SSOM.py
4.85 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
import tensorflow as tf
import numpy as np
from .SOM import SOM, load_SOM_base
from .tf_util import tf_base
from .util import check_dir
########################
# Supervised SOM layer #
########################
def init_SLSOM(path,som):
parameter = {}
with open(path+"parameters.txt") as f:
for line in f:
splitted = line.split("\t")
parameter[splitted[0]] = splitted[1]
tmp = SLSOM(som=som,
nb_label=int(parameter["nb_label"]),
k = int(parameter["k"]),
loss_type = parameter["loss_type"]
)
return tmp
class SLSOM(object):
def __init__(self,som,nb_label,loss_type='cross_entropy',verbose=True):
self.tf_object = som.tf_object
self.ulen = som.ulen
self.nb_label = nb_label
self.som = som
self.loss_type = loss_type
with self.tf_object.graph.as_default():
self.W = tf.Variable(tf.random_normal([self.ulen,self.nb_label],dtype=tf.float64))
self.W_loader = tf.placeholder(tf.float64,shape=[self.ulen,self.nb_label])
self.load_W = self.W.assign(self.W_loader)
self.biases = tf.Variable(tf.random_normal([self.nb_label],dtype=tf.float64))
self.biases_loader = tf.placeholder(tf.float64,shape=[self.nb_label])
self.load_biases = self.biases.assign(self.biases_loader)
self.it_max = tf.placeholder(tf.int32)
self.it = tf.Variable(0,dtype=tf.int32)
self.update_it = self.it.assign_add(1)
self.data = self.som.sim2units(self.som.data2pred)
self.datapred = tf.one_hot(
self.som.bmu_finder(self.som.data2pred,self.som.units),
self.som.ulen,
dtype=tf.float64
)
self.data_size = tf.placeholder(tf.int32,shape=[1])
self.lambda_penality = tf.placeholder(tf.float64,shape=[1])
self.max_it = tf.placeholder(tf.int32,shape=[1])
self.labels = tf.placeholder(tf.int32,shape=[None])
self.train_op = self.minimize_loss()
self.proba_data_op = self.proba_class_op()
self.prediction = self.prediction_op(self.proba_data_op)
self.update_it_som = self.som.it.assign_add(1)
def learning_rate(self,it):
return 1.0-tf.cast(self.it,tf.float64)/(tf.cast(self.it_max,tf.float64))
def save(self,path):
W = self.get_W()
biases = self.get_biases()
np.savetxt(path+"W.txt",np.array(W))
np.savetxt(path+"biases.txt",np.array(biases))
check_dir(path)
with open(path+"parameters.txt","w") as f:
towrite = ""
towrite += "nb_label\t"+str(self.nb_label)+"\n"
towrite += "k\t"+str(self.k)+"\n"
towrite += "loss_type\t"+str(self.loss_type)+"\n"
f.write(towrite)
def load(self,path):
W = np.loadtxt(path+"W.txt")
biases = np.loadtxt(path+"biases.txt")
self.tf_object.run([self.load_W,self.load_biases],feed_dict={self.W_loader: W, self.biases_loader: biases})
def minimize_loss(self):
x = self.data
dist = self.som.dist2units(self.som.data2pred)
bmus = tf.argmin(dist,1)
dist_bmu = self.som.dist_bmus_op(bmus)
neighbour = self.som.R(dist_bmu,tf.cast(self.learning_rate(self.it),tf.float64)*max(self.som.dim)/2.0)
x = x*tf.transpose(neighbour)
y = tf.matmul(
x,
self.W
) + self.biases
y_ = tf.one_hot(
self.labels,
self.nb_label,
dtype=tf.float64
)
if self.loss_type == 'cross_entropy':
loss = tf.reduce_mean(tf.nn.sparse_softmax_cross_entropy_with_logits(logits=y, labels=self.labels))
else :
loss = 0.5*tf.reduce_mean(tf.pow(tf.nn.softmax(y) - y_,2.0))
regularizer = tf.contrib.layers.l2_regularizer(self.lambda_penality)
penality = regularizer(self.W)
optimizer = tf.train.GradientDescentOptimizer(0.3*self.learning_rate(self.it))
optimizer2 = tf.train.GradientDescentOptimizer(self.learning_rate(self.it))
loss2 = tf.add(loss,penality)
applied = optimizer.minimize(loss2,var_list=[self.W,self.biases])
applied2 = optimizer2.minimize(loss2,var_list=[self.som.units])
return (applied,applied2)
def proba_class_op(self):
x = self.datapred
y = tf.matmul(
x,
self.W
) + self.biases
return tf.nn.softmax(y)
def prediction_op(self,y):
return tf.argmax(y,1)
def get_W(self):
return self.tf_object.run(self.W)
def get_biases(self):
return self.tf_object.run(self.biases)
def train(self,data,labels,max_it=2000,batch_size=10, penality=0.001):
it = np.array([max_it])
nb_data = data.shape[0]
pen = np.array([penality])
data2 = data
for i in range(max_it):
if self.verbose:
print("It SLSOM: "+str(i))
idx = np.random.randint(nb_data,size=batch_size)
self.tf_object.run(self.train_op,
feed_dict={
self.som.data2pred:data2[idx,:],
self.data_size:np.array([batch_size]),
self.labels:labels[idx],
self.lambda_penality:pen,
self.it_max:max_it
})
self.tf_object.run(self.update_it)
def predict(self,data):
pred,proba = self.tf_object.run([self.prediction,self.proba_data_op],
feed_dict={
self.som.data2pred:data
})
return (pred,proba)
def proba_data(self,data):
prob = self.tf_object.run(self.proba_data_op,
feed_dict={
self.data:data
})
return prob