forked from atulsinha007/Domain-adaptation
-
Notifications
You must be signed in to change notification settings - Fork 0
/
learn_transformation.py
178 lines (143 loc) · 4.96 KB
/
learn_transformation.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
import random
import numpy as np
from deap import base
from deap import creator
from deap import tools
import pickle
import copy
import time
NGEN = 1000
pop_size = 100
cxpb = .8
m_fac = .2
m_prob = .2
np.random.seed(0)
rng = np.random
pstri = "./pickle_jar/"
fs = open(pstri + "tar_data_with_dic.pickle", "rb")
tup_t = pickle.load(fs)
fs.close()
target_arr, target_label, target_dic = tup_t
dum_arr = target_label.reshape((target_label.shape[0], 1))
clumped_arr = np.concatenate((target_arr, dum_arr), axis=1)
# print(dic)
numlis = np.arange(clumped_arr.shape[0])
rng.shuffle(numlis)
clumped_arr = clumped_arr[numlis]
# clumped_arr = clumped_arr[ numlis ]
clumped_target = clumped_arr[:]
ann = int((3/4)*clumped_target.shape[0])
print(ann)
tup_t = (target_rest_arr, target_rest_label), (target_test_arr, target_rest_label) = (clumped_target[:ann, :-1], clumped_target[:ann, -1:]), (clumped_target[ann:, :-1], clumped_target[ann:, -1:])
#print(tup_t)
fs = open(pstri + "tar_tup.pickle", "wb")
pickle.dump(tup_t, fs)
fs.close()
target_dim = target_rest_arr.shape[1]
fs = open(pstri + "src_data_with_dic.pickle", "rb")
tup_s = pickle.load(fs)
fs.close()
source_arr, source_label, source_dic = tup_s
dum_arr = source_label.reshape((source_label.shape[0], 1))
clumped_arr = np.concatenate((source_arr, dum_arr), axis=1)
# print(dic)
numlis = np.arange(clumped_arr.shape[0])
rng.shuffle(numlis)
clumped_arr = clumped_arr[numlis]
# clumped_arr = clumped_arr[ numlis ]
clumped_source = clumped_arr[:]
#ann = (3//4)*clumped_source.shape[0]
ann = clumped_source.shape[0]
tup_s = (source_rest_arr, source_rest_label), (source_test_arr, source_rest_label) = (clumped_source[:ann, :-1], clumped_source[:ann, -1:]), (clumped_source[ann:, :-1], clumped_source[ann:, -1:])
fs = open(pstri + "src_tup.pickle", "wb")
pickle.dump(tup_s, fs)
fs.close()
source_dim = source_rest_arr.shape[1]
def generate_pop(pop_size, source_dim, target_dim, rng = np.random):
pop_lis = []
for individual in range(pop_size):
W = rng.random((source_dim, target_dim))
pop_lis.append(W)
return pop_lis
def dist(transformed_target, source_instance):
return np.sqrt(np.sum((transformed_target - source_instance)**2))
def closeness_cost(W):
sumi = 0
#print( source_dic)
for class_num in target_dic:
for target_instance in target_rest_arr[ target_dic[class_num][0]: target_dic[class_num][1]+1 ]:
min_dist = np.inf
target_instance = np.reshape(target_instance, (target_instance.shape[0], 1))
transformed_target = np.dot(W, target_instance)
transformed_target = np.ravel( transformed_target)
for source_instance in source_rest_arr[source_dic[class_num][0]: source_dic[class_num][1]+1 ]:
#print(transformed_target, source_instance)
min_dist = min( min_dist, dist(transformed_target, source_instance))
sumi += min_dist
return sumi
def calc_fitness(population):
cost_lis = []
for indi in population:
cost_lis.append(-closeness_cost(indi))
return cost_lis
def myCrossover(arr1, arr2, cxpb, rng = np.random):
for row in range(arr1.shape[0]):
for col in range(arr1.shape[1]):
if rng.random() < cxpb:
alpha = rng.random()
temp = copy.deepcopy(arr1[row][col])
arr1[row][col] = alpha*arr1[row][col] + (1-alpha)*arr2[row][col]
arr2[row][col] = alpha*arr2[row][col] + (1-alpha)*temp
return arr1, arr2
def myMutate(arr, m_prob, m_fac, rng = np.random):
arr = arr + rng.random(arr.shape)*m_fac
return arr
for row in range(arr1.shape[0]):
index = rng.randint(0, row)
if rng.random() < m_prob:
arr[row][index] += rng.uniform(-1,1)*m_fac
def tournament_selection(population, fitness_arr, rng = np.random):
a = rng.randint(0,len(population)-1)
b = rng.randint(0,len(population)-1)
parent1 = population[a]
parent2 = population[b]
if fitness_arr[a] < fitness_arr[b]:
parentA = parent1
else:
parentA = parent2
c = rng.randint(0,len(population)-1)
d = rng.randint(0,len(population)-1)
parent3 = population[c]
parent4 = population[d]
if fitness_arr[c] < fitness_arr[d]:
parentB = parent3
else:
parentB = parent4
return parentA, parentB
def main(pop_size):
global source_dim, target_dim, m_fac, m_prob, cxpb
population = generate_pop(pop_size, source_dim, target_dim)
for i in range(NGEN):
# print(population)
print(i)
fitness_arr = []
fitness_arr = calc_fitness(population)
#print(fitness_arr)
minn = np.inf
if np.amin(fitness_arr) < minn:
minn = np.amin(fitness_arr)
ind_min = np.argmin(fitness_arr)
print('minimum in this generation is '+ str(np.amin(fitness_arr)), "at", ind_min, "th index")
mating_pool = population
for j in range(int(pop_size/2)):
parent1, parent2 = tournament_selection(population, fitness_arr)
child1, child2 = myCrossover(parent1, parent2, cxpb)
child1 = myMutate(child1, m_fac, m_prob)
child2 = myMutate(child2, m_fac, m_prob)
ind_min = np.argmin(fitness_arr)
print(population[ ind_min ])
fs= open("./pickle_jar/dublue.pickle", "wb")
pickle.dump(population[ind_min], fs)
fs.close()
if __name__ == "__main__":
main(pop_size)