-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathGRU.py
244 lines (191 loc) · 8.37 KB
/
GRU.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
import numpy as np
def sigmoid(x):
return np.power(1+np.exp(-x), -1)
def dsigmoid(x):
t=sigmoid(x)
return (1-t)*t
def tanh(x):
return np.tanh(x)
def dtanh(x):
return 1-np.square(np.tanh(x))
def softmax(x):
xexp = np.exp(x)
esum = np.sum(xexp)
return xexp/esum
import pickle
class gru:
def __init__(self, i_size, h_size, o_size, optimize='rmsprop', wb=None):
self.i_size = i_size
self.h_size = h_size
self.o_size = o_size
self.optimize = optimize
# self.names = {'ur':0,'wr':1, 'uz':2, 'wz':3, 'u_h':4, 'w_h':5, 'wo':6}
self.names = {0:'ur',1:'wr', 2:'uz', 3:'wz', 4:'u_h', 5:'w_h', 6:'wo'}
if wb:
self.w, self.b = self.load_weights(wb)
else:
self.w={}
self.b={}
# reset weights
self.w['ur'] = np.random.normal(0,0.01,(h_size, i_size))
self.b['r'] = np.zeros((h_size, 1))
self.w['wr'] = np.random.normal(0,0.01,(h_size, h_size))
# update weights
self.w['uz'] = np.random.normal(0,0.01,(h_size, i_size))
self.b['z'] = np.zeros((h_size, 1))
self.w['wz'] = np.random.normal(0,0.01,(h_size, h_size))
# _h weights
self.w['u_h'] = np.random.normal(0,0.01,(h_size, i_size))
self.b['_h'] = np.zeros((h_size, 1))
self.w['w_h'] = np.random.normal(0,0.01,(h_size, h_size))
# out weight
self.w['wo'] = np.random.normal(0,0.01,(o_size, h_size))
self.b['o'] = np.zeros((o_size, 1))
if optimize == 'rmsprop' or optimize == 'adam':
self.m={}
self.m['ur'] = np.zeros((h_size, i_size))
self.m['wr'] = np.zeros((h_size, h_size))
self.m['uz'] = np.zeros((h_size, i_size))
self.m['wz'] = np.zeros((h_size, h_size))
self.m['u_h'] = np.zeros((h_size, i_size))
self.m['w_h'] = np.zeros((h_size, h_size))
self.m['wo'] = np.zeros((o_size, h_size))
if optimize == 'adam':
self.v={}
self.v['ur'] = np.zeros((h_size, i_size))
self.v['wr'] = np.zeros((h_size, h_size))
self.v['uz'] = np.zeros((h_size, i_size))
self.v['wz'] = np.zeros((h_size, h_size))
self.v['u_h'] = np.zeros((h_size, i_size))
self.v['w_h'] = np.zeros((h_size, h_size))
self.v['wo'] = np.zeros((o_size, h_size))
self.weight_update = adam
elif optimize == 'rmsprop':
self.weight_update = rmsprop
def forward_pass(self, inputs):
self.inputs = inputs
self.n_inp = len(inputs)
self.vr = []; self.vz = []; self.v_h = []; self.vo = [];
self.r=[]; self.z=[]; self._h=[]; self.h={}; self.o = []
self.h[-1] = np.zeros((self.h_size,1))
for i in range(self.n_inp):
# calculating reset gate value
self.vr.append(np.dot(self.w['ur'],inputs[i]) + np.dot(self.w['wr'], self.h[i-1]) + self.b['r'])
self.r.append(sigmoid(self.vr[i]))
# calculation update gate value
self.vz.append(np.dot(self.w['uz'],inputs[i]) + np.dot(self.w['wz'], self.h[i-1]) + self.b['z'])
self.z.append(sigmoid(self.vz[i]))
# applying reset gate value
self.v_h.append(np.dot(self.w['u_h'], inputs[i]) + np.dot(self.w['w_h'], np.multiply(self.h[i - 1], self.r[i])) + + self.b['_h'])
self._h.append(tanh(self.v_h[i]))
# applying update gate value
self.h[i] = np.multiply(self.z[i], self.h[i - 1]) + np.multiply(1-self.z[i], self._h[i])
# # calculating output
# self.vo.append(np.dot(self.w['wo'], self.h[i]))
# self.o.append(sigmoid(self.vo[i]))
# calculating output
self.vo.append(np.dot(self.w['wo'], self.h[i]) + self.b['o'])
self.o.append(softmax(self.vo[i]))
return self.o
def backward_pass(self, t):
# error calculation
e = self.error(t)
# dw variables
dw={}
db= {}
dw['uz'] = np.zeros((self.h_size, self.i_size))
db['z'] = np.zeros((self.h_size, 1))
dw['wz'] = np.zeros((self.h_size, self.h_size))
# reset dw
dw['ur'] = np.zeros((self.h_size, self.i_size))
db['r'] = np.zeros((self.h_size, 1))
dw['wr'] = np.zeros((self.h_size, self.h_size))
# _h dw
dw['u_h'] = np.zeros((self.h_size, self.i_size))
db['_h'] = np.zeros((self.h_size, 1))
dw['w_h'] = np.zeros((self.h_size, self.h_size))
# hidden-2-output dw
dw['wo'] = np.zeros((self.o_size, self.h_size))
db['o'] = np.zeros((self.o_size, 1))
dh = 0.0
for i in reversed(range(self.n_inp)):
# gradient at output layer
go = self.o[i] - t[i]
# hidden to outpur weight's dw
dw['wo'] += np.dot(go, self.h[i].T)
db['o'] += go
# gradient at top hidden layer
dh += np.dot(self.w['wo'].T, go)
dz = (self.h[i-1] - self._h[i]) * dh
dz__ = self.z[i] * dh
# dz_ = dsigmoid(self.vz[i]) * dz
dz_ = np.multiply((1.0 - self.z[i]), self.z[i]) * dz
d_h = (1-self.z[i]) * dh
d_h_ = (1- np.square(self._h[i])) * d_h
temp = np.dot(self.w['w_h'].T, d_h_)
dr = self.h[i-1] * temp
dr_ = np.multiply((1.0 - self.r[i]), self.r[i]) * dr
dr__ = self.r[i] * temp
# calculating reset dw
dw['ur'] += np.dot(dr_ , self.inputs[i].T)
db['r'] += dr_
dw['wr'] += np.dot(dr_ , self.h[i-1].T)
# db['wr'] += dr_
# calculating update dw
dw['uz'] += np.dot(dz_, self.inputs[i].T)
db['z'] += dz_
dw['wz'] += np.dot(dz_, self.h[i-1].T)
# db['wz'] += dz_
# calculating _h dw
dw['u_h'] += np.dot(d_h_, self.inputs[i].T)
db['_h'] += d_h_
dw['w_h'] += np.dot(d_h_, np.multiply(self.r[i], self.h[i-1]).T)
# db['w_h'] += d_h_
dh = np.dot(self.w['wr'].T, dr_) + np.dot(self.w['wz'].T, dz_) + dz__ + dr__
return dw, db, np.linalg.norm(e)
def error(self, t):
loss = np.sum(t * np.log(self.o))
return -loss
def save_model(self, fname):
pickle.dump([self.w, self.b], open(fname, 'wb'))
def load_model(self, fname):
return pickle.load(open(fname, 'rb'))
# def reset_m(self):
# if self.optimize == 'rmsprop' or self.optimize == 'adam':
# self.m = {}
# self.m['ur'] = np.zeros((self.h_size, self.i_size))
# self.m['wr'] = np.zeros((self.h_size, self.h_size))
# self.m['uz'] = np.zeros((self.h_size, self.i_size))
# self.m['wz'] = np.zeros((self.h_size, self.h_size))
# self.m['u_h'] = np.zeros((self.h_size, self.i_size))
# self.m['w_h'] = np.zeros((self.h_size, self.h_size))
# self.m['wo'] = np.zeros((self.o_size, self.h_size))
#
# if self.optimize == 'adam':
# self.v = {}
# self.v['ur'] = np.zeros((self.h_size, self.i_size))
# self.v['wr'] = np.zeros((self.h_size, self.h_size))
# self.v['uz'] = np.zeros((self.h_size, self.i_size))
# self.v['wz'] = np.zeros((self.h_size, self.h_size))
# self.v['u_h'] = np.zeros((self.h_size, self.i_size))
# self.v['w_h'] = np.zeros((self.h_size, self.h_size))
# self.v['wo'] = np.zeros((self.o_size, self.h_size))
# return
def rmsprop(self, dw, db, neta, b1=.9, b2=.0, e=1e-8):
for wpi, g in dw.items():
self.m[wpi] = b1 * self.m[wpi] + (1 - b1) * np.square(g)
self.w[wpi] -= neta * np.divide(g, (np.sqrt(self.m[wpi]) + e))
for wpi in db:
self.b[wpi] -= neta * db[wpi]
return
def adam(self, dw, db, neta, b1=0.9, b2=0.99, e=1e-8):
for wpi, g in dw.items():
self.m[wpi] = (b1 * self.m[wpi]) + ((1. - b1) * g)
self.v[wpi] = (b2 * self.v[wpi]) + ((1. - b2) * np.square(g))
m_h = self.m[wpi]/(1.-b1)
v_h = self.v[wpi]/(1.-b2)
# w[wpi] -= neta * (m_h/(np.sqrt(v_h) + e) + regu * w[wpi])
self.w[wpi] -= neta * m_h/(np.sqrt(v_h) + e)
for wpi in db:
self.b[wpi] -= neta * db[wpi]
return