forked from Ronnypetson/tfoe_vo
-
Notifications
You must be signed in to change notification settings - Fork 0
/
opt_slocal.py
222 lines (197 loc) · 7.08 KB
/
opt_slocal.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
import numpy as np
from scipy.optimize import minimize, least_squares
from reproj import gen_pts
from reproj import reproj_tc_foe
from reproj import reproj_tc_foe_slocal
from liegroups import SE3
import torch
import torch.nn.functional as F
from liegroups.torch import SE3 as SE3tc
from hessian import jacobian, hessian
from utils import compose_slocal
class OptSingle:
def __init__(self, x, x_, T0ij, c, g):
''' x and x_ are (i, j) -> 3xN
'''
self.g = g
self.x = x # (i, j) -> x
self.x_ = x_ # (i, j) -> x_
self.T0ij = T0ij
#self.f = {} # (i, j) -> f
#for k in x:
# self.f[k] = x_[k] - x[k]
self.T = np.zeros(6)
self.foe = np.zeros(2)
self.c = torch.from_numpy(c) #.float()
self.c_ = torch.inverse(self.c)
self.min_obj = np.inf
def obj_tc(self, Tfoe):
T = Tfoe[:6]
foe = Tfoe[6:]
foe = foe.unsqueeze(-1)
T = SE3tc.exp(T.clone())
T = T.as_matrix()
c = self.c
c_ = self.c_
x_rep = reproj_tc_foe(torch.from_numpy(self.x),
torch.from_numpy(self.x_),
T, foe, c)
y = c_ @ torch.from_numpy(self.x_)-x_rep
y = torch.mean(y**2.0)
return y
def obj_npy(self, Tfoe):
Tfoe = torch.from_numpy(Tfoe)
Tfoe = Tfoe.requires_grad_(True)
Tfoe.retain_grad()
Tfoe_ = Tfoe.clone()
y = self.obj_tc(Tfoe_)
y = y.detach().numpy()
return y
def jac_npy(self, Tfoe):
Tfoe = torch.from_numpy(Tfoe)
Tfoe = Tfoe.requires_grad_(True)
jac = jacobian(self.obj_tc(Tfoe),Tfoe)
jac = jac.detach().numpy()
return jac
def hess_npy(self, Tfoe):
Tfoe = torch.from_numpy(Tfoe)
Tfoe = Tfoe.requires_grad_(True)
hess = hessian(self.obj_tc(Tfoe), Tfoe)
hess = hess.detach().numpy()
return hess
def objective(self, Tfoe, grad=True, residuals=False):
''' Tfoe is n*9
'''
Tfoe = torch.from_numpy(Tfoe)
Tfoe = Tfoe.requires_grad_(True)
Tfoe.retain_grad()
Tfoe_ = Tfoe.clone()
Tfoe_ = Tfoe_.reshape(-1, 10)
g = self.g # bundle graph
c = self.c
c_ = self.c_
T = Tfoe_[:, :6]
foe = Tfoe_[:, 6:9]
scale = Tfoe_[:, 9:]
foe = foe.unsqueeze(-1)
T = SE3tc.exp(T.clone())
T = T.as_matrix()
y = 0.0
#resid = []
for ij in g:
Tij, foeij, ep_ = compose_slocal(ij[0], ij[1],
T.clone(), foe.clone(),
scale.clone(), c, base=self.base)
x_rep, den = reproj_tc_foe_slocal(torch.from_numpy(self.x[ij]),
torch.from_numpy(self.x_[ij]),
Tij, foeij, c)
yij = F.smooth_l1_loss(c_ @ torch.from_numpy(self.x_[ij]), x_rep)
#yij = (c_ @ torch.from_numpy(self.x_[ij]) - x_rep)**2
#yij = den * yij
yij = torch.mean(yij)
if self.x[ij].shape[1] == 32:
yij *= 1e-14
elif False:
T0ij = torch.from_numpy(self.T0ij[ij])
yt_ij = F.smooth_l1_loss(Tij[:3, :3], T0ij[:3, :3])
yij += 1e-2 * yt_ij
#yij = yt_ij
#t0ij = T0ij[:3, 3:]
#ep0ij = (c @ (t0ij / (t0ij[-1] + 1e-10))) / 1e3
#ep0ij = ep0ij[:2]
#x_rep0 = reproj_tc_foe_local(torch.from_numpy(self.x[ij]),
# torch.from_numpy(self.x_[ij]),
# T0ij, ep0ij, c)
#yij0 = F.smooth_l1_loss(c_ @ torch.from_numpy(self.x_[ij]), x_rep0)
#yt_ij_t = torch.sum(torch.abs(Tij[:2, 3] - T0ij[:2, 3]))
#y_ep = F.smooth_l1_loss(foeij / (foeij[-1] + 1e-10),
# ep_ / (ep_[-1] + 1e-10))
#yt_ij = F.smooth_l1_loss(Tij, T0ij)
#if ij[1] - ij[0] == 2 and False:
# print(ij)
# #print(Tij)
# #print(T0ij)
# #print(yt_ij_t)
# print((foeij / foeij[-1]).detach().numpy())
# print((ep_ / ep_[-1]).detach().numpy())
# print(yij.item()) # , yt_ij.item()
# input()
#if abs(ij[1] - ij[0]) > 1:
# y = y + yt_ij # + 1e-2*yt_ij # + 1e-4*yt_ij #
#else:
# y = y + yij
y = y + yij
#input()
#resid = torch.cat(resid, dim=0)
y = y / len(g)
y.backward()
gradTfoe = Tfoe.grad.detach().numpy()
y = y.detach().numpy()
self.min_obj = min(self.min_obj, y)
if grad:
#print(gradTfoe)
#print(Tfoe.detach().numpy())
#input()
return y, gradTfoe
elif residuals:
return None
else:
return y
def optimize(self, T0, foe0, scale, freeze=True):
''' T0 is n,6
foe0 is n,2
scale is n,1
'''
self.base = np.min([ij[0] for ij in self.g])
self.min_obj = np.inf
Tfoe0 = np.concatenate([T0, foe0, scale], axis=-1) # n,10
Tfoe0 = Tfoe0.reshape((-1,))
#Tfoe0 = np.expand_dims(Tfoe0, axis=-1)
bounds = []
if freeze:
for par in Tfoe0:
bounds.append((par-1e-10, par+1e-10))
#for par in Tfoe0[:6]:
# bounds.append((par-1e-10, par+1e-10))
#for par in Tfoe0[6:]:
# bounds.append((None, None))
else:
for i, par in enumerate(Tfoe0):
#if i % 9 > 5 or i % 9 in [1, 3, 5]:
if i % 10 == 9:
bounds.append((par - 1e-10, par + 1e-10))
elif (i % 10 < 3 or i % 10 in [6, 7, 8]) and False:
bounds.append((par - 1e-10, par + 1e-10))
else:
bounds.append((None, None))
res = minimize(self.objective,
Tfoe0, method='L-BFGS-B',
jac=True,
bounds=bounds,
tol=1e-14,
options={'disp': False,
'maxiter': 1e2,
'gtol': 1e-12,
'ftol': 1e-12,
'maxcor': len(Tfoe0)})
#res = least_squares(self.objective,
# Tfoe0, method='lm',
# kwargs={'grad': False,
# 'residuals': True})
return res.x
if __name__ == '__main__':
x, x_, d, R, t = gen_pts(3)
opt = OptSingle(x, x_)
T0 = np.zeros(6)
foe0 = np.zeros(2)
with torch.autograd.set_detect_anomaly(False):
Tfoe = opt.optimize(T0, foe0)
T = Tfoe[:6]
foe = Tfoe[6:]
T = SE3.exp(T).as_matrix()
print(T[:3, :3])
print(R)
print(foe)
print()
print(T[:3, 3:])
print(t)