app.py

import gradio as gr
import pixeltable as pxt
from pixeltable.functions.huggingface import clip_image, clip_text
from pixeltable.iterators import FrameIterator
import PIL.Image
import os

# Process video and create index
def process_video(video_file, progress=gr.Progress()):
    progress(0, desc="Initializing...")

    # Pixeltable setup
    pxt.drop_dir('video_search', force=True)
    pxt.create_dir('video_search')

    # Update type declaration to use simpler syntax
    video_table = pxt.create_table('video_search.videos', {'video': pxt.Video})

    frames_view = pxt.create_view(
        'video_search.frames', 
        video_table, 
        iterator=FrameIterator.create(video=video_table.video, fps=1)
    )

    progress(0.2, desc="Inserting video...")
    video_table.insert([{'video': video_file.name}])
    
    progress(0.4, desc="Creating embedding index...")
    # Updated embedding pattern using .using()
    frames_view.add_embedding_index(
        'frame',
        string_embed=clip_text.using(model_id='openai/clip-vit-base-patch32'),
        image_embed=clip_image.using(model_id='openai/clip-vit-base-patch32')
    )

    progress(1.0, desc="Processing complete")
    return "Good news! Your video has been processed. Easily find the moments you need by searching with text or images."

# Perform similarity search
def similarity_search(query, search_type, num_results, progress=gr.Progress()):
    frames_view = pxt.get_table('video_search.frames')
    
    progress(0.5, desc="Performing search...")
    if search_type == "Text":
        sim = frames_view.frame.similarity(query)
    else:  # Image search
        sim = frames_view.frame.similarity(query)
    
    results = frames_view.order_by(sim, asc=False).limit(num_results).select(frames_view.frame, sim=sim).collect()
    
    progress(1.0, desc="Search complete")
    return [row['frame'] for row in results]
    
# Gradio interface
with gr.Blocks(theme=gr.themes.Base()) as demo:
    gr.Markdown(
        """
        <div style= margin-bottom: 20px;">
            <img src="https://raw.githubusercontent.com/pixeltable/pixeltable/main/docs/resources/pixeltable-logo-large.png" alt="Pixeltable" style="max-width: 150px;" />
            <h2>Text and Image similarity search on video frames with embedding indexes</h2>
        </div>
        """
    )
    gr.HTML(
    """
    <p>
        <a href="https://github.com/pixeltable/pixeltable" target="_blank" style="color: #F25022; text-decoration: none; font-weight: bold;">Pixeltable</a> is a declarative interface for working with text, images, embeddings, and even video, enabling you to store, transform, index, and iterate on data.
    </p>
    """
    )

    
    with gr.Row():
        with gr.Column(scale=1):

            gr.Markdown(
            """
            <h3>1. Insert video</h3>
            """)
            
            video_file = gr.File(label="Upload Video")
            process_button = gr.Button("Process Video")
            process_output = gr.Textbox(label="Status", lines=2)
            
            gr.Markdown(
            """
            <h3>2. Search video frames</h3>
            """)
            
            search_type = gr.Radio(["Text", "Image"], label="Search Type", value="Text")
            text_input = gr.Textbox(label="Text Query")
            image_input = gr.Image(label="Image Query", type="pil", visible=False)
            num_results = gr.Slider(minimum=1, maximum=20, value=5, step=1, label="Number of Results")
            search_button = gr.Button("Search")
        
        with gr.Column(scale=2):

            gr.Markdown(
            """
            <h3>3. Visualize results</h3>
            """)
            
            results_gallery = gr.Gallery(label="Search Results", columns=3)
       
            gr.Examples(
            examples=[
                ["bangkok.mp4"],
                ["lotr.mp4"],
                ["mi.mp4"],
            ],
            label="Click one of the examples below to get started",
            inputs=[video_file],
            fn=process_video
            )
    
    def update_search_input(choice):
        return gr.update(visible=choice=="Text"), gr.update(visible=choice=="Image")

    search_type.change(update_search_input, search_type, [text_input, image_input])
    
    process_button.click(
        process_video,
        inputs=[video_file],
        outputs=[process_output]
    )
    
    def perform_search(search_type, text_query, image_query, num_results):
        query = text_query if search_type == "Text" else image_query
        return similarity_search(query, search_type, num_results)

    search_button.click(
        perform_search,
        inputs=[search_type, text_input, image_input, num_results],
        outputs=[results_gallery]
    )

if __name__ == "__main__":
    demo.launch()