forked from cgpotts/cs224u
-
Notifications
You must be signed in to change notification settings - Fork 0
/
nli_rnn.py
executable file
·266 lines (220 loc) · 9.01 KB
/
nli_rnn.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
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
import sys
import numpy as np
import random
from collections import defaultdict
import copy
from utils import randvec, randmatrix, d_tanh, softmax, progress_bar
__author__ = "Christopher Potts"
__version__ = "CS224u, Stanford, Spring 2016"
class ClassifierRNN:
"""Very simple Recurrent Neural Network for classification
problems. The structure of the network is as follows:
y
/|
b | W_hy
|
h_0 -- W_hh -- h_1 -- W_hh -- h_2 -- W_hh -- h_3
| | |
| W_xh | W_xh | W_xh
| | |
x_1 x_2 x_3
where x_i are the inputs, h_j are the hidden units, and y is a
one-hot vector indicating the true label for this sequence. The
parameters are W_xh, W_hh, W_hy, and the bias b. The inputs x_i
come from a user-supplied embedding space for the vocabulary. These
can either be random or pretrained. The network equations in brief:
h[t] = tanh(x[t].dot(W_xh) + h[t-1].dot(W_hh))
y = softmax(h[-1].dot(W_hy) + b)
The network will work for any kind of classification task. For
NLI, we process the premise and hypothesis in order and then
use the final hidden state as the basis for the predictions:
[1, 0, 0] (entailment, contradiction, neutral)
/ |
b |
|
h0 - h1 - h2 - h3 - h4 - h5 - h6
| | | | | |
x3 x2 x1 x3 x2 x4
| | | | | | look-up in embedding space
every dog danced every dog moved
"""
def __init__(self,
vocab,
embedding,
hidden_dim=20,
eta=0.01,
maxiter=100,
epsilon=1.5e-8,
display_progress=True):
"""
Parameters
----------
vocab : list of str
This should be the vocabulary. It needs to be aligned with
`embedding` in the sense that the ith element of vocab
should be represented by the ith row of `embedding`.
embedding : np.array
Each row represents a word in `vocab`, as described above.
hidden_dim : int (default: 10)
Dimensionality for the hidden layer.
eta : float (default: 0.05)
Learning rate.
maxiter : int (default: 100)
Maximum number of training epochs for SGD.
epsilon : float (default: 1.5e-8)
Training terminates if the error reaches this point (or
`maxiter` is met).
display_progress : bool (default: True)
Whether to print progress reports to stderr.
All of the above are set as attributes. In addition, `self.word_dim`
is set to the dimensionality of the input representations.
"""
self.vocab = dict(zip(vocab, range(len(vocab))))
self.embedding = embedding
self.hidden_dim = hidden_dim
self.eta = eta
self.maxiter = maxiter
self.epsilon = epsilon
self.display_progress = display_progress
self.word_dim = len(embedding[0])
def get_word_rep(self, w):
"""For getting the input representation of word `w` from `self.embedding`."""
word_index = self.vocab[w]
return self.embedding[word_index]
def fit(self, training_data):
"""Train the network.
Parameters
----------
training_data : list of pairs
In each pair, the first element should be a list of items
from the vocabulary (for the NLI task, this is the
concatenation of the premise and hypothesis), and the
second element should be the one-hot label vector.
Attributes
----------
self.output_dim : int
Set based on the length of the labels in `training_data`.
self.W_xh : np.array
Dense connections between the word representations
and the hidden layers. Random initialization.
self.W_hh : np.array
Dense connections between the hidden representations.
Random initialization.
self.W_hy : np.array
Dense connections from the final hidden layer to
the output layer. Random initialization.
self.b : np.array
Output bias. Initialized to all 0.
"""
self.output_dim = len(training_data[0][1])
self.W_xh = randmatrix(self.word_dim, self.hidden_dim)
self.W_hh = randmatrix(self.hidden_dim, self.hidden_dim)
self.W_hy = randmatrix(self.hidden_dim, self.output_dim)
self.b = np.zeros(self.output_dim)
# SGD:
iteration = 0
error = sys.float_info.max
while error > self.epsilon and iteration < self.maxiter:
error = 0.0
random.shuffle(training_data)
for seq, labels in training_data:
self._forward_propagation(seq)
# Cross-entropy error reduces to log(prediction-for-correct-label):
error += -np.log(self.y[np.argmax(labels)])
# Back-prop:
d_W_hy, d_b, d_W_hh, d_W_xh = self._backward_propagation(seq, labels)
# Updates:
self.W_hy -= self.eta * d_W_hy
self.b -= self.eta * d_b
self.W_hh -= self.eta * d_W_hh
self.W_xh -= self.eta * d_W_xh
iteration += 1
if self.display_progress:
# Report the average error:
error /= len(training_data)
progress_bar("Finished epoch %s of %s; error is %s" % (iteration, self.maxiter, error))
if self.display_progress:
sys.stderr.write('\n')
def _forward_propagation(self, seq):
"""
Parameters
----------
seq : list
Variable length sequence of elements in the vocabulary.
Attributes
----------
self.h : np.array
Each row is for a hidden representation. The first row
is an all-0 initial state. The others correspond to
the inputs in seq.
self.y : np.array
The vector of predictions.
"""
self.h = np.zeros((len(seq)+1, self.hidden_dim))
for t in range(1, len(seq)+1):
word_rep = self.get_word_rep(seq[t-1])
self.h[t] = np.tanh(word_rep.dot(self.W_xh) + self.h[t-1].dot(self.W_hh))
self.y = softmax(self.h[-1].dot(self.W_hy) + self.b)
def _backward_propagation(self, seq, y_):
"""
Parameters
----------
seq : list
Variable length sequence of elements in the vocabulary. This
is needed both for its lengths and for its input representations.
y_ : list
The label vector.
Returns
-------
tuple
The matrices of derivatives (d_W_hy, d_b, d_W_hh, d_W_xh).
"""
# Output errors:
y_err = self.y
y_err[np.argmax(y_)] -= 1
h_err = y_err.dot(self.W_hy.T) * d_tanh(self.h[-1])
d_W_hy = np.outer(self.h[-1], y_err)
d_b = y_err
# For accumulating the gradients through time:
d_W_hh = np.zeros(self.W_hh.shape)
d_W_xh = np.zeros(self.W_xh.shape)
# Back-prop through time; the +1 is because the 0th
# hidden state is the all-0s initial state.
num_steps = len(seq)+1
for t in reversed(range(1, num_steps)):
d_W_hh += np.outer(self.h[t], h_err)
word_rep = self.get_word_rep(seq[t-1])
d_W_xh += np.outer(word_rep, h_err)
h_err = h_err.dot(self.W_hh.T) * d_tanh(self.h[t])
return (d_W_hy, d_b, d_W_hh, d_W_xh)
def predict(self, seq):
"""
Parameters
----------
seq : list
Variable length sequence of elements in the vocabulary.
Returns
-------
int
The index of the highest probability class according to
the model.
"""
self._forward_propagation(seq)
return np.argmax(self.y)
######################################################################
if __name__ == '__main__':
T = 'T'
F = 'F'
train = [
# p q XOR
([T ,T], [1., 0.]),
([T, F], [0., 1.]),
([F, T], [0., 1.]),
([F, F], [1., 0.])]
vocab = [T, F]
embedding = np.array([randvec(10) for _ in vocab])
mod = ClassifierRNN(vocab=vocab, embedding=embedding, maxiter=1000)
mod.fit(copy.copy(train))
for x, y in train:
p = mod.predict(x)
print(p == np.argmax(y), mod.y, y)