app.py

import sys
import os
import OpenGL.GL as gl
os.environ["PYOPENGL_PLATFORM"] = "egl"
sys.argv = ['VQ-Trans/GPT_eval_multi.py']
os.umask(0)
os.makedirs('output', exist_ok=True)
os.chdir('VQ-Trans')
os.umask(0)
os.makedirs('checkpoints', exist_ok=True)
os.system('gdown --fuzzy https://drive.google.com/file/d/1chKRmW2vNX5Lz6dPO45WY1qVNUkoJJ5D/view -O checkpoints/')
os.system('gdown --fuzzy https://drive.google.com/file/d/1cgUVeGwPU7PzTDx20CAijt18fNhtnOPq/view -O checkpoints/')
# os.system('gdown --fuzzy https://drive.google.com/file/d/1o7RTDQcToJjTm9_mNWTyzvZvjTWpZfug/view -O checkpoints/')
# os.system('gdown --fuzzy https://drive.google.com/file/d/1tX79xk0fflp07EZ660Xz1RAFE33iEyJR/view -O checkpoints/')
os.system('unzip checkpoints/t2m.zip')
os.system('unzip checkpoints/kit.zip')
os.system('mv kit checkpoints')
os.system('mv t2m checkpoints')
os.system('rm checkpoints/t2m.zip')
os.system('rm checkpoints/kit.zip')
sys.path.append('/home/user/app/VQ-Trans')

import options.option_transformer as option_trans
from huggingface_hub import snapshot_download
model_path = snapshot_download(repo_id="vumichien/T2M-GPT")

args = option_trans.get_args_parser()

args.dataname = 't2m'
args.resume_pth = f'{model_path}/VQVAE/net_last.pth'
args.resume_trans = f'{model_path}/VQTransformer_corruption05/net_best_fid.pth'
args.down_t = 2
args.depth = 3
args.block_size = 51

import clip
import torch
import numpy as np
import models.vqvae as vqvae
import models.t2m_trans as trans
from utils.motion_process import recover_from_ric
import visualization.plot_3d_global as plot_3d
from models.rotation2xyz import Rotation2xyz
import numpy as np
from trimesh import Trimesh
import gc

import torch
from visualize.simplify_loc2rot import joints2smpl
import pyrender
# import matplotlib.pyplot as plt

import io
import imageio
from shapely import geometry
import trimesh
from pyrender.constants import RenderFlags
import math
# import ffmpeg
# from PIL import Image
import hashlib
import gradio as gr

## load clip model and datasets
is_cuda = torch.cuda.is_available()
device = torch.device("cuda" if is_cuda else "cpu")
print(device)
clip_model, clip_preprocess = clip.load("ViT-B/32", device=device, jit=False, download_root='./')  # Must set jit=False for training
clip.model.convert_weights(clip_model)  # Actually this line is unnecessary since clip by default already on float16
clip_model.eval()
for p in clip_model.parameters():
    p.requires_grad = False

net = vqvae.HumanVQVAE(args, ## use args to define different parameters in different quantizers
                       args.nb_code,
                       args.code_dim,
                       args.output_emb_width,
                       args.down_t,
                       args.stride_t,
                       args.width,
                       args.depth,
                       args.dilation_growth_rate)


trans_encoder = trans.Text2Motion_Transformer(num_vq=args.nb_code, 
                                embed_dim=1024, 
                                clip_dim=args.clip_dim, 
                                block_size=args.block_size, 
                                num_layers=9, 
                                n_head=16, 
                                drop_out_rate=args.drop_out_rate, 
                                fc_rate=args.ff_rate)


print('loading checkpoint from {}'.format(args.resume_pth))
ckpt = torch.load(args.resume_pth, map_location='cpu')
net.load_state_dict(ckpt['net'], strict=True)
net.eval()
    
print('loading transformer checkpoint from {}'.format(args.resume_trans))
ckpt = torch.load(args.resume_trans, map_location='cpu')
trans_encoder.load_state_dict(ckpt['trans'], strict=True)
trans_encoder.eval()
mean = torch.from_numpy(np.load('./checkpoints/t2m/VQVAEV3_CB1024_CMT_H1024_NRES3/meta/mean.npy'))
std = torch.from_numpy(np.load('./checkpoints/t2m/VQVAEV3_CB1024_CMT_H1024_NRES3/meta/std.npy'))
if is_cuda:
    net.cuda()
    trans_encoder.cuda()
    mean = mean.cuda()
    std = std.cuda()


def render(motions, device_id=0, name='test_vis'):
    frames, njoints, nfeats = motions.shape
    MINS = motions.min(axis=0).min(axis=0)
    MAXS = motions.max(axis=0).max(axis=0)

    height_offset = MINS[1]
    motions[:, :, 1] -= height_offset
    trajec = motions[:, 0, [0, 2]]
    is_cuda = torch.cuda.is_available()
    # device = torch.device("cuda" if is_cuda else "cpu")
    j2s = joints2smpl(num_frames=frames, device_id=0, cuda=is_cuda)
    rot2xyz = Rotation2xyz(device=device)
    faces = rot2xyz.smpl_model.faces

    if not os.path.exists(f'output/{name}_pred.pt'): 
        print(f'Running SMPLify, it may take a few minutes.')
        motion_tensor, opt_dict = j2s.joint2smpl(motions)  # [nframes, njoints, 3]

        vertices = rot2xyz(torch.tensor(motion_tensor).clone(), mask=None,
                                        pose_rep='rot6d', translation=True, glob=True,
                                        jointstype='vertices',
                                        vertstrans=True)
        vertices = vertices.detach().cpu()
        torch.save(vertices, f'output/{name}_pred.pt')
    else:
        vertices = torch.load(f'output/{name}_pred.pt')
    frames = vertices.shape[3] # shape: 1, nb_frames, 3, nb_joints
    print(vertices.shape)
    MINS = torch.min(torch.min(vertices[0], axis=0)[0], axis=1)[0]
    MAXS = torch.max(torch.max(vertices[0], axis=0)[0], axis=1)[0]

    out_list = []
    
    minx = MINS[0] - 0.5
    maxx = MAXS[0] + 0.5
    minz = MINS[2] - 0.5 
    maxz = MAXS[2] + 0.5
    polygon = geometry.Polygon([[minx, minz], [minx, maxz], [maxx, maxz], [maxx, minz]])
    polygon_mesh = trimesh.creation.extrude_polygon(polygon, 1e-5)

    vid = []
    for i in range(frames):
        if i % 10 == 0:
            print(i)

        mesh = Trimesh(vertices=vertices[0, :, :, i].squeeze().tolist(), faces=faces)

        base_color = (0.11, 0.53, 0.8, 0.5)
        ## OPAQUE rendering without alpha
        ## BLEND rendering consider alpha 
        material = pyrender.MetallicRoughnessMaterial(
            metallicFactor=0.7,
            alphaMode='OPAQUE',
            baseColorFactor=base_color
        )


        mesh = pyrender.Mesh.from_trimesh(mesh, material=material)

        polygon_mesh.visual.face_colors = [0, 0, 0, 0.21]
        polygon_render = pyrender.Mesh.from_trimesh(polygon_mesh, smooth=False)

        bg_color = [1, 1, 1, 0.8]
        scene = pyrender.Scene(bg_color=bg_color, ambient_light=(0.4, 0.4, 0.4))
        
        sx, sy, tx, ty = [0.75, 0.75, 0, 0.10]

        camera = pyrender.PerspectiveCamera(yfov=(np.pi / 3.0))

        light = pyrender.DirectionalLight(color=[1,1,1], intensity=300)

        scene.add(mesh)

        c = np.pi / 2

        scene.add(polygon_render, pose=np.array([[ 1, 0, 0, 0],

        [ 0, np.cos(c), -np.sin(c), MINS[1].cpu().numpy()],

        [ 0, np.sin(c), np.cos(c), 0],

        [ 0, 0, 0, 1]]))

        light_pose = np.eye(4)
        light_pose[:3, 3] = [0, -1, 1]
        scene.add(light, pose=light_pose.copy())

        light_pose[:3, 3] = [0, 1, 1]
        scene.add(light, pose=light_pose.copy())

        light_pose[:3, 3] = [1, 1, 2]
        scene.add(light, pose=light_pose.copy())


        c = -np.pi / 6

        scene.add(camera, pose=[[ 1, 0, 0, (minx+maxx).cpu().numpy()/2],

                                [ 0, np.cos(c), -np.sin(c), 1.5],

                                [ 0, np.sin(c), np.cos(c), max(4, minz.cpu().numpy()+(1.5-MINS[1].cpu().numpy())*2, (maxx-minx).cpu().numpy())],

                                [ 0, 0, 0, 1]
                                ])
        
        # render scene
        r = pyrender.OffscreenRenderer(960, 960)

        color, _ = r.render(scene, flags=RenderFlags.RGBA)
        # Image.fromarray(color).save(outdir+'/'+name+'_'+str(i)+'.png')

        vid.append(color)

        r.delete()

    out = np.stack(vid, axis=0)
    imageio.mimwrite(f'output/results.gif', out, fps=20)
    del out, vertices
    return f'output/results.gif'

def predict(clip_text, method='fast'):
    gc.collect()
    if torch.cuda.is_available():
        text = clip.tokenize([clip_text], truncate=True).cuda()
    else:
        text = clip.tokenize([clip_text], truncate=True)
    feat_clip_text = clip_model.encode_text(text).float()
    index_motion = trans_encoder.sample(feat_clip_text[0:1], False)
    pred_pose = net.forward_decoder(index_motion)
    pred_xyz = recover_from_ric((pred_pose*std+mean).float(), 22)
    output_name = hashlib.md5(clip_text.encode()).hexdigest()
    if method == 'fast':
        xyz = pred_xyz.reshape(1, -1, 22, 3)
        pose_vis = plot_3d.draw_to_batch(xyz.detach().cpu().numpy(), title_batch=None, outname=[f'output/results.gif'])
        return f'output/results.gif'
    elif method == 'slow':
        output_path = render(pred_xyz.detach().cpu().numpy().squeeze(axis=0), device_id=0, name=output_name)
        return output_path

        
# ---- Gradio Layout -----
text_prompt = gr.Textbox(label="Text prompt", lines=1, interactive=True)
video_out = gr.Video(label="Motion", mirror_webcam=False, interactive=False) 
demo = gr.Blocks()
demo.encrypt = False

with demo:
    gr.Markdown('''
            <div>
            <h1 style='text-align: center'>Generating Human Motion from Textual Descriptions with Discrete Representations (T2M-GPT)</h1>
            This space uses <a href='https://mael-zys.github.io/T2M-GPT/' target='_blank'><b>T2M-GPT models</b></a> based on Vector Quantised-Variational AutoEncoder (VQ-VAE) and Generative Pre-trained Transformer (GPT) for human motion generation from textural descriptions🤗
            <figure>
              <img src="https://huggingface.co/vumichien/T2M-GPT/resolve/main/Teaser.png" alt="T2M-GPT", width="425", height=850>
              <figcaption>Audio-visual HuBERT
              </figcaption>
            </figure>
            </div>
        ''')
    with gr.Row():
            gr.Markdown('''
            ### Generate human motion by **T2M-GPT**
            ##### Step 1. Give prompt text describing human motion
            ##### Step 2. Choice method to generate output (Fast: Sketch skeleton; Slow: SMPL mesh)
            ##### Step 3. Generate output and enjoy
            ''')
    with gr.Row():         
        gr.Markdown('''
            ### You can test by following examples:
            ''')
    examples = gr.Examples(examples=
            [ "a person jogs in place, slowly at first, then increases speed. they then back up and squat down.", 
              "a man steps forward and does a handstand", 
              "a man rises from the ground, walks in a circle and sits back down on the ground"],
          label="Examples", inputs=[text_prompt])
    
    with gr.Column():
        with gr.Row():
            text_prompt.render()
            method = gr.Dropdown(["slow", "fast"], label="Method", value="fast")
        with gr.Row():
            generate_btn = gr.Button("Generate")
            generate_btn.click(predict, [text_prompt, method], [video_out])
            print(video_out)
        with gr.Row():  
            video_out.render()
        
demo.launch(debug=True)