-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathlstm.py
More file actions
317 lines (206 loc) · 9.79 KB
/
lstm.py
File metadata and controls
317 lines (206 loc) · 9.79 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
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
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
import numpy as np
from utils import *
import collections # for dL
class depTreeLSTMModel:
def __init__(self, wvecDim):
self.wvecDim = wvecDim
self.defaultVec = lambda : np.zeros((wvecDim,))
def initialParams(self, word2vecs, rng):
# Word vectors
self.numWords = word2vecs.shape[1]
self.L = word2vecs[:self.wvecDim, :]
self.Wi = rng.uniform(
low=-np.sqrt(6. / (self.wvecDim + self.wvecDim)),
high=np.sqrt(6. / (self.wvecDim + self.wvecDim)),
size=(self.wvecDim, self.wvecDim)
)
self.Ui = rng.uniform(
low=-np.sqrt(6. / (self.wvecDim + self.wvecDim)),
high=np.sqrt(6. / (self.wvecDim + self.wvecDim)),
size=(self.wvecDim, self.wvecDim)
)
self.bi = np.zeros((self.wvecDim))
self.Wf = rng.uniform(
low=-np.sqrt(6. / (self.wvecDim + self.wvecDim)),
high=np.sqrt(6. / (self.wvecDim + self.wvecDim)),
size=(self.wvecDim, self.wvecDim)
)
self.Uf = rng.uniform(
low=-np.sqrt(6. / (self.wvecDim + self.wvecDim)),
high=np.sqrt(6. / (self.wvecDim + self.wvecDim)),
size=(self.wvecDim, self.wvecDim)
)
self.bf = np.zeros((self.wvecDim))
self.Wo = rng.uniform(
low=-np.sqrt(6. / (self.wvecDim + self.wvecDim)),
high=np.sqrt(6. / (self.wvecDim + self.wvecDim)),
size=(self.wvecDim, self.wvecDim)
)
self.Uo = rng.uniform(
low=-np.sqrt(6. / (self.wvecDim + self.wvecDim)),
high=np.sqrt(6. / (self.wvecDim + self.wvecDim)),
size=(self.wvecDim, self.wvecDim)
)
self.bo = np.zeros((self.wvecDim))
self.Wu = rng.uniform(
low=-np.sqrt(6. / (self.wvecDim + self.wvecDim)),
high=np.sqrt(6. / (self.wvecDim + self.wvecDim)),
size=(self.wvecDim, self.wvecDim)
)
self.Uu = rng.uniform(
low=-np.sqrt(6. / (self.wvecDim + self.wvecDim)),
high=np.sqrt(6. / (self.wvecDim + self.wvecDim)),
size=(self.wvecDim, self.wvecDim)
)
self.bu = np.zeros((self.wvecDim))
self.stack = [self.L, self.Wi, self.Ui, self.bi, self.Wf, self.Uf, self.bf,
self.Wo, self.Uo, self.bo, self.Wu, self.Uu, self.bu ]
self.dWi = np.empty((self.wvecDim, self.wvecDim))
self.dUi = np.empty((self.wvecDim, self.wvecDim))
self.dbi = np.empty(self.wvecDim)
self.dWf = np.empty((self.wvecDim, self.wvecDim))
self.dUf = np.empty((self.wvecDim, self.wvecDim))
self.dbf = np.empty(self.wvecDim)
self.dWo = np.empty((self.wvecDim, self.wvecDim))
self.dUo = np.empty((self.wvecDim, self.wvecDim))
self.dbo = np.empty(self.wvecDim)
self.dWu = np.empty((self.wvecDim, self.wvecDim))
self.dUu = np.empty((self.wvecDim, self.wvecDim))
self.dbu = np.empty(self.wvecDim)
def clearDerivativeSharedMemory(self):
self.dWi[:] = 0
self.dUi[:] = 0
self.dbi[:] = 0
self.dWf[:] = 0
self.dUf[:] = 0
self.dbf[:] = 0
self.dWo[:] = 0
self.dUo[:] = 0
self.dbo[:] = 0
self.dWu[:] = 0
self.dUu[:] = 0
self.dbu[:] = 0
self.dL = collections.defaultdict(self.defaultVec)
self.dstack = [self.dL, self.dWi, self.dUi, self.dbi, self.dWf, self.dUf, self.dbf,
self.dWo, self.dUo, self.dbo, self.dWu, self.dUu, self.dbu ]
def forwardProp(self, tree):
#because many training epoch.
tree.resetFinished()
to_do = []
to_do.append(tree.root)
while to_do:
curr = to_do.pop(0)
curr.vec = self.L[:, curr.index]
# node is leaf
if len(curr.kids) == 0:
# j is node id
x_j = curr.vec
# hj is zero
i_j = sigmoid( np.dot(self.Wi, x_j) + self.bi)
# due to k is zero
f_jk = sigmoid( np.dot(self.Wf, x_j) + self.bf)
o_j = sigmoid( np.dot(self.Wo, x_j) + self.bo)
u_j = np.tanh( np.dot(self.Wu, x_j) + self.bu)
c_j = i_j * u_j
curr.i_j = i_j
curr.u_j = u_j
curr.o_j = o_j
curr.c_j = c_j
curr.h_j= o_j * np.tanh(c_j)
curr.finished=True
else:
#check if all kids are finished
all_done = True
for index, rel in curr.kids:
node = tree.nodes[index]
if not node.finished:
to_do.append(node)
all_done = False
if all_done:
x_j = curr.vec
h_j_hat = np.zeros((self.wvecDim))
sum_f_jk_C_k = np.zeros((self.wvecDim))
for i, rel in curr.kids:
h_j_hat += tree.nodes[i].h_j
f_jk = sigmoid( np.dot(self.Wf, x_j) + np.dot(self.Uf, tree.nodes[i].h_j )+ self.bf)
sum_f_jk_C_k += f_jk * tree.nodes[i].c_j
i_j = sigmoid( np.dot(self.Wi, x_j) + np.dot(self.Ui, h_j_hat)+ self.bi)
o_j = sigmoid( np.dot(self.Wo, x_j) + np.dot(self.Uo, h_j_hat)+ self.bo)
u_j = np.tanh( np.dot(self.Wu, x_j) + np.dot(self.Uu, h_j_hat)+ self.bu)
c_j = i_j * u_j + sum_f_jk_C_k
curr.c_j = c_j
curr.i_j = i_j
curr.u_j = u_j
curr.o_j = o_j
curr.h_j= o_j * np.tanh(c_j)
curr.h_j_hat = h_j_hat
curr.finished = True
else:
to_do.append(curr)
return tree.root.h_j
def backProp(self, tree, deltas):
to_do = []
to_do.append(tree.root)
tree.root.deltas = deltas
while to_do:
curr = to_do.pop(0)
if len(curr.kids) == 0:
x_j = curr.vec
delta_h_j = curr.deltas
delta_o_j = delta_h_j * np.tanh(curr.c_j)
delta_o_j *= derivative_sigmoid(x_j)
self.dWo += np.outer(delta_o_j, x_j)
self.dbo += delta_o_j
#self.dL[:, curr.index] += np.dot(self.Wo, delta_o_j)
self.dL[curr.index] += np.dot(self.Wo, delta_o_j)
delta_c_j = curr.o_j * delta_h_j * derivative_tanh(curr.h_j)
delta_i_j = delta_c_j * curr.u_j
delta_i_j *= derivative_sigmoid(x_j)
self.dWi += np.outer(delta_i_j, x_j)
self.dbi += delta_i_j
#self.dL[:, curr.index] += np.dot(self.Wi, delta_i_j)
self.dL[curr.index] += np.dot(self.Wi, delta_i_j)
delta_u_j = delta_c_j * curr.i_j
delta_u_j *= derivative_sigmoid(x_j)
self.dWu += np.outer(delta_u_j, x_j)
self.dbu += delta_u_j
#self.dL[:, curr.index] += np.dot(self.Wu, delta_u_j)
self.dL[curr.index] += np.dot(self.Wu, delta_u_j)
else:
x_j = curr.vec
delta_h_j = curr.deltas
delta_o_j = delta_h_j * np.tanh(curr.c_j)
delta_o_j_1 = delta_o_j * derivative_sigmoid(x_j)
delta_o_j_2 = delta_o_j * derivative_sigmoid(curr.h_j_hat)
self.dWo += np.outer(delta_o_j_1, x_j)
self.dUo += np.outer(delta_o_j_2, curr.h_j_hat)
self.dbo += delta_o_j_1
#self.dL[:, curr.index] += np.dot(self.Wo, delta_o_j_1)
self.dL[curr.index] += np.dot(self.Wo, delta_o_j_1)
delta_c_j = curr.o_j * delta_h_j * derivative_tanh(curr.h_j)
delta_i_j = delta_c_j
delta_i_j_1 = delta_i_j * derivative_sigmoid(x_j)
delta_i_j_2 = delta_i_j * derivative_sigmoid(curr.h_j_hat)
self.dWi += np.outer(delta_i_j_1, x_j)
self.dUi += np.outer(delta_i_j_2, curr.h_j_hat)
self.dbi += delta_i_j_1
#self.dL[:, curr.index] += np.dot(self.Wi, delta_i_j_1)
self.dL[curr.index] += np.dot(self.Wi, delta_i_j_1)
delta_u_j = delta_c_j * curr.i_j
delta_u_j_1 = delta_u_j * derivative_sigmoid(x_j)
delta_u_j_2 = delta_u_j * derivative_sigmoid(curr.h_j_hat)
self.dWu += np.outer(delta_u_j_1, x_j)
self.dUu += np.outer(delta_u_j_2, curr.h_j_hat)
self.dbu += delta_u_j_1
#self.dL[:, curr.index] += np.dot(self.Wu, delta_u_j_1)
self.dL[curr.index] += np.dot(self.Wu, delta_u_j_1)
for i, rel in curr.kids:
kid = tree.nodes[i]
to_do.append(kid)
delta_f_jk = delta_c_j * kid.c_j
delta_f_jk_1 = delta_f_jk * derivative_sigmoid(x_j)
self.dWf += np.outer(delta_f_jk_1, x_j)
self.dbf += delta_f_jk_1
delta_f_jk_2 = delta_f_jk * derivative_sigmoid(kid.h_j)
self.dUf += np.outer(delta_f_jk_2, kid.h_j)
kid.deltas = np.dot(self.Uf, delta_f_jk)