st.markdown("### 🔧 System Information")
    col1, col2 = st.columns(2)
    
    with col1:
        st.write(f"**Python Version:** {platform.python_version()}")
        st.write(f"**FFmpeg:** {shutil.which('ffmpeg') or 'Not found'}")
        st.write(f"**FFprobe:** {shutil.which('ffprobe') or 'Not found'}")
    
    with col2:
        try:
            st.write(f"**OpenCV Version:** {cv2.__version__}")
        except:
            st.write("**OpenCV:** Not available")
        try:
            import torch
            st.write(f"**PyTorch:** {torch.__version__}")
            st.write(f"**Keras:** {keras.__version__}")
        except:
            st.write("**PyTorch/Keras:** Not available")
    
    # Team contact information
    st.markdown(
        """
        ### 📞 Contact Information
        
        **TechMatrix Solvers Team:**
        
        | Name | Role | Email | Phone |
        |------|------|-------|---------|
        | **Abhay Gupta** | Team Lead | contact2abhaygupta6187@gmail.com | 8115814535 |
        | **Kripanshu Gupta** | Backend Developer | guptakripanshu83@gmail.com | 7067058400 |
        | **Dipanshu Patel** | UI/UX Designer | dipanshupatel43@gmail.com | 9294526404 |
        | **Bhumika Patel** | Deployment & Presenter | bp7249951@gmail.com | 9302271422 |
        
        **Institution:** Shri Ram Group of Institutions
        
        ### 📚 Documentation
        
        For detailed technical documentation and implementation details, please refer to our 
        [comprehensive documentation](https://docs.google.com/document/d/1mzr2KGHRJT5heUjFF20NQ3Gb89urpjZJ/edit?usp=sharing).
        
        ---
        
        **© 2024 TechMatrix Solvers - Innovating Accessible Technology Solutions**
        """
    )

elif app_mode == 'Test Video Translation':
    # Video selection interface
    st.markdown("## 🎥 Test Video Translation")
    
    # Load test data dynamically
    with st.spinner("Loading test data..."):
        try:
            testing_df, test_files_df = load_test_data()
            st.success("✅ Test data loaded successfully!")
        except Exception as e:
            st.error(f"❌ Failed to load test data: {e}")
            st.stop()
    
    category = st.sidebar.selectbox(
        'Choose Category',
        np.sort(test_files_df['Category'].unique(), axis=-1, kind='mergesort')
    )
    
    # Filter by category
    category_mask = (test_files_df['Category'] == category)
    test_files_category = test_files_df[category_mask]
    
    class_name = st.sidebar.selectbox(
        'Choose Class',
        np.sort(test_files_category['Class'].unique(), axis=-1, kind='mergesort')
    )
    
    # Filter by class
    class_mask = (test_files_df['Class'] == class_name)
    filename = st.sidebar.selectbox(
        'Choose File',
        np.sort(test_files_category[class_mask]['Filename'].unique(), axis=-1, kind='mergesort')
    )
    
    # Display selection info
    st.info(f"📂 Selected: {category} → {class_name} → {filename}")
    
    if st.sidebar.button("🚀 Start Translation", type="primary"):
        # Filter test data for selected video
        data_mask = ((testing_df['FileName'] == filename) & 
                    (testing_df['Type'] == category) & 
                    (testing_df['Expression'] == class_name))
        
        window_size = 20
        current_test_data = testing_df[data_mask]

        if current_test_data.empty:
            st.error(f"⚠️ No matching data found for: {filename} | {category} | {class_name}")
            st.stop()
        else:
            st.success(f"✅ Loaded {current_test_data.shape[0]} frames for processing")
        
        # Create time series data
        X_test_processed, y_test_processed = create_time_series_sequences(
            current_test_data, feature_columns_processed, label_columns, window_size=window_size
        )
        X_test_processed = np.array(X_test_processed)

        # Configure Streamlit display options
        st.set_option('deprecation.showfileUploaderEncoding', False)

        st.sidebar.markdown('---')
        st.markdown(
            """
            <style>
            [data-testid="stSidebar"][aria-expanded="true"] > div:first-child {
                width: 400px;
            }
            [data-testid="stSidebar"][aria-expanded="false"] > div:first-child {
                width: 400px;
                margin-left: -400px;
            }
            </style>
            """,
            unsafe_allow_html=True,
        )

        st.sidebar.markdown('---')
        st.markdown('## 📊 Translation Results')

        # Progress tracking container
        progress_container = st.empty()

        with progress_container.container():
            progress_df = pd.DataFrame([['--', '--']], 
                                     columns=['Frames Processed', 'Detected Sign'])
            progress_table = st.table(progress_df)
            
        # Video display container
        video_display = st.empty()
        st.markdown("<hr/>", unsafe_allow_html=True)
        frame_display = st.empty()

        # Download test video
        video_file_path = hf_hub_download(
            repo_id="sunilsarolkar/isl-test-data",
            filename=f'test/{category}/{class_name}/{filename}',
            repo_type="dataset"
        )

        if not os.path.exists(video_file_path):
            st.error(f"⚠️ Video file not found: {video_file_path}")
            st.stop()

        # Initialize video capture
        video_capture = cv2.VideoCapture(video_file_path)

        # Get video metadata
        probe_result = probe_video_info(video_file_path)
        video_info = json.loads(probe_result.json)
        video_stream = [stream for stream in video_info["streams"] if stream["codec_type"] == "video"][0]
        
        input_fps = video_stream["avg_frame_rate"]
        input_pix_fmt = video_stream["pix_fmt"]
        input_vcodec = video_stream["codec_name"]
        format_name = video_info["format"]["format_name"].split(",")[0]

        # Video properties
        width = int(video_capture.get(cv2.CAP_PROP_FRAME_WIDTH))
        height = int(video_capture.get(cv2.CAP_PROP_FRAME_HEIGHT))
        fps_input = int(video_capture.get(cv2.CAP_PROP_FPS))
        
        # Processing variables
        total_frames = int(video_capture.get(cv2.CAP_PROP_FRAME_COUNT))
        frame_buffer = []
        
        # Output video configuration
        output_file = f"/tmp/techmatrix_output_{uuid.uuid4().hex}.{format_name}"
        video_writer = None
        weighted_predictions = {}
        frame_predictions = {}  # Reset for each video session
        
        frame_idx = 0
        
        try:
            # Process each frame
            for _, frame_data in current_test_data.iterrows():
                if not video_capture.isOpened():
                    st.error(f"❌ Could not open video: {video_file_path}")
                    break
                
                if video_capture.isOpened():
                    ret, frame = video_capture.read()
                    
                    if len(frame_buffer) < window_size:
                        # Initial frames - build up buffer
                        if utils is not None:
                            visualization_canvas = utils.render_stick_model(
                                frame,
                                eval(frame_data['bodypose_circles']),
                                eval(frame_data['bodypose_sticks']),
                                eval(frame_data['handpose_edges']),
                                eval(frame_data['handpose_peaks'])
                            )
                        else:
                            visualization_canvas = frame  # Use original frame if utils not available
                        
                        # Add prediction plots
                        if utils is not None:
                            canvas_with_predictions = utils.create_bar_plot_visualization(
                                visualization_canvas, {}, 
                                f'Building Buffer - Frame {frame_idx + 1} [No Predictions Yet]',
                                visualization_canvas
                            )
                            canvas_with_predictions = utils.create_bar_plot_visualization(
                                canvas_with_predictions, weighted_predictions,
                                f'Weighted Average - Frame {frame_idx + 1} [No Predictions Yet]',
                                visualization_canvas
                            )
                            canvas_with_predictions = utils.add_bottom_padding(
                                canvas_with_predictions, (255, 255, 255), 100
                            )
                        else:
                            canvas_with_predictions = visualization_canvas  # Use base canvas if utils not available
                        
                        # Initialize video writer
                        if video_writer is None:
                            input_framesize = canvas_with_predictions.shape[:2]
                            video_writer = VideoWriter(output_file, input_fps, input_framesize, 
                                                     input_pix_fmt, input_vcodec)

                        video_writer.write_frame(canvas_with_predictions)
                        
                        # Update progress display
                        with progress_container.container():
                            progress_df = pd.DataFrame(
                                [[f'{frame_idx + 1}/{current_test_data.shape[0]}', 
                                  '<Building 20-frame buffer>']],
                                columns=['Frames Processed', 'Detected Sign']
                            )
                            progress_table = st.table(progress_df)
                            
                        frame_buffer.append(frame)
                        
                        # Display current frame
                        with video_display.container():
                            st.image(canvas_with_predictions, channels='BGR', use_column_width=True)
                    else:
                        # Process with full buffer - make predictions
                        frame_buffer[:-1] = frame_buffer[1:]
                        frame_buffer[-1] = frame
                        
                        # Load translation model
                        translation_model = load_translation_model()
                        
                        # Check if model loaded successfully
                        sequence_idx = frame_idx - 20  # Define sequence_idx for both cases
                        if translation_model is None:
                            st.error("❌ Translation model failed to load. Cannot make predictions.")
                            # Use dummy predictions to keep the visualization working
                            current_predictions = {"model_not_available": 0.0}
                            top_3_signs = ["model_not_available"]
                            top_3_probabilities = [0.0]
                        else:
                            # Make prediction on current window
                            prediction_output = translation_model(
                                X_test_processed[sequence_idx].reshape(
                                    1, X_test_processed[sequence_idx].shape[0], 
                                    X_test_processed[sequence_idx].shape[1]
                                )
                            )
                            
                            # Handle both PyTorch and Keras/TensorFlow models
                            try:
                                # Try PyTorch tensor operations first
                                prediction_output = prediction_output[0].cpu().detach().numpy()
                            except AttributeError:
                                # If it's a Keras model, it already returns NumPy arrays
                                prediction_output = prediction_output[0]
                            
                            # Get top predictions
                            top_prediction_idx = np.argmax(prediction_output)
                            top_3_indices = prediction_output.argsort()[-3:][::-1]
                            top_3_signs = [get_sign_label(i) for i in top_3_indices]
                            top_3_probabilities = prediction_output[top_3_indices]
                            
                            # Current frame predictions
                            current_predictions = {}
                            for sign, prob in zip(top_3_signs, top_3_probabilities):
                                current_predictions[sign] = prob
                        
                        # Update frame-wise predictions for weighted average
                        for sign, prob in zip(top_3_signs, top_3_probabilities):
                            if sign not in frame_predictions:
                                frame_predictions[sign] = []
                            frame_predictions[sign].append(prob)

                        # Calculate weighted averages
                        for sign in frame_predictions:
                            sign_predictions = frame_predictions[sign]
                            sign_weights = [len(sign_predictions) for _ in range(len(sign_predictions))]
                            weighted_predictions[sign] = calculate_weighted_average(
                                sign_predictions, sign_weights
                            )

                        # Sort predictions by confidence
                        sorted_predictions = dict(
                            sorted(weighted_predictions.items(), key=lambda item: item[1], reverse=True)
                        )
                        
                        # Create visualization
                        if utils is not None:
                            visualization_canvas = utils.render_stick_model(
                                frame,
                                eval(frame_data['bodypose_circles']),
                                eval(frame_data['bodypose_sticks']),
                                eval(frame_data['handpose_edges']),
                                eval(frame_data['handpose_peaks'])
                            )
                        else:
                            visualization_canvas = frame  # Use original frame if utils not available
                        
                        # Add prediction visualizations
                        if utils is not None:
                            canvas_with_predictions = utils.create_bar_plot_visualization(
                                visualization_canvas, current_predictions,
                                f'Current Window Prediction (Frames {sequence_idx + 1}-{frame_idx + 1})',
                                visualization_canvas
                            )
                            canvas_with_predictions = utils.create_bar_plot_visualization(
                                canvas_with_predictions, weighted_predictions,
                                f'Cumulative Weighted Average - Frame {frame_idx + 1}',
                                visualization_canvas
                            )
                            canvas_with_predictions = utils.add_bottom_padding(
                                canvas_with_predictions, (255, 255, 255), 100
                            )
                        else:
                            canvas_with_predictions = visualization_canvas  # Use base canvas if utils not available
                        
                        video_writer.write_frame(canvas_with_predictions)
                        
                        # Get best prediction for display
                        if weighted_predictions:
                            best_sign = max(weighted_predictions, key=weighted_predictions.get)
                            best_confidence = weighted_predictions[best_sign]
                        else:
                            best_sign = "no_predictions"
                            best_confidence = 0.0
                        
                        # Update progress display
                        with progress_container.container():
                            progress_df = pd.DataFrame(
                                [[f'{frame_idx + 1}/{current_test_data.shape[0]}',
                                  f'{best_sign} ({best_confidence * 100:.2f}%)']],
                                columns=['Frames Processed', 'Detected Sign']
                            )
                            progress_table = st.table(progress_df)
                            
                        # Display current frame
                        with video_display.container():
                            st.image(canvas_with_predictions, channels='BGR', use_column_width=True)

                    frame_idx += 1

            # Finalize video processing
            st.success("✅ Video processing completed!")
            
            with video_display.container():
                if video_writer is not None:
                    video_writer.close()
                    with open(output_file, 'rb') as video_file:
                        output_video_bytes = video_file.read()
                    st.video(output_video_bytes)
                    st.info(f"💾 Processed video saved: {output_file}")
                else:
                    st.warning("⚠️ No video output generated")

        finally:
            # Clean up resources
            if 'video_capture' in locals() and video_capture is not None:
                video_capture.release()
            if 'video_writer' in locals() and video_writer is not None:
                video_writer.close()
            # Note: cv2.destroyAllWindows() removed for headless compatibility

# Footer
st.markdown(
    """
    ---
    <div style="text-align: center; color: #666;">
    <p><strong>TechMatrix Solvers</strong> | Shri Ram Group of Institutions</p>
    <p>Innovating Accessible Technology Solutions for Everyone 🚀</p>
    </div>
    """, 
    unsafe_allow_html=True
)