-
Notifications
You must be signed in to change notification settings - Fork 1
/
Illustrative_experiments.py
192 lines (155 loc) · 7.9 KB
/
Illustrative_experiments.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
import argparse
import os
import tensorflow as tf
import numpy as np
import matplotlib.pyplot as plt
import matplotlib.patches as patches
from tqdm import tqdm
from lib.utils import *
from lib.ops import conditioner
from itertools import combinations
ds = tf.contrib.distributions
from models.model_toy import generator,discriminator
parser = argparse.ArgumentParser('')
parser.add_argument('--batch_size', '-b',type=int, default=64)
parser.add_argument('--data_size',type=int, default=64)
parser.add_argument('--num_agent',type=int, default=3)
parser.add_argument('--dataset', '-d', type=str, default='cifar10') #'mnist', 'cifar10'
parser.add_argument('--distribution', type=str, default='small') #'small', 'big'
parser.add_argument('--objective', '-o', type=str, default='gan') #gan, hinge, wgan-gp
parser.add_argument('--model', '-m', type=str, default='dcgan') #dcgan, resnet
parser.add_argument('--gen_type', '-c', type=str, default='independent') #conditional, independent
parser.add_argument('--reg', '-r', type=str, default='d_reg') #d_reg, g_reg
parser.add_argument('--z_dim', '-z', type=int, default=128)
parser.add_argument('--scale', type=float, default=10.0)
parser.add_argument('--n_dis', type=int, default=1, help='number of discriminator update per generator update')
parser.add_argument('--max_iter', type=int, default=5000)
parser.add_argument('--decay', type=float, default=0.999)
#parser.add_argument('--sn', type=bool, default=False)
parser.add_argument('--gpu', '-g', type=int, default=0, help='GPU ID (negative value indicates CPU)')
#parser.add_argument('--out', default='logs_dogs_cats_mt_64', help='Directory to output the result')
parser.add_argument('--viz_size', type=int, default=25, help='number of images to display')
parser.add_argument('--cal_every', type=int, default=10000, help='Interval of evaluation')
parser.add_argument('--viz_every', type=int, default=200,help='Interval of display')
# args = parser.parse_args()
args = parser.parse_args("--gpu 0 ".split())
os.environ['CUDA_VISIBLE_DEVICES'] = str(args.gpu)
print(args)
bn_g = False
sn_g = False
sn_d = False
params = dict(disc_learning_rate=0.0002,gen_learning_rate=0.0002,beta1=0.0,beta2=0.9)
num_modes = np.power(2,args.num_agent) -1
batch_size = np.power(2,args.num_agent-1) * args.batch_size
batch_size_z = args.batch_size
#params['batch_size_z'] = num_modes * params['batch_size_each_mode']
slim = tf.contrib.slim
connection_map = [[0,3,5,6],
[1,4,5,6],
[2,3,4,6]]
list_loc= [[-1 ,-1], [0.0, 1], [1 ,-1], [0 ,-1], [1.0, 0], [-1.0,0], [0,0]]
def sample_mog(batch_size, num_mixt=4,num_agent=3, std=0.01):
data_sample=[]
cat = ds.Categorical(tf.zeros(num_mixt))
for i in range(num_agent):
loc = [list_loc[r] for r in connection_map[i]]
comps = [ds.MultivariateNormalDiag([float(p[0]),float(p[1])], [std, std]) for p in loc]
data_sample.append(ds.Mixture(cat, comps).sample(batch_size))
return data_sample
def disc_loss(real,fake):
return tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=real, labels=tf.ones_like(real))
+tf.nn.sigmoid_cross_entropy_with_logits(logits=fake, labels=tf.zeros_like(fake)))
def gen_loss(fake):
return tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=fake, labels=tf.ones_like(fake)))
tf.reset_default_graph()
data_samples=sample_mog(batch_size,std=0.1)
noise = ds.Normal(tf.zeros(args.z_dim), tf.ones(args.z_dim)).sample(num_modes*batch_size_z)
#noise = ds.Normal(tf.zeros(args.z_dim), tf.ones(args.z_dim)).sample(batch_size_z)
#noise = tf.concat([noise]*num_modes,axis=0)
modes = tf.concat([i*tf.ones(batch_size_z, dtype='int32') for i in range(num_modes)],axis=0)
is_training_pl = tf.placeholder(tf.bool, [], name='is_training_pl')
#generator
if args.gen_type == 'conditional':
gen_out = generator(noise, is_training=is_training_pl ,mode=modes, num_modes=num_modes, sn=sn_g, bn= bn_g, name="gen")
gen_list = tf.split(gen_out,num_modes,axis=0)
elif args.gen_type == 'independent':
gen_list = []
noise_list = tf.split(noise,num_modes,axis=0)
for i in range(num_modes):
gen_list.append(generator(noise_list[i], is_training=is_training_pl , sn=sn_g, bn= bn_g, name="gen_{}".format(i)))
#standard discriminators
dis_list=[[0]*2 for j in range(args.num_agent)]
for i in range(args.num_agent):
dis_list[i][0] = discriminator(data_samples[i], sn=sn_d, name = 'dis_'+str(i))
gen_samples=tf.concat([gen_list[x] for x in connection_map[i]],axis=0)
dis_list[i][1] = discriminator(gen_samples, sn=sn_d, reuse=True, name = 'dis_'+str(i))
#loss functions
if args.objective == 'gan':
d_loss = sum([disc_loss(*dis) for dis in dis_list])
g_loss = sum([gen_loss(dis[-1]) for dis in dis_list])
g_vars = tf.global_variables(scope="gen")
d_vars = tf.global_variables(scope="dis")
else:
raise NotImplementedError
#optimizers
if True:#args.optimizer == 'adam':
g_train_opt = tf.train.AdamOptimizer(params['gen_learning_rate'],params['beta1'],params['beta2'])
d_train_opt = tf.train.AdamOptimizer(params['disc_learning_rate'],params['beta1'],params['beta2'])
d_train_op = d_train_opt.minimize(d_loss, var_list=d_vars)
g_train_op = g_train_opt.minimize(g_loss, var_list=g_vars)
else:
raise NotImplementedError
sess = tf.InteractiveSession()
#init input pipelines first
sess.run(tf.global_variables_initializer())
err_list = []
for i in tqdm(range(args.max_iter+1),disable=True):
for _ in range(args.n_dis):
f, _ = sess.run([d_loss, d_train_op],{is_training_pl:True})
sess.run([g_train_op],{is_training_pl:True})
if ((i) % args.viz_every == 0):
print('step: ',i)
x=[sess.run(g,{is_training_pl:False}) for g in gen_list]
print(x[0].shape)
#fig, axarr = plt.subplots(1,2, sharex='col', sharey='row', figsize=(12,4))
fig = plt.figure(1, figsize=(4, 4))
[plt.plot(v[:,0],v[:,1],'.',alpha=0.5) for v in x]
plt.axis([-1.5,1.5,-1.5,1.5])
plt.show()
mean_region = [np.mean(region,axis=0) for region in x]
err_list.append(np.mean([np.mean(np.abs(xx-yy)) for xx,yy in zip(list_loc,mean_region)]))
#raise ValueError('bad things!')
fig.savefig("/home/***/Pictures/multi_agent_gan/gaussian_toy.pdf", bbox_inches='tight')
'''
print('step: ',i)
x=[sess.run(g,{is_training_pl:False}) for g in gen_list]
y=[sess.run(d) for d in data_samples]
labels = ['p_{data_1}','p_{data_2}','p_{data_3}']
print(x[0].shape)
fig, axarr = plt.subplots(1,3, sharex='col', sharey='row', figsize=(14,4))
for i,v in enumerate(y):
axarr[i].plot(v[:,0],v[:,1],'.',alpha=0.5)
axarr[i].set_xlim([-1.5,1.5])
axarr[i].set_ylim([-1.5,1.5])
axarr[i].set_xlabel(r'$'+labels[i]+'$',fontsize=20)
axarr.show()
rect_o = patches.Rectangle((-1.42,-0.5),2.80,1.9,linewidth=2,edgecolor='orange',facecolor='none')
rect_g = patches.Rectangle((-0.5,-1.38),1.92,1.92,linewidth=2,edgecolor='g',facecolor='none')
rect_b = patches.Rectangle((-1.38,-1.42),1.92,1.88,linewidth=2,edgecolor='b',facecolor='none')
axarr[0].add_patch(rect_o)
axarr[0].add_patch(rect_g)
axarr[0].add_patch(rect_b)
if ((i) % args.viz_every == 0):
print('step: ',i)
x=[sess.run(g,{is_training_pl:False}) for g in gen_list]
y=[sess.run(d) for d in data_samples]
print(x[0].shape)
fig, axarr = plt.subplots(1,2, sharex='col', sharey='row', figsize=(12,4))
[axarr[1].plot(v[:,0],v[:,1],'.',alpha=0.5) for v in x]
axarr[1].set_xlim([-1.5,1.5]);axarr[0].set_ylim([-1.5,1.5])
[axarr[0].plot(v[:,0],v[:,1],'.',alpha=0.5) for v in y]
axarr[0].set_xlim([-1.5,1.5]);axarr[1].set_ylim([-1.5,1.5])
axarr.show()
mean_region = [np.mean(region,axis=0) for region in x]
err_list.append(np.mean([np.mean(np.abs(xx-yy)) for xx,yy in zip(list_loc,mean_region)]))
'''