質問編集履歴
7
test
CHANGED
File without changes
|
test
CHANGED
@@ -8,9 +8,7 @@
|
|
8
8
|
|
9
9
|
time_layer.pyをいくら変更しても、反映されません。
|
10
10
|
|
11
|
-
このプログラムは使われてないのでしょうか。
|
12
|
-
|
13
|
-
|
11
|
+
rnnlm.pyのpredict関数で使われてるforward関数は、time_layer.pyのTimeEmbeddingクラスのものではないのでしょうか?
|
14
12
|
|
15
13
|
|
16
14
|
|
6
test
CHANGED
File without changes
|
test
CHANGED
@@ -100,7 +100,7 @@
|
|
100
100
|
|
101
101
|
for layer in self.layers:
|
102
102
|
|
103
|
-
xs = layer.forward
|
103
|
+
xs = layer.forward(xs)
|
104
104
|
|
105
105
|
|
106
106
|
|
5
test
CHANGED
File without changes
|
test
CHANGED
@@ -108,7 +108,7 @@
|
|
108
108
|
|
109
109
|
|
110
110
|
|
111
|
-
def forward
|
111
|
+
def forward(self, xs, ts):
|
112
112
|
|
113
113
|
|
114
114
|
|
4
test
CHANGED
File without changes
|
test
CHANGED
@@ -10,6 +10,8 @@
|
|
10
10
|
|
11
11
|
このプログラムは使われてないのでしょうか。
|
12
12
|
|
13
|
+
それとも、他のプログラムが見えないところで使われているのでしょうか。
|
14
|
+
|
13
15
|
|
14
16
|
|
15
17
|
|
3
test
CHANGED
File without changes
|
test
CHANGED
File without changes
|
2
タイトルの変更
test
CHANGED
@@ -1 +1 @@
|
|
1
|
-
|
1
|
+
python コードを変更しても何の反映もない
|
test
CHANGED
File without changes
|
1
コードの追加
test
CHANGED
File without changes
|
test
CHANGED
@@ -12,7 +12,217 @@
|
|
12
12
|
|
13
13
|
|
14
14
|
|
15
|
+
|
16
|
+
|
17
|
+
rnnlm.py
|
18
|
+
|
15
|
-
python
|
19
|
+
```python
|
20
|
+
|
21
|
+
import sys
|
22
|
+
|
23
|
+
sys.path.append('..')
|
24
|
+
|
25
|
+
import numpy as np
|
26
|
+
|
27
|
+
from common.time_layers import TimeLSTM,TimeAffine,TimeSoftmaxWithLoss,TimeEmbedding
|
28
|
+
|
29
|
+
#from common.base_model import BaseModel
|
30
|
+
|
31
|
+
|
32
|
+
|
33
|
+
|
34
|
+
|
35
|
+
class Rnnlm:
|
36
|
+
|
37
|
+
def __init__(self, vocab_size=10000, wordvec_size=100, hidden_size=100):
|
38
|
+
|
39
|
+
|
40
|
+
|
41
|
+
V, D, H = vocab_size, wordvec_size, hidden_size
|
42
|
+
|
43
|
+
rn = np.random.randn
|
44
|
+
|
45
|
+
|
46
|
+
|
47
|
+
# 重みの初期化
|
48
|
+
|
49
|
+
embed_W = (rn(V, D) / 100).astype('f')
|
50
|
+
|
51
|
+
|
52
|
+
|
53
|
+
lstm_Wx = (rn(D, 4 * H) / np.sqrt(D)).astype('f')
|
54
|
+
|
55
|
+
lstm_Wh = (rn(H, 4 * H) / np.sqrt(H)).astype('f')
|
56
|
+
|
57
|
+
lstm_b = np.zeros(4 * H).astype('f')
|
58
|
+
|
59
|
+
affine_W = (rn(H, V) / np.sqrt(H)).astype('f')
|
60
|
+
|
61
|
+
affine_b = np.zeros(V).astype('f')
|
62
|
+
|
63
|
+
|
64
|
+
|
65
|
+
# レイヤの生成
|
66
|
+
|
67
|
+
self.layers = [
|
68
|
+
|
69
|
+
TimeEmbedding(embed_W),
|
70
|
+
|
71
|
+
TimeLSTM(lstm_Wx, lstm_Wh, lstm_b, stateful=True),
|
72
|
+
|
73
|
+
TimeAffine(affine_W, affine_b)
|
74
|
+
|
75
|
+
]
|
76
|
+
|
77
|
+
|
78
|
+
|
79
|
+
self.loss_layer = TimeSoftmaxWithLoss()
|
80
|
+
|
81
|
+
self.lstm_layer = self.layers[1]
|
82
|
+
|
83
|
+
|
84
|
+
|
85
|
+
# すべての重みと勾配をリストにまとめる
|
86
|
+
|
87
|
+
self.params, self.grads = [], []
|
88
|
+
|
89
|
+
for layer in self.layers:
|
90
|
+
|
91
|
+
self.params += layer.params
|
92
|
+
|
93
|
+
self.grads += layer.grads
|
94
|
+
|
95
|
+
|
96
|
+
|
97
|
+
def predict(self, xs):
|
98
|
+
|
99
|
+
for layer in self.layers:
|
100
|
+
|
101
|
+
xs = layer.forward2(xs)
|
102
|
+
|
103
|
+
|
104
|
+
|
105
|
+
return xs
|
106
|
+
|
107
|
+
|
108
|
+
|
109
|
+
def forward2(self, xs, ts):
|
110
|
+
|
111
|
+
|
112
|
+
|
113
|
+
score = self.predict(xs)
|
114
|
+
|
115
|
+
loss = self.loss_layer.forward(score, ts)
|
116
|
+
|
117
|
+
|
118
|
+
|
119
|
+
return loss
|
120
|
+
|
121
|
+
|
122
|
+
|
123
|
+
def backward(self, dout=1):
|
124
|
+
|
125
|
+
dout = self.loss_layer.backward(dout)
|
126
|
+
|
127
|
+
for layer in reversed(self.layers):
|
128
|
+
|
129
|
+
dout = layer.backward(dout)
|
130
|
+
|
131
|
+
return dout
|
132
|
+
|
133
|
+
|
134
|
+
|
135
|
+
def reset_state(self):
|
136
|
+
|
137
|
+
self.lstm_layer.reset_state()
|
138
|
+
|
139
|
+
```
|
140
|
+
|
141
|
+
|
142
|
+
|
143
|
+
上のコードのself.layersでTimeEmbeddingオブジェクトを生成してると思うのですが、
|
144
|
+
|
145
|
+
|
146
|
+
|
147
|
+
time_layer.py
|
148
|
+
|
149
|
+
```python
|
150
|
+
|
151
|
+
class TimeEmbedding:
|
152
|
+
|
153
|
+
def __init__(self, W):
|
154
|
+
|
155
|
+
|
156
|
+
|
157
|
+
self.params = [W]
|
158
|
+
|
159
|
+
self.grads = [np.zeros_like(W)]
|
160
|
+
|
161
|
+
self.layers = None
|
162
|
+
|
163
|
+
self.W = W
|
164
|
+
|
165
|
+
|
166
|
+
|
167
|
+
|
168
|
+
|
169
|
+
def forward(self, xs):
|
170
|
+
|
171
|
+
N, T = xs.shape
|
172
|
+
|
173
|
+
V, D = self.W.shape
|
174
|
+
|
175
|
+
|
176
|
+
|
177
|
+
out = np.empty((N, T, D), dtype='f')
|
178
|
+
|
179
|
+
self.layers = []
|
180
|
+
|
181
|
+
|
182
|
+
|
183
|
+
for t in range(T):
|
184
|
+
|
185
|
+
layer = Embedding(self.W)
|
186
|
+
|
187
|
+
out[:, t, :] = layer.forward(xs[:, t])
|
188
|
+
|
189
|
+
self.layers.append(layer)
|
190
|
+
|
191
|
+
|
192
|
+
|
193
|
+
return out
|
194
|
+
|
195
|
+
|
196
|
+
|
197
|
+
def backward(self, dout):
|
198
|
+
|
199
|
+
N, T, D = dout.shape
|
200
|
+
|
201
|
+
|
202
|
+
|
203
|
+
grad = 0
|
204
|
+
|
205
|
+
for t in range(T):
|
206
|
+
|
207
|
+
layer = self.layers[t]
|
208
|
+
|
209
|
+
layer.backward(dout[:, t, :])
|
210
|
+
|
211
|
+
grad += layer.grads[0]
|
212
|
+
|
213
|
+
|
214
|
+
|
215
|
+
self.grads[0][...] = grad
|
216
|
+
|
217
|
+
return None
|
218
|
+
|
219
|
+
|
220
|
+
|
221
|
+
```
|
222
|
+
|
223
|
+
上のtime_layer.pyをいくらいじっても、何の反映もありません。
|
224
|
+
|
225
|
+
|
16
226
|
|
17
227
|
https://github.com/oreilly-japan/deep-learning-from-scratch-2/blob/master/ch07/generate_text.py
|
18
228
|
|