app.py

import os
import sys
import os.path as osp
from pathlib import Path
import cv2
import gradio as gr
import torch
import math
import spaces

try:
    import mmpose
except:
    os.system('pip install /home/user/app/main/transformer_utils')

os.system('cp -rf /home/user/app/assets/conversions.py /home/user/.pyenv/versions/3.9.18/lib/python3.9/site-packages/torchgeometry/core/conversions.py')
DEFAULT_MODEL='postometro'
OUT_FOLDER = '/home/user/app/demo_out'
os.makedirs(OUT_FOLDER, exist_ok=True)

@spaces.GPU(enable_queue=True)
def infer(image_input, in_threshold=0.5, num_people="Single person", render_mesh=False):
    num_gpus = 1 if torch.cuda.is_available() else -1
    print("!!! torch.cuda.is_available: ", torch.cuda.is_available())      
    print("!!! torch.cuda.device_count: ", torch.cuda.device_count())   
    print("CUDA version: ", torch.version.cuda)  
    index = torch.cuda.current_device()
    print("CUDA current_device: ", index)  
    print("CUDA device_name: ", torch.cuda.get_device_name(index))
    from main.inference import Inferer
    inferer = Inferer(DEFAULT_MODEL, num_gpus, OUT_FOLDER)
    os.system(f'rm -rf {OUT_FOLDER}/*')
    multi_person = False if (num_people == "Single person") else True
    vis_img, bbox = inferer.infer(image_input, in_threshold, 0, multi_person, not(render_mesh))

    # cap = cv2.VideoCapture(video_input)
    # fps = math.ceil(cap.get(5))
    # width = int(cap.get(3))
    # height = int(cap.get(4))
    # fourcc = cv2.VideoWriter_fourcc(*'mp4v')
    # video_path = osp.join(OUT_FOLDER, f'out.m4v')
    # final_video_path = osp.join(OUT_FOLDER, f'out.mp4')
    # video_output = cv2.VideoWriter(video_path, fourcc, fps, (width, height))
    # success = 1
    # frame = 0
    # while success:
    #     success, original_img = cap.read()
    #     if not success:
    #         break
    #     frame += 1
    #     img, mesh_paths, smplx_paths = inferer.infer(original_img, in_threshold, frame, multi_person, not(render_mesh))
    #     video_output.write(img)
    #     yield img, None, None, None
    # cap.release()
    # video_output.release()
    # cv2.destroyAllWindows()
    # os.system(f'ffmpeg -i {video_path} -c copy {final_video_path}')

    # #Compress mesh and smplx files
    # save_path_mesh = os.path.join(OUT_FOLDER, 'mesh')
    # save_mesh_file = os.path.join(OUT_FOLDER, 'mesh.zip')
    # os.makedirs(save_path_mesh, exist_ok= True)
    # save_path_smplx = os.path.join(OUT_FOLDER, 'smplx')
    # save_smplx_file = os.path.join(OUT_FOLDER, 'smplx.zip')
    # os.makedirs(save_path_smplx, exist_ok= True)
    # os.system(f'zip -r {save_mesh_file} {save_path_mesh}')
    # os.system(f'zip -r {save_smplx_file} {save_path_smplx}')
    # yield img, video_path, save_mesh_file, save_smplx_file
    return vis_img, "bbox meta: {}".format(bbox)

TITLE = '''<h1 align="center">PostoMETRO: Pose Token Enhanced Mesh Transformer for Robust 3D Human Mesh Recovery</h1>'''
DESCRIPTION = '''
<b>Official Gradio demo</b> for <b>PostoMETRO: Pose Token Enhanced Mesh Transformer for Robust 3D Human Mesh Recovery</b>.<br>
<p>
Note: You can drop a image at the panel (or select one of the examples) 
to obtain the 3D parametric reconstructions of the detected humans.
</p>
'''

with gr.Blocks(title="PostoMETRO", css=".gradio-container") as demo:

    gr.Markdown(TITLE)
    gr.Markdown(DESCRIPTION)

    with gr.Row():
        with gr.Column():
            image_input = gr.Image(label="Input image", elem_classes="Image")
            threshold = gr.Slider(0, 1.0, value=0.5, label='BBox detection threshold')
            num_people = gr.Radio(
                choices=["Single person", "Multiple people"],
                value="Single person",
                label="Number of people",
                info="Choose how many people are there in the video. Choose 'single person' for faster inference.",
                interactive=True,
                scale=1,)
            mesh_as_vertices = gr.Checkbox(
                label="Render as mesh", 
                info="By default, the estimated SMPL-X parameters are rendered as vertices for faster visualization. Check this option if you want to visualize meshes instead.",
                interactive=True,
                scale=1,)
            send_button = gr.Button("Infer")
        with gr.Column():
            processed_frames = gr.Image(label="Rendered Results")
            debug_textbox = gr.Textbox(label="Debug information")

    # example_images = gr.Examples([])
    send_button.click(fn=infer, inputs=[image_input, threshold, num_people, mesh_as_vertices], outputs=[processed_frames, debug_textbox])
    # with gr.Row():
    example_images = gr.Examples([
        ['/home/user/app/assets/01.jpg'], 
        ['/home/user/app/assets/02.jpg'], 
        ['/home/user/app/assets/03.jpg'], 
        ], 
        inputs=[image_input, 0.5])

#demo.queue()
demo.queue().launch(debug=True)