-
Notifications
You must be signed in to change notification settings - Fork 77
/
Copy pathconvert_real_robot_data.py
194 lines (152 loc) · 7.57 KB
/
convert_real_robot_data.py
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
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
import os
import zarr
import pickle
import tqdm
import numpy as np
import torch
import pytorch3d.ops as torch3d_ops
import torchvision
from termcolor import cprint
import re
import time
import numpy as np
import torch
import pytorch3d.ops as torch3d_ops
import torchvision
import socket
import pickle
def farthest_point_sampling(points, num_points=1024, use_cuda=True):
K = [num_points]
if use_cuda:
points = torch.from_numpy(points).cuda()
sampled_points, indices = torch3d_ops.sample_farthest_points(points=points.unsqueeze(0), K=K)
sampled_points = sampled_points.squeeze(0)
sampled_points = sampled_points.cpu().numpy()
else:
points = torch.from_numpy(points)
sampled_points, indices = torch3d_ops.sample_farthest_points(points=points.unsqueeze(0), K=K)
sampled_points = sampled_points.squeeze(0)
sampled_points = sampled_points.numpy()
return sampled_points, indices
def preprocess_point_cloud(points, use_cuda=True):
num_points = 1024
extrinsics_matrix = np.array([[ 0.5213259, -0.84716441, 0.10262438, 0.04268034],
[ 0.25161211, 0.26751035, 0.93012341, 0.15598059],
[-0.81542053, -0.45907589, 0.3526169, 0.47807532],
[ 0., 0., 0., 1. ]])
WORK_SPACE = [
[0.65, 1.1],
[0.45, 0.66],
[-0.7, 0]
]
# scale
point_xyz = points[..., :3]*0.0002500000118743628
point_homogeneous = np.hstack((point_xyz, np.ones((point_xyz.shape[0], 1))))
point_homogeneous = np.dot(point_homogeneous, extrinsics_matrix)
point_xyz = point_homogeneous[..., :-1]
points[..., :3] = point_xyz
# crop
points = points[np.where((points[..., 0] > WORK_SPACE[0][0]) & (points[..., 0] < WORK_SPACE[0][1]) &
(points[..., 1] > WORK_SPACE[1][0]) & (points[..., 1] < WORK_SPACE[1][1]) &
(points[..., 2] > WORK_SPACE[2][0]) & (points[..., 2] < WORK_SPACE[2][1]))]
points_xyz = points[..., :3]
points_xyz, sample_indices = farthest_point_sampling(points_xyz, num_points, use_cuda)
sample_indices = sample_indices.cpu()
points_rgb = points[sample_indices, 3:][0]
points = np.hstack((points_xyz, points_rgb))
return points
def preproces_image(image):
img_size = 84
image = image.astype(np.float32)
image = torch.from_numpy(image).cuda()
image = image.permute(2, 0, 1) # HxWx4 -> 4xHxW
image = torchvision.transforms.functional.resize(image, (img_size, img_size))
image = image.permute(1, 2, 0) # 4xHxW -> HxWx4
image = image.cpu().numpy()
return image
expert_data_path = '/home/zhanggu/3D-Diffusion-Policy/3D-Diffusion-Policy/data/realdex_roll'
save_data_path = '/home/zhanggu/3D-Diffusion-Policy/3D-Diffusion-Policy/data/realdex_roll.zarr'
demo_dirs = [os.path.join(expert_data_path, d, 'data.pkl') for d in os.listdir(expert_data_path) if os.path.isdir(os.path.join(expert_data_path, d))]
# storage
total_count = 0
img_arrays = []
point_cloud_arrays = []
depth_arrays = []
state_arrays = []
action_arrays = []
episode_ends_arrays = []
if os.path.exists(save_data_path):
cprint('Data already exists at {}'.format(save_data_path), 'red')
cprint("If you want to overwrite, delete the existing directory first.", "red")
cprint("Do you want to overwrite? (y/n)", "red")
user_input = 'y'
if user_input == 'y':
cprint('Overwriting {}'.format(save_data_path), 'red')
os.system('rm -rf {}'.format(save_data_path))
else:
cprint('Exiting', 'red')
exit()
os.makedirs(save_data_path, exist_ok=True)
for demo_dir in demo_dirs:
dir_name = os.path.dirname(demo_dir)
cprint('Processing {}'.format(demo_dir), 'green')
with open(demo_dir, 'rb') as f:
demo = pickle.load(f)
pcd_dirs = os.path.join(dir_name, 'pcd')
if not os.path.exists(pcd_dirs):
os.makedirs(pcd_dirs)
demo_length = len(demo['point_cloud'])
# dict_keys(['point_cloud', 'rgbd', 'agent_pos', 'action'])
for step_idx in tqdm.tqdm(range(demo_length)):
total_count += 1
obs_image = demo['image'][step_idx]
obs_depth = demo['depth'][step_idx]
obs_image = preproces_image(obs_image)
obs_depth = preproces_image(np.expand_dims(obs_depth, axis=-1)).squeeze(-1)
obs_pointcloud = demo['point_cloud'][step_idx]
robot_state = demo['agent_pos'][step_idx]
action = demo['action'][step_idx]
obs_pointcloud = preprocess_point_cloud(obs_pointcloud, use_cuda=True)
img_arrays.append(obs_image)
action_arrays.append(action)
point_cloud_arrays.append(obs_pointcloud)
depth_arrays.append(obs_depth)
state_arrays.append(robot_state)
episode_ends_arrays.append(total_count)
# create zarr file
zarr_root = zarr.group(save_data_path)
zarr_data = zarr_root.create_group('data')
zarr_meta = zarr_root.create_group('meta')
img_arrays = np.stack(img_arrays, axis=0)
if img_arrays.shape[1] == 3: # make channel last
img_arrays = np.transpose(img_arrays, (0,2,3,1))
point_cloud_arrays = np.stack(point_cloud_arrays, axis=0)
depth_arrays = np.stack(depth_arrays, axis=0)
action_arrays = np.stack(action_arrays, axis=0)
state_arrays = np.stack(state_arrays, axis=0)
episode_ends_arrays = np.array(episode_ends_arrays)
compressor = zarr.Blosc(cname='zstd', clevel=3, shuffle=1)
img_chunk_size = (100, img_arrays.shape[1], img_arrays.shape[2], img_arrays.shape[3])
point_cloud_chunk_size = (100, point_cloud_arrays.shape[1], point_cloud_arrays.shape[2])
depth_chunk_size = (100, depth_arrays.shape[1], depth_arrays.shape[2])
if len(action_arrays.shape) == 2:
action_chunk_size = (100, action_arrays.shape[1])
elif len(action_arrays.shape) == 3:
action_chunk_size = (100, action_arrays.shape[1], action_arrays.shape[2])
else:
raise NotImplementedError
zarr_data.create_dataset('img', data=img_arrays, chunks=img_chunk_size, dtype='uint8', overwrite=True, compressor=compressor)
zarr_data.create_dataset('point_cloud', data=point_cloud_arrays, chunks=point_cloud_chunk_size, dtype='float64', overwrite=True, compressor=compressor)
zarr_data.create_dataset('depth', data=depth_arrays, chunks=depth_chunk_size, dtype='float64', overwrite=True, compressor=compressor)
zarr_data.create_dataset('action', data=action_arrays, chunks=action_chunk_size, dtype='float32', overwrite=True, compressor=compressor)
zarr_data.create_dataset('state', data=state_arrays, chunks=(100, state_arrays.shape[1]), dtype='float32', overwrite=True, compressor=compressor)
zarr_meta.create_dataset('episode_ends', data=episode_ends_arrays, chunks=(100,), dtype='int64', overwrite=True, compressor=compressor)
# print shape
cprint(f'img shape: {img_arrays.shape}, range: [{np.min(img_arrays)}, {np.max(img_arrays)}]', 'green')
cprint(f'point_cloud shape: {point_cloud_arrays.shape}, range: [{np.min(point_cloud_arrays)}, {np.max(point_cloud_arrays)}]', 'green')
cprint(f'depth shape: {depth_arrays.shape}, range: [{np.min(depth_arrays)}, {np.max(depth_arrays)}]', 'green')
cprint(f'action shape: {action_arrays.shape}, range: [{np.min(action_arrays)}, {np.max(action_arrays)}]', 'green')
cprint(f'state shape: {state_arrays.shape}, range: [{np.min(state_arrays)}, {np.max(state_arrays)}]', 'green')
cprint(f'episode_ends shape: {episode_ends_arrays.shape}, range: [{np.min(episode_ends_arrays)}, {np.max(episode_ends_arrays)}]', 'green')
cprint(f'total_count: {total_count}', 'green')
cprint(f'Saved zarr file to {save_data_path}', 'green')