-
Notifications
You must be signed in to change notification settings - Fork 5
/
nms2_nowidth.py
268 lines (220 loc) · 9.22 KB
/
nms2_nowidth.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
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
#!/usr/bin/env python
# -*- coding: utf-8 -*-
# @Date : 2019-03-10 21:04:24
# @Author : Chaozheng Wu ([email protected])
# @Version : $Id$
import os
import numpy as np
import argparse
import time
import shutil
# import pybullet
import torch
from torch.autograd import Variable
import torch.nn as nn
import torch.optim as optim
import torch.utils.data
import torch.nn.functional as F
import matplotlib.pyplot as plt
from tools.nms import nms2
def readShapePoses(fname):
f = open(fname, 'r')
lines = f.readlines()
num = len(lines)
shape_poses = {}
for l in range(0, num, 11):
shape = lines[l].strip()
grasps = lines[l+1: l+11]
grasps = np.array([g.strip().split(',') for g in grasps]).astype(float).reshape(-1, 8)
shape_poses[shape] = grasps
# print(shape, grasps.shape, grasps.dtype)
return shape_poses
parser = argparse.ArgumentParser()
parser.add_argument('--GPU', dest='GPUid', help='set GPU id', default=0, type=str)
parser.add_argument('-dp', '--data_path', type=str, default=None, help='data root dir')
parser.add_argument('--posi_only', dest='posi_only', default=False, help='whether save all prediction', action='store_true')
parser.add_argument('--anti_score', dest='anti_score', default=False, help='whether to use antipodal score', action='store_true')
opt = parser.parse_args()
def main():
pred_data_path = opt.data_path
outputdir = os.path.dirname(pred_data_path)
if opt.posi_only:
if opt.anti_score:
LOG_FOUT = open(os.path.join(outputdir, 'test_top10_poses_%s_anti_score_posi_only.txt'%(os.path.basename(pred_data_path))), 'w')
log_all = open(os.path.join(outputdir, 'test_all_poses_%s_anti_score_posi_only.txt'%(os.path.basename(pred_data_path))), 'w')
else:
LOG_FOUT = open(os.path.join(outputdir, 'test_top10_poses_%s_posi_only.txt'%(os.path.basename(pred_data_path))), 'w')
log_all = open(os.path.join(outputdir, 'test_all_poses_%s_posi_only.txt'%(os.path.basename(pred_data_path))), 'w')
else:
if opt.anti_score:
LOG_FOUT = open(os.path.join(outputdir, 'test_top10_poses_%s_anti_score.txt'%(os.path.basename(pred_data_path))), 'w')
log_all = open(os.path.join(outputdir, 'test_all_poses_%s_anti_score.txt'%(os.path.basename(pred_data_path))), 'w')
else:
LOG_FOUT = open(os.path.join(outputdir, 'test_top10_poses_%s.txt'%(os.path.basename(pred_data_path))), 'w')
log_all = open(os.path.join(outputdir, 'test_all_poses_%s.txt'%(os.path.basename(pred_data_path))), 'w')
def log_string(out_str):
LOG_FOUT.write(out_str+'\n')
LOG_FOUT.flush()
print(out_str)
keep_num = []
pp = []
rr = []
n = 21
th = 0.03
q_th = np.pi/4
files = os.listdir(opt.data_path)
for i, f in enumerate(files):
shape = f.split('.')[0]
# if shape == '64d97464f86b591caf17412e945e52f4':
# continue
print(shape, i)
# seen_num += 1
if not os.path.exists(os.path.join(pred_data_path, shape+'.npz')):
continue
f = np.load(os.path.join(pred_data_path, shape+'.npz'))
pred_cent = f['centers']
pred_quat = f['quaternions']
# widths = f['widths']
scores = f['scores']
if opt.anti_score:
anti_scores = f['anti_scores']
print('antipodal scores max-min :', anti_scores.max(), anti_scores.min())
print('grasp scores max-min :', scores.max(), scores.min())
if opt.anti_score:
idx = anti_scores > 0.5
if idx.sum() == 0:
idx = scores > 0
scores = scores[idx]
pred_cent = pred_cent[idx]
pred_quat = pred_quat[idx]
if opt.posi_only:
idx = scores > 0.5
if idx.sum() == 0:
idx = scores > 0
scores = scores[idx]
pred_cent = pred_cent[idx]
pred_quat = pred_quat[idx]
keep = nms2(pred_cent, pred_quat, scores, cent_th=0.04, ang_th=np.pi/3)
keep = np.array(keep, dtype=np.int32)
pred_cent = pred_cent[keep]
pred_quat = pred_quat[keep]
log_all.write(shape + '\n')
for i in range(pred_cent.shape[0]):
c = pred_cent[i]
q = pred_quat[i]
log_all.write('%f,%f,%f,%f,%f,%f,%f\n'%(c[0], c[1], c[2], q[0], q[1], q[2], q[3]))
log_all.flush()
keep_num = keep.shape[0]
print('keep_num: ', keep_num)
if keep_num < 10:
idx = np.concatenate([np.arange(keep_num), np.random.choice(keep_num, 10 - keep_num, replace=True)], 0)
pred_cent = pred_cent[idx]
pred_quat = pred_quat[idx]
log_string(shape)
for i in range(10):
c = pred_cent[i]
q = pred_quat[i]
log_string('%f,%f,%f,%f,%f,%f,%f'%(c[0], c[1], c[2], q[0], q[1], q[2], q[3]))
def check_pred_pose(pred_poses, gt_poses, cent_th=0.035, ang_th=5*np.pi/18):
k = pred_poses.size(0)
topk_correct = []
topk_ang_correct = []
topk_cent_correct = []
gt_cent = gt_poses[:, 1:4]
gt_quat = gt_poses[:, 4:]
cover = []
for i in range(k):
p = pred_poses[i]
cent = p[1:4]
quat = p[4:]
# print(p.size())
cent_correct = (torch.sum((gt_cent - cent)**2, 1) <= cent_th**2).float()
quat_diff = torch.abs(gt_quat.matmul(quat))
# print(quat_diff.max().item(), quat_diff.min().item())
quat_correct = (2 * torch.acos(quat_diff) <= ang_th).float()
all_correct = cent_correct * quat_correct
correct_idx = torch.nonzero(quat_correct).view(-1)
cover.append(correct_idx)
if all_correct.sum()>0:
topk_correct.append(True)
else:
topk_correct.append(False)
if quat_correct.sum()>0:
topk_ang_correct.append(True)
else:
topk_ang_correct.append(False)
if cent_correct.sum()>0:
topk_cent_correct.append(True)
else:
topk_cent_correct.append(False)
cover = torch.unique(torch.cat(cover, 0))
return topk_correct, topk_ang_correct, topk_cent_correct, cover
def dist_matrix_torch(x, y):
x2 = torch.sum(x**2, -1, keepdim=True)
y2 = torch.sum(y**2, -1, keepdim=True)
xy = torch.matmul(x, y.transpose(-1, -2))
matrix = x2 - 2*xy + y2.transpose(-1, -2)
matrix[matrix<=0] = 1e-10
return torch.sqrt(matrix)
def check_pred_pose_batch(pred_poses, gt_poses, cent_th=0.035, ang_th=5*np.pi/18, GPU=False):
if GPU:
pred_poses = pred_poses.cuda(0)
gt_poses = gt_poses.cuda(0)
gt_cent = gt_poses[:, 1:4]
gt_quat = gt_poses[:, 4:]
pred_cent = pred_poses[:, :3]
print(pred_cent.size(), gt_cent.size())
pred_quat = pred_poses[:, 3:]
pred_num = pred_cent.size(0)
if pred_num > 50000:
pred_correct = []
pred_cent_correct = []
pred_quat_correct = []
num = pred_num // 20000 + 1
delta = pred_num // num
for i in range(num):
s = delta * i
e = delta * (i+1)
if i+1 == num:
e = max(e, pred_num)
cent_dist_ = dist_matrix_torch(pred_cent[s:e], gt_cent)
quat_diff_ = torch.abs(torch.matmul(pred_quat[s:e], gt_quat.t()))
cent_correct = (cent_dist_ <= cent_th).float()
quat_correct = (2 * torch.acos(quat_diff_) <= ang_th).float()
all_correct = cent_correct * quat_correct
pred_correct_ = (all_correct.sum(-1) > 0).cpu().float().numpy()
pred_cent_correct_ = (cent_correct.sum(-1) > 0).cpu().float().numpy()
pred_quat_correct_ = (quat_correct.sum(-1) > 0).cpu().float().numpy()
pred_correct.append(pred_correct_)
pred_cent_correct.append(pred_cent_correct_)
pred_quat_correct.append(pred_quat_correct_)
pred_correct = np.concatenate(pred_correct, 0)
pred_cent_correct = np.concatenate(pred_cent_correct, 0)
pred_quat_correct = np.concatenate(pred_quat_correct, 0)
assert pred_correct.shape[0] == pred_num
else:
cent_dist = dist_matrix_torch(pred_cent, gt_cent)
quat_diff = torch.abs(torch.matmul(pred_quat, gt_quat.t()))
cent_correct = (cent_dist <= cent_th).float()
quat_correct = (2 * torch.acos(quat_diff) <= ang_th).float()
all_correct = cent_correct * quat_correct
pred_correct = (all_correct.sum(-1) > 0).cpu().float().numpy()
pred_cent_correct = (cent_correct.sum(-1) > 0).cpu().float().numpy()
pred_quat_correct = (quat_correct.sum(-1) > 0).cpu().float().numpy()
return pred_correct, pred_cent_correct, pred_quat_correct
def EulerPose2QuaternionPose(poses):
poses[:,4] = poses[:, 4] * 2*np.pi - np.pi
poses[:,5] = poses[:, 5] * np.pi - np.pi/2
poses[:,6] = poses[:, 6] * 2*np.pi - np.pi
q_poses = []
for p in poses:
Q = pybullet.getQuaternionFromEuler([p[4], p[5], p[6]])
Q = np.array([Q[3], Q[0], Q[1], Q[2]])
pose = np.zeros(8)
pose[:4] = p[:4]
pose[4:] = Q
q_poses.append(pose)
q_poses = torch.FloatTensor(np.array(q_poses)).view(1, -1, 8)
return q_poses
if __name__ == '__main__':
main()