app.py

import gradio as gr
import torch
import os
import tempfile
import shutil
import time
import ffmpeg
import numpy as np
from PIL import Image
import moviepy.editor as mp
from infer import lotus, load_models
import logging

# Set up logging
logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)

# Set device to use the L40s GPU explicitly
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")

# Load models once
task_name = 'depth'
pipe_g, pipe_d = load_models(task_name, device)

# Preprocess the video to adjust resolution and frame rate
def preprocess_video(video_path, target_fps=24, max_resolution=(512, 512)):
    """Preprocess the video to resize and adjust its frame rate."""
    video = mp.VideoFileClip(video_path)

    # Resize video if it's larger than the target resolution
    if video.size[0] > max_resolution[0] or video.size[1] > max_resolution[1]:
        video = video.resize(height=max_resolution[1])

    # Adjust FPS if target_fps is specified
    if target_fps > 0:
        video = video.set_fps(target_fps)

    return video

# Process a batch of frames through the depth model
def process_frames_batch(frames_batch, seed=0, target_size=(512, 512)):
    """Process a batch of frames and return depth maps."""
    try:
        torch.cuda.empty_cache()  # Clear GPU cache

        # Resize frames to the target size
        images_batch = [Image.fromarray(frame).convert('RGB').resize(target_size, Image.BILINEAR) for frame in frames_batch]

        # Run batch inference
        depth_maps = lotus(images_batch, 'depth', seed, device, pipe_g, pipe_d)

        return depth_maps

    except Exception as e:
        logger.error(f"Error processing batch: {e}")
        return [None] * len(frames_batch)

# Process video frames and generate depth maps
def process_video(video_path, fps=0, seed=0, batch_size=4):
    """Process video frames in batches and generate depth maps."""
    # Create a persistent temporary directory
    temp_dir = tempfile.mkdtemp()
    try:
        start_time = time.time()

        # Preprocess the video
        video = preprocess_video(video_path, target_fps=fps, max_resolution=(512, 512))

        # Use original video FPS if not specified
        if fps == 0:
            fps = video.fps

        frames = list(video.iter_frames(fps=video.fps))
        total_frames = len(frames)

        logger.info(f"Processing {total_frames} frames at {fps} FPS...")

        # Create directory for frame sequence and outputs
        frames_dir = os.path.join(temp_dir, "frames")
        os.makedirs(frames_dir, exist_ok=True)

        processed_frames = []

        # Process frames in batches
        for i in range(0, total_frames, batch_size):
            current_batch_size = batch_size
            while current_batch_size > 0:
                try:
                    frames_batch = frames[i:i+current_batch_size]
                    depth_maps = process_frames_batch(frames_batch, seed)
                    break
                except RuntimeError as e:
                    if 'out of memory' in str(e):
                        current_batch_size = current_batch_size // 2
                        logger.warning(f"Reducing batch size to {current_batch_size} due to out of memory error.")
                        torch.cuda.empty_cache()
                    else:
                        raise e

            for j, depth_map in enumerate(depth_maps):
                frame_index = i + j
                if depth_map is not None:
                    # Save frame
                    frame_path = os.path.join(frames_dir, f"frame_{frame_index:06d}.png")
                    depth_map.save(frame_path)

                    # Update live preview every 10% progress
                    if frame_index % max(1, total_frames // 10) == 0:
                        elapsed_time = time.time() - start_time
                        progress = (frame_index / total_frames) * 100
                        yield depth_map, None, None, f"Processed {frame_index}/{total_frames} frames... ({progress:.2f}%) Elapsed: {elapsed_time:.2f}s"
                else:
                    logger.error(f"Error processing frame {frame_index}")

        logger.info("Creating output files...")

        # Create ZIP of frame sequence
        zip_filename = f"depth_frames_{int(time.time())}.zip"
        zip_path = os.path.join(temp_dir, zip_filename)
        shutil.make_archive(zip_path[:-4], 'zip', frames_dir)

        # Create MP4 video
        video_filename = f"depth_video_{int(time.time())}.mp4"
        output_video_path = os.path.join(temp_dir, video_filename)

        try:
            # FFmpeg settings for high-quality MP4
            (
                ffmpeg
                .input(os.path.join(frames_dir, 'frame_%06d.png'), pattern_type='sequence', framerate=fps)
                .output(output_video_path, vcodec='libx264', pix_fmt='yuv420p', crf=17)
                .run(overwrite_output=True, quiet=True)
            )
            logger.info("MP4 video created successfully!")

        except ffmpeg.Error as e:
            logger.error(f"Error creating video: {e.stderr.decode() if e.stderr else str(e)}")
            output_video_path = None

        total_time = time.time() - start_time
        logger.info("Processing complete!")

        # Yield the file paths
        yield None, zip_path, output_video_path, f"Processing complete! Total time: {total_time:.2f} seconds"

    except Exception as e:
        logger.error(f"Error: {e}")
        yield None, None, None, f"Error processing video: {e}"

# Wrapper function with error handling
def process_wrapper(video, fps=0, seed=0, batch_size=4):
    if video is None:
        raise gr.Error("Please upload a video.")
    try:
        outputs = []
        # Use video directly, since it's the file path
        for output in process_video(video, fps, seed, batch_size):
            outputs.append(output)
            yield output
        return outputs[-1]
    except Exception as e:
        raise gr.Error(f"Error processing video: {str(e)}")

# Custom CSS for styling (unchanged)
custom_css = """
    .title-container {
        text-align: center;
        padding: 10px 0;
    }

    #title {
        font-family: -apple-system, BlinkMacSystemFont, "Segoe UI", Roboto, Helvetica, Arial, sans-serif;
        font-size: 36px;
        font-weight: bold;
        color: #000000;
        padding: 10px;
        border-radius: 10px;
        display: inline-block;
        background: linear-gradient(
            135deg,
            #e0f7fa, #e8f5e9, #fff9c4, #ffebee,
            #f3e5f5, #e1f5fe, #fff3e0, #e8eaf6
        );
        background-size: 400% 400%;
        animation: gradient-animation 15s ease infinite;
    }

    @keyframes gradient-animation {
        0% { background-position: 0% 50%; }
        50% { background-position: 100% 50%; }
        100% { background-position: 0% 50%; }
    }
"""

# Gradio Interface
with gr.Blocks(css=custom_css) as demo:
    gr.HTML('''
        <div class="title-container">
            <div id="title">Video Depth Estimation</div>
        </div>
    ''')

    with gr.Row():
        with gr.Column():
            video_input = gr.Video(label="Upload Video", interactive=True)
            fps_slider = gr.Slider(minimum=0, maximum=60, step=1, value=0, label="Output FPS (0 for original)")
            seed_slider = gr.Number(value=0, label="Seed")
            batch_size_slider = gr.Slider(minimum=1, maximum=16, step=1, value=4, label="Batch Size")
            btn = gr.Button("Process Video")

        with gr.Column():
            preview_image = gr.Image(label="Live Preview")
            output_frames_zip = gr.File(label="Download Frame Sequence (ZIP)")
            output_video = gr.File(label="Download Video (MP4)")
            time_textbox = gr.Textbox(label="Status", interactive=False)

    btn.click(
        fn=process_wrapper,
        inputs=[video_input, fps_slider, seed_slider, batch_size_slider],
        outputs=[preview_image, output_frames_zip, output_video, time_textbox]
    )

    demo.queue()

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