-
Notifications
You must be signed in to change notification settings - Fork 1
/
toy_quad.py
253 lines (220 loc) · 9.14 KB
/
toy_quad.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
# code mostley based on "Complex Momentum for Optimization in Games"
import numpy as np
import matplotlib.pyplot as plt
from scipy.stats import norm
import matplotlib as mpl
from tqdm import tqdm
def get_spectral_radius(g, eta, beta, is_eg=False, is_eg_and_cm=False,
is_og=False):
if is_eg:
eig_vals, eig_vecs = np.linalg.eig(g)
eg_eig_vals = 1 - eta * (eig_vals * (1 - beta.real * eig_vals))
eig_norms = np.real(np.sqrt(eg_eig_vals * np.conjugate(eg_eig_vals)))
max_norm = np.max(eig_norms)
return max_norm
elif is_eg_and_cm:
eig_vals, eig_vecs = np.linalg.eig(g)
eg_eig_vals = eig_vals * (1 - eta * eig_vals)
g = np.diag(eg_eig_vals)
elif is_og:
n = len(g)
R_t_min_1 = np.concatenate((np.eye(n) * 0.0, np.eye(n)))
R_t = np.concatenate((beta.real * g,
np.eye(n) - 2.0 * eta.real * g))
R = np.concatenate([R_t_min_1, R_t], axis=1)
eig_vals, eig_vecs = np.linalg.eig(R)
eig_norms = np.real(np.sqrt(eig_vals * np.conjugate(eig_vals)))
max_norm = np.max(eig_norms)
return max_norm
n = len(g)
R_re = np.concatenate((beta.real * np.eye(n),
-beta.imag * np.eye(n), -g))
R_im = np.concatenate([(beta.imag) * np.eye(n),
beta.real * np.eye(n),
0 * np.eye(n)])
R_param = np.concatenate([
(eta * beta).real * np.eye(n),
-(eta * beta).imag * np.eye(n),
np.eye(n) - eta.real * g])
R = np.concatenate([R_re, R_im, R_param], axis=1)
eig_vals, eig_vecs = np.linalg.eig(R)
eig_norms = np.real(np.sqrt(eig_vals * np.conjugate(eig_vals)))
max_norm = np.max(eig_norms)
return max_norm
def get_spectral_radius_LEAD(nabla_v, M, D, P, eta, beta, alpha):
n = nabla_v.shape[0] // 2
j11 = (1 + beta) * np.eye(2 * n) - eta * nabla_v - alpha * np.dot(D, P)
j12 = - beta * np.eye(2 * n) + alpha * nabla_v + alpha * np.dot(D, P)
j21 = np.eye(2 * n)
j22 = np.zeros((2 * n, 2 * n))
row_1 = np.concatenate([j11, j12], 1)
row_2 = np.concatenate([j21, j22], 1)
j = np.concatenate([row_1, row_2])
eig_vals, eig_vecs = np.linalg.eig(j)
eig_norms = np.real(np.sqrt(eig_vals * np.conjugate(eig_vals)))
max_norm = np.max(eig_norms)
return max_norm
np.random.seed(0)
n_game = 14
sample_constant_adv = 4.0
rand_vals_1 = np.linspace(
1.0 / sample_constant_adv, sample_constant_adv, num=n_game // 2)
matrix_A = np.diag(rand_vals_1)
adv_component = np.zeros((n_game, n_game))
adv_component[:n_game // 2, n_game // 2:] = -matrix_A
adv_component[n_game // 2:, :n_game // 2] = matrix_A
sample_constant_coop = sample_constant_adv
matrix_B = np.linspace(
1.0 / sample_constant_coop, sample_constant_coop, num=n_game)
matrix_B[:n_game // 2] = np.linspace(
1.0 / sample_constant_coop, sample_constant_coop, num=n_game // 2)
matrix_B[-n_game // 2:] = matrix_B[:n_game // 2]
coop_component = np.diag(matrix_B)
num_hist_bins = 50
dist_flag = False
g = None
plt.figure(figsize=(8, 5))
ax2 = plt.subplot(1, 1, 1)
cmap = mpl.cm.get_cmap('viridis')
# beta is the momentum
# eta is the learning rate that we do the for-loop over
for beta_phase, beta_phase_string in (
[(None, 'LEAD'),
(0.0, 'PM'),
(np.pi / 2.0, r'CM (pi/2)'),
(np.pi, r'NM'),
(None, r'EG'),
(None, r'OG'),
(None, r'GDA')]):
mixture_scores = []
mixture_space = np.linspace(0, 1, num=20)
alpha_space = np.linspace(-0.5, 0.5, num=num_hist_bins)
for mixture in tqdm(mixture_space):
np.random.seed(0)
gamma = np.zeros(n_game)
gamma[:n_game // 2] = np.random.uniform(
0, mixture, size=n_game // 2)
gamma[0] = 0
gamma[n_game // 2 - 1] = mixture
gamma[-n_game // 2:] = gamma[:n_game // 2]
g = gamma * adv_component + (1 - gamma) * coop_component
beta_mags, eta_mags, rates = [], [], []
if beta_phase is not None:
beta_mag_space = np.linspace(0.0, 1.0, num=num_hist_bins)
for beta_mag in beta_mag_space:
beta = beta_mag * np.exp(1j * beta_phase)
eta_space = np.linspace(.0, 1.0, num=num_hist_bins)
for eta in eta_space:
is_eg_and_cm = False
spectral_radius = get_spectral_radius(g=g, eta=eta, beta=beta, is_eg_and_cm=is_eg_and_cm)
beta_mags += [beta_mag]
eta_mags += [eta]
if is_eg_and_cm:
rates += [np.sqrt(spectral_radius)]
else:
rates += [spectral_radius]
else:
if beta_phase_string == 'LEAD':
n = adv_component.shape[0]
gamma = np.diag(gamma)
M = (np.dot(np.eye(n) - gamma, coop_component) +
np.dot(coop_component.T, (np.eye(n) - gamma).T)) / 2
row_1 = np.concatenate([M, np.dot(gamma, adv_component)], 1)
row_2 = np.concatenate([-np.dot(adv_component.T, gamma.T), M], 1)
nabla_v = np.concatenate([row_1, row_2])
row_1 = np.concatenate([np.dot(gamma, adv_component), np.zeros((n, n))], 1)
row_2 = np.concatenate([np.zeros((n, n)), -np.dot(adv_component.T, gamma.T)], 1)
D = np.concatenate([row_1, row_2])
row_1 = np.concatenate([np.zeros((n, n)), np.eye(n)], 1)
row_2 = np.concatenate([np.eye(n), np.zeros((n, n))], 1)
P = np.concatenate([row_1, row_2])
eta_space = np.linspace(0.0, 1.75, num=num_hist_bins)
alphas = []
for eta in eta_space:
beta_mag_space = eta_space
for beta_mag in beta_mag_space:
beta = beta_mag
if beta_phase_string == r'EG':
spectral_radius = get_spectral_radius(g=g, eta=eta, beta=beta, is_eg=True)
beta_mags += [beta_mag]
eta_mags += [eta]
rates += [np.sqrt(spectral_radius)]
elif beta_phase_string == r'OG':
spectral_radius = get_spectral_radius(g=g, eta=eta, beta=beta, is_og=True)
beta_mags += [beta_mag]
eta_mags += [eta]
rates += [spectral_radius]
elif beta_phase_string == 'LEAD':
for alpha in alpha_space:
spectral_radius = get_spectral_radius_LEAD(nabla_v, M, D, P, eta, beta, alpha)
beta_mags += [beta_mag]
eta_mags += [eta]
alphas += [alpha]
rates += [spectral_radius]
else:
spectral_radius = get_spectral_radius(g=g, eta=eta, beta=0.0, is_eg=False)
beta_mags += [0.0]
eta_mags += [eta]
rates += [spectral_radius]
best_rate = np.min(rates)
mixture_scores += [best_rate]
best_index = np.argmin(rates)
print('Method:', beta_phase_string)
if len(alphas) > best_index:
print('best alpha:', alphas[best_index])
if len(eta_mags) > best_index:
print('best eta:', eta_mags[best_index])
if len(beta_mags) > best_index:
print('best beta:', beta_mags[best_index])
label = None
if beta_phase_string is not None:
label = beta_phase_string
c = None
marker = None
markersize = None
if beta_phase is not None:
c = cmap(beta_phase / np.pi)
if beta_phase_string == 'PM':
c = 'tab:purple'
marker = 'x'
markersize = 12
elif beta_phase_string == 'LEAD':
c = 'tab:orange'
marker = 'x'
markersize = 12
elif beta_phase_string == r'NM':
c = 'tab:green'
marker = 'x'
markersize = 12
elif beta_phase_string == r'EG':
c = 'tab:red'
marker = 'o'
label = beta_phase_string
elif beta_phase_string == r'OG':
c = 'tab:gray'
marker = 'o'
label = beta_phase_string
elif beta_phase_string == r'GDA':
c = 'tab:blue'
marker = '+'
markersize = 12
label = beta_phase_string
ax2.set_xlim([0.0, 1.0])
ax2.semilogy(mixture_space, mixture_scores,
label=label, c=c, alpha=0.75,
marker=marker, markersize=markersize,
linewidth=3, zorder=10)
ax2.legend(framealpha=1.0)
ax2.set_xlabel('Max adversarialness γmax')
ax2.set_ylabel('# grad. eval. to converge')
ax2.xaxis.set_ticks_position('none')
ax2.yaxis.set_ticks_position('none')
ax2.grid(color="k", linestyle="--", linewidth=0.5, alpha=0.3)
if dist_flag:
ax3 = ax2.twinx()
cur_xs = np.linspace(0.0, 1.0, 10)
cur_ys = norm.pdf(cur_xs, 0.25, 0.27)
ax3.fill_between(cur_xs, cur_ys, alpha=0.25, zorder=0)
ax3.set_ylim([0.0, 1.5])
dist_flag = False
plt.savefig('phase_compare.pdf', transparent=True, bbox_inches='tight')