-
Notifications
You must be signed in to change notification settings - Fork 562
Expand file tree
/
Copy pathtest_optim.py
More file actions
262 lines (192 loc) · 7.28 KB
/
Copy pathtest_optim.py
File metadata and controls
262 lines (192 loc) · 7.28 KB
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
260
261
262
"""Tests for module optim fro OT optimization"""
# Author: Remi Flamary <remi.flamary@unice.fr>
#
# License: MIT License
import numpy as np
import ot
def test_conditional_gradient(nx):
n_bins = 100 # nb bins
# bin positions
x = np.arange(n_bins, dtype=np.float64)
# Gaussian distributions
a = ot.datasets.make_1D_gauss(n_bins, m=20, s=5) # m= mean, s= std
b = ot.datasets.make_1D_gauss(n_bins, m=60, s=10)
# loss matrix
M = ot.dist(x.reshape((n_bins, 1)), x.reshape((n_bins, 1)))
M /= M.max()
def f(G):
return 0.5 * np.sum(G**2)
def df(G):
return G
def fb(G):
return 0.5 * nx.sum(G**2)
ab, bb, Mb = nx.from_numpy(a, b, M)
reg = 1e-1
G, log = ot.optim.cg(a, b, M, reg, f, df, verbose=True, log=True)
Gb, log = ot.optim.cg(ab, bb, Mb, reg, fb, df, verbose=True, log=True)
Gb = nx.to_numpy(Gb)
np.testing.assert_allclose(Gb, G)
np.testing.assert_allclose(a, Gb.sum(1))
np.testing.assert_allclose(b, Gb.sum(0))
def test_conditional_gradient_itermax(nx):
n = 100 # nb samples
mu_s = np.array([0, 0])
cov_s = np.array([[1, 0], [0, 1]])
mu_t = np.array([4, 4])
cov_t = np.array([[1, -0.8], [-0.8, 1]])
xs = ot.datasets.make_2D_samples_gauss(n, mu_s, cov_s)
xt = ot.datasets.make_2D_samples_gauss(n, mu_t, cov_t)
a, b = np.ones((n,)) / n, np.ones((n,)) / n
# loss matrix
M = ot.dist(xs, xt)
M /= M.max()
def f(G):
return 0.5 * np.sum(G**2)
def df(G):
return G
def fb(G):
return 0.5 * nx.sum(G**2)
ab, bb, Mb = nx.from_numpy(a, b, M)
reg = 1e-1
G, log = ot.optim.cg(
a, b, M, reg, f, df, numItermaxEmd=10000, verbose=True, log=True
)
Gb, log = ot.optim.cg(
ab, bb, Mb, reg, fb, df, numItermaxEmd=10000, verbose=True, log=True
)
Gb = nx.to_numpy(Gb)
np.testing.assert_allclose(Gb, G)
np.testing.assert_allclose(a, Gb.sum(1))
np.testing.assert_allclose(b, Gb.sum(0))
def test_generalized_conditional_gradient(nx):
n_bins = 100 # nb bins
# bin positions
x = np.arange(n_bins, dtype=np.float64)
# Gaussian distributions
a = ot.datasets.make_1D_gauss(n_bins, m=20, s=5) # m= mean, s= std
b = ot.datasets.make_1D_gauss(n_bins, m=60, s=10)
# loss matrix
M = ot.dist(x.reshape((n_bins, 1)), x.reshape((n_bins, 1)))
M /= M.max()
def f(G):
return 0.5 * np.sum(G**2)
def df(G):
return G
def fb(G):
return 0.5 * nx.sum(G**2)
reg1 = 1e-3
reg2 = 1e-1
ab, bb, Mb = nx.from_numpy(a, b, M)
G, log = ot.optim.gcg(a, b, M, reg1, reg2, f, df, verbose=True, log=True)
Gb, log = ot.optim.gcg(ab, bb, Mb, reg1, reg2, fb, df, verbose=True, log=True)
Gb = nx.to_numpy(Gb)
np.testing.assert_allclose(Gb, G, atol=1e-12)
np.testing.assert_allclose(a, Gb.sum(1), atol=1e-05)
np.testing.assert_allclose(b, Gb.sum(0), atol=1e-05)
def test_gcg_no_reg():
n = 100
X = ot.datasets.make_2D_samples_gauss(n, m=[0, 0], sigma=np.eye(2))
Y = ot.datasets.make_2D_samples_gauss(n, m=[1, 1], sigma=np.eye(2))
a = ot.unif(n)
b = ot.unif(n)
M = ot.dist(X, Y)
M /= M.max()
eps = 1e-2
G_gcg = ot.optim.gcg(
a,
b,
M,
eps,
0.0,
lambda x: 0.0,
lambda x: np.zeros(x.shape),
numInnerItermax=1000,
numItermax=10,
)
G_sinkhorn = ot.sinkhorn(a, b, M, eps, numItermax=1000)
np.testing.assert_allclose(G_gcg, G_sinkhorn, atol=1e-12)
np.testing.assert_allclose(a, G_gcg.sum(1), atol=1e-5)
np.testing.assert_allclose(b, G_gcg.sum(0), atol=1e-5)
def test_solve_1d_linesearch_quad_funct():
np.testing.assert_allclose(ot.optim.solve_1d_linesearch_quad(1, -1), 0.5)
np.testing.assert_allclose(ot.optim.solve_1d_linesearch_quad(-1, 5), 0)
np.testing.assert_allclose(ot.optim.solve_1d_linesearch_quad(-1, 0.5), 1)
def test_line_search_armijo(nx):
xk = np.array([[0.25, 0.25], [0.25, 0.25]])
pk = np.array([[-0.25, 0.25], [0.25, -0.25]])
gfk = np.array([[23.04273441, 23.0449082], [23.04273441, 23.0449082]])
old_fval = -123.0
xkb, pkb, gfkb = nx.from_numpy(xk, pk, gfk)
def f(x):
return 1.0
# Should not throw an exception and return 0. for alpha
alpha, a, b = ot.optim.line_search_armijo(f, xkb, pkb, gfkb, old_fval)
alpha_np, anp, bnp = ot.optim.line_search_armijo(f, xk, pk, gfk, old_fval)
assert a == anp
assert b == bnp
assert alpha == 0.0
# check line search armijo
def f(x):
return nx.sum((x - 5.0) ** 2)
def grad(x):
return 2 * (x - 5.0)
xk = nx.from_numpy(np.array([[[-5.0, -5.0]]]))
pk = nx.from_numpy(np.array([[[100.0, 100.0]]]))
gfk = grad(xk)
old_fval = f(xk)
# chech the case where the optimum is on the direction
alpha, _, _ = ot.optim.line_search_armijo(f, xk, pk, gfk, old_fval)
np.testing.assert_allclose(alpha, 0.1)
# check the case where the direction is not far enough
pk = nx.from_numpy(np.array([[[3.0, 3.0]]]))
alpha, _, _ = ot.optim.line_search_armijo(f, xk, pk, gfk, old_fval, alpha0=1.0)
np.testing.assert_allclose(alpha, 1.0)
# check the case where checking the wrong direction
alpha, _, _ = ot.optim.line_search_armijo(f, xk, -pk, gfk, old_fval)
assert alpha <= 0
# check the case where the point is not a vector
xk = nx.from_numpy(np.array(-5.0))
pk = nx.from_numpy(np.array(100.0))
gfk = grad(xk)
old_fval = f(xk)
alpha, _, _ = ot.optim.line_search_armijo(f, xk, pk, gfk, old_fval)
np.testing.assert_allclose(alpha, 0.1)
def test_line_search_armijo_dtype_device(nx):
for tp in nx.__type_list__:
def f(x):
return nx.sum((x - 5.0) ** 2)
def grad(x):
return 2 * (x - 5.0)
xk = np.array([[[-5.0, -5.0]]])
pk = np.array([[[100.0, 100.0]]])
xkb, pkb = nx.from_numpy(xk, pk, type_as=tp)
gfkb = grad(xkb)
old_fval = f(xkb)
# chech the case where the optimum is on the direction
alpha, _, fval = ot.optim.line_search_armijo(f, xkb, pkb, gfkb, old_fval)
alpha = nx.to_numpy(alpha)
np.testing.assert_allclose(alpha, 0.1)
nx.assert_same_dtype_device(old_fval, fval)
# check the case where the direction is not far enough
pk = np.array([[[3.0, 3.0]]])
pkb = nx.from_numpy(pk, type_as=tp)
alpha, _, fval = ot.optim.line_search_armijo(
f, xkb, pkb, gfkb, old_fval, alpha0=1.0
)
alpha = nx.to_numpy(alpha)
np.testing.assert_allclose(alpha, 1.0)
nx.assert_same_dtype_device(old_fval, fval)
# check the case where checking the wrong direction
alpha, _, fval = ot.optim.line_search_armijo(f, xkb, -pkb, gfkb, old_fval)
alpha = nx.to_numpy(alpha)
assert alpha <= 0
nx.assert_same_dtype_device(old_fval, fval)
# check the case where the point is not a vector
xkb = nx.from_numpy(np.array(-5.0), type_as=tp)
pkb = nx.from_numpy(np.array(100), type_as=tp)
gfkb = grad(xkb)
old_fval = f(xkb)
alpha, _, fval = ot.optim.line_search_armijo(f, xkb, pkb, gfkb, old_fval)
alpha = nx.to_numpy(alpha)
np.testing.assert_allclose(alpha, 0.1)
nx.assert_same_dtype_device(old_fval, fval)