File size: 7,146 Bytes
45950ff
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
import sys
sys.path.append('/mnt/shenzhen2cephfs/capybarali/codes/humanoid')

import torch, yaml, os
from tqdm import tqdm
from smplx import SMPLX
from src.utils.rotation_conversions import quaternion_to_matrix, matrix_to_rotation_6d, matrix_to_axis_angle, rotation_6d_to_matrix, matrix_to_quaternion
import numpy as np
from argparse import ArgumentParser
import joblib
from data.vis import vis_3d_motion
from data.vis_g1 import vis_3d_g1
from copy import deepcopy

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

_debug_save_count = 0
_debug_save_dir = 'data/debug_transform'

parser = ArgumentParser(description="Launch MoCap processing")
parser.add_argument('--save_root', type=str, default="data/final_data")
parser.add_argument('--start_idx', type=int, default=0)
parser.add_argument('--interval', type=int, default=1)
args = parser.parse_args()

os.makedirs(args.save_root, exist_ok=True)   
os.makedirs(os.path.join(args.save_root, 'motions'), exist_ok=True)   

smplx_model = SMPLX(
    model_path='checkpoints/human_model/SMPLX_NEUTRAL.npz', 
    use_pca=False, num_expression_coeffs=100, num_betas=10, ext='npz'
).to(device)

def extract_g1_component(x):
    vel_xy = x[:, :2]
    dof = x[:, -29:]
    root_rot_mat = rotation_6d_to_matrix(x[:, 2:8])
    trans_xyz = torch.cat([torch.cumsum(vel_xy, dim=0), x[:, 10]])

    rot_mat = torch.tensor([[1, 0, 0], [0, 0, -1], [0, 1, 0]]).float()
    global_orient_mat = torch.from_numpy(root_rot_mat).squeeze(1)
    global_orient_mat = torch.einsum('ij,tjk->tik', rot_mat, global_orient_mat)
    rot_quat = matrix_to_quaternion(global_orient_mat)  # (T, 4) wxyz order

    transl = torch.from_numpy(trans_xyz).float()
    transl = torch.einsum('ij,tj->ti', rot_mat, transl)

    return dof, rot_quat, trans_xyz

def get_smplx_motion(data_path):
    smplx_data = joblib.load(data_path)

    transl = torch.from_numpy(smplx_data['smplx_transl'])
    betas = torch.from_numpy(np.load('checkpoints/humanoid_model/g1/betas.npy'))
    global_orient = torch.from_numpy(smplx_data['smplx_global_orient'])
    body_pose = torch.from_numpy(smplx_data['smplx_body_pose'])

    N = transl.shape[0]
    motion_params = dict(
        transl=transl, 
        global_orient=global_orient, 
        body_pose=body_pose, 
        betas=betas.unsqueeze(0).repeat(N, 1).float()
    )
    # 1. process positions
    frame_params = {k: v.to(device) for k, v in motion_params.items()}
    frame_params['leye_pose'] = torch.zeros((N, 3)).to(device)
    frame_params['reye_pose'] = torch.zeros((N, 3)).to(device)
    frame_params['left_hand_pose'] = torch.zeros((N, 45)).to(device)
    frame_params['right_hand_pose'] = torch.zeros((N, 45)).to(device)
    frame_params['jaw_pose'] = torch.zeros((N, 3)).to(device)
    frame_params['expression'] = torch.zeros((N, 100)).to(device)
    output = smplx_model(**frame_params)
    position_data = output.joints.detach().cpu()[:, :22] # T, 22 ,3
    position_val_data = position_data[1:] - position_data[:-1]

    root_idx = 0
    # put on floor and put root on origin for the first frame
    ori = deepcopy(position_data[0, root_idx]) # first frame root position
    y_min = torch.min(position_data[:, :, 1])
    ori[1] = y_min
    position_data = position_data - ori
    velocities_root = position_data[1:, root_idx, :] - position_data[:-1, root_idx, :]
    position_data[:,:,0] -= position_data[:,0:1,0]
    position_data[:,:,2] -= position_data[:,0:1,2]

    T, njoint, _ = position_data.shape
    final_x = torch.zeros((T, 2 + 6 + njoint * 3 + njoint * 3))
    final_x[1:, 0] = velocities_root[:, 0]
    final_x[1:, 1] = velocities_root[:, 1]
    final_x[:, 2:2+6] = matrix_to_rotation_6d(global_orient)
    final_x[:, 8:8+njoint*3] = position_data.flatten(1, 2)
    final_x[1:, 8+njoint*3:8+njoint*6] = position_val_data.flatten(1, 2) # T, 140
    return final_x

def get_g1_motion(data_path):
    global _debug_save_count
    rotation_matrix = torch.tensor([[1.0, 0, 0], [0, 0, -1], [0, 1, 0]]).inverse()

    g1_data = joblib.load(data_path)
    dof = g1_data['g1_dof'] # T, 29
    global_orient = g1_data['g1_root_ori'] # T, 4, wxyz
    joints = g1_data['g1_joints'] # T, 30, 3

    # save before-transform data for first 3 motions
    if _debug_save_count < 3:
        os.makedirs(_debug_save_dir, exist_ok=True)
        np.savez(
            os.path.join(_debug_save_dir, f'motion_{_debug_save_count:02d}_before.npz'),
            g1_trans=joints,
            g1_root_rot=global_orient,
            g1_dof=dof,
        )

    global_orient_mat = quaternion_to_matrix(torch.from_numpy(global_orient)).float()
    global_orient_mat = torch.einsum('ij,tjk->tik', rotation_matrix, global_orient_mat)
    global_orient = matrix_to_axis_angle(global_orient_mat)
    position_data = torch.einsum('ij,tkj->tki', rotation_matrix, torch.from_numpy(joints).float())
    position_val_data = position_data[1:] - position_data[:-1]

    # save after-transform data for first 3 motions
    if _debug_save_count < 3:
        np.savez(
            os.path.join(_debug_save_dir, f'motion_{_debug_save_count:02d}_after.npz'),
            g1_trans=position_data,
            g1_root_rot=global_orient.numpy(),
            g1_dof=dof,
        )

    root_idx = 0
    # put on floor and put root on origin for the first frame
    ori = deepcopy(position_data[0, root_idx]) # first frame root position
    y_min = torch.min(position_data[:, :, 1])
    ori[1] = y_min
    position_data = position_data - ori
    velocities_root = position_data[1:, root_idx, :] - position_data[:-1, root_idx, :]
    position_data[:,:,0] -= position_data[:,0:1,0]
    position_data[:,:,2] -= position_data[:,0:1,2]

    T, njoint, _ = position_data.shape
    final_x = torch.zeros((T, 2 + 6 + njoint * 3 + njoint * 3))
    final_x[1:, 0] = velocities_root[:, 0]
    final_x[1:, 1] = velocities_root[:, 1]
    final_x[:, 2:2+6] = matrix_to_rotation_6d(global_orient)
    final_x[:, 8:8+njoint*3] = position_data.flatten(1, 2)
    final_x[1:, 8+njoint*3:8+njoint*6] = position_val_data.flatten(1, 2) # T, 140
    final_x = torch.concat([final_x, torch.from_numpy(dof)], dim=-1)
    return final_x # 217

# python -m data.motionmillion.tools.process --save_root "data/motionmillion/final_data"
if __name__ == '__main__':
    with open('data/merge_gmr_retarget_smplx_path.txt', 'r') as f:
        paths = f.readlines()

    for line in tqdm(paths[args.start_idx::args.interval]):
        data_path = line.strip()
        smplx_motion = get_smplx_motion(data_path)
        g1_motion = get_g1_motion(data_path)

        if smplx_motion.shape[0] != g1_motion.shape[0]:
            min_len = min(smplx_motion.shape[0], g1_motion.shape[0])
            smplx_motion = smplx_motion[:min_len]
            g1_motion = g1_motion[:min_len]
        
        data = dict(
            g1_motion=g1_motion, 
            smplx_motion=smplx_motion, 
        )
        data_path = data_path.replace('.npz', '.pkl')
        save_path = args.save_root + '/motions/' + '/'.join(data_path.split('/')[2:])
        # os.makedirs(os.path.dirname(save_path), exist_ok=True)
        # joblib.dump(data, save_path)