app.py

import os
import sys
import subprocess
import pkg_resources

def is_package_installed(package_name):
    try:
        pkg_resources.get_distribution(package_name)
        return True
    except pkg_resources.DistributionNotFound:
        return False

if is_package_installed("mmcv"):
    print("MMCV is installed.")
else:
    print("MMCV is not installed. Build it from the source.")
    os.environ["MMCV_WITH_OPS"] = "1"
    os.environ["FORCE_MLU"] = "1"
    subprocess.run(["pip", "install", "-e", "./mmcv"], check=True)
    subprocess.run(["pip", "list"], check=True)

if is_package_installed("pytorch3d"):
    print("pytorch3d is installed.")
else:
    print("pytorch3d is not installed. Build it from the source.")
    subprocess.run(["pip", "install", "-e", "./pytorch3d"], check=True)

if is_package_installed("MultiScaleDeformableAttention"):
    print("MultiScaleDeformableAttention is installed.")
else:
    print("MultiScaleDeformableAttention is not installed. Build it from the source.")
    subprocess.run(["pip", "install", "-e", "./models/aios/ops"], check=True)
    
import os.path as osp
from pathlib import Path
import cv2
import gradio as gr
import torch
import math
import spaces
from huggingface_hub import hf_hub_download

hf_hub_download(repo_id="ttxskk/AiOS", filename="aios_checkpoint.pth", local_dir="/home/user/app/pretrained_models")

OUT_FOLDER = '/home/user/app/demo_out'
os.makedirs(OUT_FOLDER, exist_ok=True)

DEMO_CONFIG = '/home/user/app/config/aios_smplx_demo.py'
MODEL_PATH = '/home/user/app/pretrained_models/aios_checkpoint.pth'
@spaces.GPU(enable_queue=True, duration=300)
def infer(video_input, batch_size, threshold=0.3, num_person=1):
    os.system(f'rm -rf {OUT_FOLDER}/*')
    os.system(f'torchrun --nproc_per_node 1 \
             main.py \
            -c {DEMO_CONFIG} \
            --options batch_size={batch_size} backbone="resnet50" num_person={num_person} threshold={threshold} \
            --resume {MODEL_PATH} \
            --eval \
            --inference \
            --inference_input {video_input} \
            --to_vid \
            --output_dir {OUT_FOLDER}')
    
    video_path = os.path.join(OUT_FOLDER, 'demo_vid.mp4')
    save_path_img = os.path.join(OUT_FOLDER, 'res_img')
    save_path_mesh = os.path.join(OUT_FOLDER, 'mesh')
    save_mesh_file = os.path.join(OUT_FOLDER, 'mesh.zip')
    os.system(f'zip -r {save_mesh_file} {save_path_mesh}')
    yield video_path, save_mesh_file

TITLE = """

<div style="display: flex; justify-content: center; align-items: center; text-align: center;">
  <div>
    <h1 align="center">AiOS: All-in-One-Stage Expressive Human Pose and Shape Estimation</h1>
  </div>
</div>

<div style="display: flex; justify-content: center; align-items: center; text-align: center;">
    <div style="display:flex; gap: 0.25rem;" align="center">
        <a href="https://ttxskk.github.io/AiOS/" target="_blank"><img src='https://img.shields.io/badge/Project-Page-Green'></a>
        <a href="https://github.com/ttxskk/AiOS" target="_blank"><img src='https://img.shields.io/badge/Github-Code-blue'></a>
        <a href="https://ttxskk.github.io/AiOS/assets/aios_cvpr24.pdf" target="_blank"><img src='https://img.shields.io/badge/Paper-Arxiv-red'></a>
    </div>
</div>
<div style="font-size: 1.1rem; color: #555; max-width: 800px; margin: 1rem auto; line-height: 1.5; justify-content: center; align-items: center; text-align: center;">
  <div>
    <p>Recover multiple expressive human pose and shape from an RGB image without any additional requirements, such as an off-the-shelf detection model.</h1>
  </div>
</div>
"""
VIDEO = '''
<center>
    <iframe width="960" height="540" 
    src="https://www.youtube.com/embed/yzCL7TYpzvc?si=EoxWNE6VPBxsy7Go" 
    title="AiOS: All-in-One-Stage Expressive Human Pose and Shape Estimation" 
    frameborder="0" 
    allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture; web-share" 
    allowfullscreen>
    </iframe>
</center><br/>
'''

DESCRIPTION = '''
<p>
Note: Score threshold defines the minimum confidence level for person detection. The default value is 0.3. 
If the confidence score of a detected person falls below this score threshold, the detection will be discarded.
</p>
'''
with gr.Blocks(title="AiOS", theme=gr.themes.Soft(primary_hue="blue", secondary_hue="gray")) as demo:
    
    gr.Markdown(TITLE)
    gr.HTML(VIDEO)
    gr.Markdown(DESCRIPTION)
    with gr.Row():
        with gr.Column(scale=2):
            video_input = gr.Video(label="Input video", elem_classes="video")
        with gr.Column(scale=1):
            batch_size = gr.Textbox(label="Batch Size", type="text", value=8)
            num_person = gr.Textbox(label="Number of Person", type="text", value=1)
            threshold = gr.Slider(0, 1.0, value=0.3, label='Score Threshold')
            send_button = gr.Button("Infer")
    gr.HTML("""<br/>""")
        
    with gr.Row():
        with gr.Column():
            # processed_frames = gr.Image(label="Last processed frame")
            video_output = gr.Video(elem_classes="video")
        with gr.Column():
            meshes_output = gr.File(label="3D meshes")
    
    send_button.click(fn=infer, inputs=[video_input, batch_size, threshold, num_person], outputs=[video_output, meshes_output])
    # example_videos = gr.Examples([
    #     ['./assets/01.mp4'], 
    #     ['./assets/02.mp4'], 
    #     ['./assets/03.mp4'],
    #     ['./assets/04.mp4'], 
    #     ['./assets/05.mp4'], 
    #     ['./assets/06.mp4'], 
    #     ['./assets/07.mp4'], 
    #     ['./assets/08.mp4'], 
    #     ['./assets/09.mp4'], 
    #     ], 
    #     inputs=[video_input, 0.5])
    
demo.queue().launch(debug=True)