-
Notifications
You must be signed in to change notification settings - Fork 1
/
one_shot_recognition.py
265 lines (187 loc) · 8.63 KB
/
one_shot_recognition.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
260
261
262
263
264
265
from sklearn import linear_model
import pickle
import numpy as np
from six.moves import urllib
from six.moves import xrange # pylint: disable=redefined-builtin
from numpy import random
import data_prepare as dp
import math
import tensorflow as tf
import os
import base64
import time
from numpy import linalg as LA
import os.path
import argparse
from numpy import *
import numpy as np
from os import path
"""updated @ 2017-8-31"""
## Define Random Initilization
def xavier_init(size,std=1):
in_dim = size[0]
xavier_stddev = 1. / tf.sqrt(in_dim / 2.)
return tf.random_normal(shape=size, stddev=xavier_stddev*std)
def generator(z, c, way_Gw, opt_nml, Wz, Wc, alpha = 65):
# inputs = tf.concat(axis=1, values=[z, c])
if way_Gw==1:
G_prob = tf.matmul(z, tf.diag(Wz))+tf.matmul(c, Wc)
elif way_Gw==0:
G_prob = tf.matmul(z, Wz)+tf.matmul(c, Wc)
if opt_nml==1:
G_prob = G_prob/tf.norm(G_prob)*alpha
G_prob = tf.nn.relu(G_prob)
return G_prob
def discriminator(X, Wg, Ws_bs, Ws_nl):
Og = tf.nn.sigmoid(tf.matmul(X, Wg))
Os = tf.matmul(X, tf.concat([Ws_bs, Ws_nl],1))
return Og, Os
def sample_z(m, n):
return np.random.uniform(-1., 1., size=[m, n])
def cross_entropy(logit, y):
return -tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=logit, labels=y))
def w_gp_ac_gan(cmd):
lambda2 = cmd.w ## weight between Wc*c and Wz*z
"""Different Ways to Initialize the weights for Generator"""
G_w = ["random+"+str(lambda2)+"weight", "random vector"+str(lambda2)+"weight"]
Opt = ["Momentum", "Adam", "RMS", "SGD"]
way_Gw = cmd.i ## 0: Wz is matrix; 1: Wz is vector
way_opt = 0 ## choose optimizor
wgan = cmd.g ## 0: without WGAN; 1: with WGAN
wz_s = cmd.s ## scale of Wz
opt_nml = cmd.n ## 0: without Normalization; 1: with normalization
## Parameter
z_dim = 512
X_dim = 512
y_dim = 21000
SEED = 66478
BATCH_SIZE = 512
num_epochs = 20
LAMBDA = 1e1
num_classes = 21000
## prepare index and shuffle file
trSf = dp.readfile("/data/21K/TrainValData_300Max-ps.train.r100.shuffle")
vlSf = dp.readfile("/data/21K/TrainValData_300Max-ps.val.base.shuffle")
tsSf = dp.readfile("/data/21K/TrainValData_300Max-ps.val.lowshot.shuffle")
feaIdx = dp.readfile("/data/21K/TrainValData_300Max-ps.LineNumber.resnet34-b.0.pool5.lineidx")
lblIdx = dp.readfile("/data/21K/TrainValData_300Max-ps.LineNumber.resnet34-b.0.pool5.label.lineidx")
tsvFea = "/data/21K/TrainValData_300Max-ps.LineNumber.resnet34-b.0.pool5.tsv"
tsvLbl = "/data/21K/TrainValData_300Max-ps.LineNumber.resnet34-b.0.pool5.label.tsv"
train_size = len(trSf)
## Define Input Variables
X = tf.placeholder(tf.float32, shape=[None, X_dim])
y = tf.placeholder(tf.float32, shape=[None, y_dim])
z = tf.placeholder(tf.float32, shape=[None, z_dim])
## Initialize Weights of Class center
weights = np.array(np.load('/data/weights/Wg_class_center.npy'), dtype=np.float32)
Wc = tf.Variable(lambda2*weights, name="G2")
if way_Gw ==0:
Wz = tf.Variable(xavier_init([z_dim, X_dim], wz_s), name="G1")
elif way_Gw==1:
Wz = tf.Variable(wz_s*xavier_init([X_dim],1), name="G1")
"""Initialization for Discriminator: Binary Discriminatot & Softmax Classifier"""
Wd = tf.Variable(xavier_init([X_dim, 1],1), name="Dg")
W_sf = np.load('/data/weights/Wc_iter100000.npy').transpose()
Ws_nl = tf.Variable(W_sf[:,20000:21000], name="Dc1")
Ws_bs = tf.Variable(W_sf[:,0:20000], name="Dc2")
theta_G = [Wz, Wc]
theta_D = [Wd, Ws_nl]
G_sample = generator(z, y, way_Gw, opt_nml, Wz, Wc)
D_real, C_real = discriminator(X, Wd, Ws_bs, Ws_nl)
D_fake, C_fake = discriminator(G_sample, Wd, Ws_bs, Ws_nl)
# Cross entropy aux loss
Cr_loss = cross_entropy(C_real, y)
Cf_loss = cross_entropy(C_fake, y)
C_loss = Cr_loss + Cf_loss
# GAN D loss
D_loss = tf.reduce_mean(tf.log(D_real)) + tf.reduce_mean(tf.log(1. - D_fake))
## See if use W-GAN
if wgan==1:
## Gradient Panelty
alpha = tf.random_uniform(shape=[BATCH_SIZE,1], minval=0., maxval=1.)
differences = G_sample - X
interpolates = X + (alpha*differences)
Ig, _ = discriminator(interpolates, Wd, Ws_bs, Ws_nl)
gradients = tf.gradients(Ig, [interpolates])[0]
slopes = tf.sqrt(tf.reduce_sum(tf.square(gradients), reduction_indices=[1]))
gradient_penalty = tf.reduce_mean((slopes-1.)**2)
else:
gradient_penalty = 0
weight_decay = 0.0001
DC_loss = - (D_loss + C_loss) + LAMBDA*gradient_penalty+ weight_decay*(tf.nn.l2_loss(Ws_nl)+tf.nn.l2_loss(Wd))+tf.nn.l2_loss(Ws_bs-W_sf[:,0:20000])
# GAN's G loss
G_loss = tf.reduce_mean(tf.log(D_fake))
GC_loss = - (G_loss + C_loss)
## Prediction Accuracy
P_real = tf.nn.softmax(C_real)
correct_prediction = tf.equal(tf.argmax(C_real,1), tf.argmax(y,1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
correct_prediction_f = tf.equal(tf.argmax(C_fake,1), tf.argmax(y,1))
accuracy_f = tf.reduce_mean(tf.cast(correct_prediction_f, tf.float32))
# Gradient Descent
## Choose learning rate
learning_rate = cmd.lr
scale = 100
if way_opt==0:
D_solver = tf.train.MomentumOptimizer(learning_rate, 0.9).minimize(DC_loss, var_list=theta_D)
G_solver = tf.train.MomentumOptimizer(learning_rate*scale, 0.9).minimize(GC_loss, var_list=theta_G)
elif way_opt==1:
D_solver = (tf.train.AdamOptimizer(learning_rate=learning_rate).minimize(DC_loss, var_list=theta_D))
G_solver = (tf.train.AdamOptimizer(learning_rate=learning_rate*scale).minimize(GC_loss, var_list=theta_G))
elif way_opt==2:
D_solver = (tf.train.RMSPropOptimizer(learning_rate=learning_rate).minimize(DC_loss, var_list=theta_D))
G_solver = (tf.train.RMSPropOptimizer(learning_rate=learning_rate*scale).minimize(GC_loss, var_list=theta_G))
elif way_opt==3:
D_solver = tf.train.GradientDescentOptimizer(learning_rate=learning_rate).minimize(DC_loss, var_list=theta_D)
G_solver = tf.train.GradientDescentOptimizer(learning_rate=learning_rate*scale).minimize(GC_loss, var_list=theta_G)
sess = tf.Session()
# Initializing the variables
init = tf.global_variables_initializer()
## Load data file
pos = 0
fea_tsv = open(tsvFea,"rb")
lbl_tsv = open(tsvLbl,"rb")
## Load validation data and test data
val_data, val_labels, _, _ = dp.nextbatch(pos,vlSf,len(vlSf),fea_tsv,lbl_tsv,feaIdx,lblIdx)
test_data, test_labels, _, _ = dp.nextbatch(pos,tsSf,len(tsSf),fea_tsv,lbl_tsv,feaIdx,lblIdx)
if opt_nml==1:
val_data = dp.normalization(val_data)
test_data = dp.normalization(test_data)
merged = tf.summary.merge_all()
fline = G_w[way_Gw], Opt[way_opt], learning_rate, "WGAN:", wgan, "G Scale:", scale
print fline
with tf.Session() as sess:
sess.run(init)
for step in xrange(int(num_epochs * train_size) // BATCH_SIZE):
## Generate Batches
if pos+BATCH_SIZE >= train_size:
pos = 0
X_mb, y_mb, _ , _ = dp.nextbatch(pos,trSf,BATCH_SIZE,fea_tsv,lbl_tsv,feaIdx,lblIdx)
if opt_nml==1:
X_mb = dp.normalization(X_mb)
pos += BATCH_SIZE
z_mb = sample_z(BATCH_SIZE, z_dim)
## Optimize Disriminator
if step%2==0:
_ = sess.run(D_solver, feed_dict={X: X_mb, y: y_mb, z: z_mb})
_ = sess.run(G_solver, feed_dict={X: X_mb, y: y_mb, z: z_mb})
if step % 2000 == 0:
val_acc, val_pre = sess.run([accuracy, P_real], feed_dict={X: val_data, y: val_labels})
ts_acc, ts_pre = sess.run([accuracy, P_real], feed_dict={X: test_data, y: test_labels})
fline2 = 'Iter_acc: {}; Validation Accuracy: {}; Test Accuracy: {}'.format(step, val_acc, ts_acc)
print fline2
fea_tsv.close()
lbl_tsv.close()
def parse_args():
parser = argparse.ArgumentParser()
parser.add_argument('--n', required=True, type=int, help='normalization')
parser.add_argument('--g', required=True, type=int, help='wgan')
parser.add_argument('--s', required=True, type=int, help='Wz_scale')
parser.add_argument('--w', required=True, type=np.float32, help='learning rate')
parser.add_argument('--lr', required=True, type=np.float32, help='weights')
parser.add_argument('--i', required=True, type=int, help='inilization')
return parser.parse_args()
if __name__ == "__main__":
cmd = parse_args()
w_gp_ac_gan(cmd)
## python one_shot_recognition.py --n 1 --g 1 --lr 1e-4 --s 200 --w 0.7 --i 1