-
Notifications
You must be signed in to change notification settings - Fork 23
/
model.py
245 lines (211 loc) · 9.64 KB
/
model.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
import pickle
import numpy as np
import os
import json
from random import shuffle
import sys
sys.path.append('..')
from config import data_path, experiment_path
from train.data import Vocab
import time
def sigmoid(x):
return 1 / (np.exp(-x) + 1)
def softmax(w):
assert w.ndim == 2 or w.ndim == 1, 'softmax dim error %d' % w.ndim
w = np.expand_dims(w, axis=0) if w.ndim == 1 else w
e = np.exp(w - np.amax(w, axis=1, keepdims=True))
dist = e / np.sum(e, axis=1, keepdims=True)
return dist
def tanh(x):
return np.tanh(x)
def find_top_N(a,N):
return np.argsort(a)[::-1][:N]
def sample(a, temperature=1.0):
# helper function to sample an index from a probability array
# from https://github.com/fchollet/keras/blob/master/examples/lstm_text_generation.py
a = np.log(a) / temperature
a = np.exp(a) / np.sum(np.exp(a))
return np.argmax(np.random.multinomial(1, a, 1))
class LSTM_Model():
"""The numpy implementation of the NN Language Model"""
def __init__(self, experiment_id=0, comp=0):
print('LSTM model: exp {} comp {}'.format(experiment_id, comp))
self.config = json.loads(open(os.path.join(experiment_path, str(experiment_id), "config.json"), "rt").read())
self.weights = self._load_model(experiment_id, comp)
self.embed_size = self.config['embed_size']
self.hidden_size = self.config['hidden_size']
self.share_embedding = self.config['share_embedding']
self.hidden = np.zeros((1, self.hidden_size))
self.cell = np.zeros((1, self.hidden_size))
if self.config['D_softmax']:
self.blocks = self.weights['LM']
self.embed_size = sum([x[0] for x in self.config['embedding_seg']])
self.weights['LM'] = np.zeros((self.weights['b2'].shape[0], self.embed_size))
col_s = 0
for i, (size, s, e) in enumerate(self.config['embedding_seg']):
self.weights['LM'][s:e, col_s:col_s + size] = self.blocks[i]
col_s += size
if self.config['V_table']:
self.blocks = []
self.v_tables = []
embeddings = []
for i, seg in enumerate(self.config['embedding_seg']):
block = self.weights['LM{}'.format(i)]
self.blocks.append(block)
if i != 0:
v_table = self.weights['VT{}'.format(i)]
self.v_tables.append(v_table)
embeddings.append(np.dot(block, v_table))
else:
self.v_tables.append(None)
embeddings.append(block)
self.weights['LM'] = np.concatenate(embeddings, axis=0)
def _load_model(self, experiment_id=0, comp=0):
if comp:
file = 'lstm_weights_comp_{}.pkl'.format(comp)
print('use compressed model, comp_{}'.format(comp))
else:
file = 'lstm_weights.pkl'
weights = pickle.load(open(os.path.join(experiment_path, str(experiment_id), 'weights', file), 'rb'))
# temp code for shu compression algorithm
# make this a model parameter once done
use_hash_code = False
if use_hash_code:
M = 32
K = 16
code = np.loadtxt(open(os.path.join(experiment_path, str(experiment_id), 'weights', 'LM.codes'), 'r'), dtype=np.int)
codebook = np.load(open(os.path.join(experiment_path, str(experiment_id), 'weights', 'LM.codebook.npy'), 'rb'))
LM = np.zeros((code.shape[0], codebook[0].shape[0]))
for row in range(LM.shape[0]):
c = code[row]
LM[row, :] = np.sum([codebook[i*K+x] for i, x in enumerate(c)], axis=0)
print('hash coded embedding loaded')
print('distance {}'.format(np.mean(np.linalg.norm(LM - weights['LM'], axis=1).tolist())))
weights['LM'] = LM
np.save('LM.npy', LM)
return weights
def predict(self, index, vocab=None, reset=False):
if reset: # hidden and cell should be set before using this function
self.hidden = np.zeros(shape=self.hidden.shape)
self.cell = np.zeros(shape=self.cell.shape)
start_time = time.time()
self._lstm_cell(index)
log_lstm_time = time.time() - start_time
start_time = time.time()
y = self.project(self.hidden, vocab)
if self.config['self_norm']:
pred = np.exp(y)
else:
pred = softmax(y)
log_softmax_time = time.time() - start_time
return pred, y, log_lstm_time, log_softmax_time
def _lstm_cell(self, index):
# embedding lookup
e = self.weights['LM'][index,:]
i = np.dot(self.hidden, self.weights['HMi']) + np.dot(e, self.weights['IMi']) + self.weights['bi']
f = np.dot(self.hidden, self.weights['HMf']) + np.dot(e, self.weights['IMf']) + self.weights['bf']
o = np.dot(self.hidden, self.weights['HMo']) + np.dot(e, self.weights['IMo']) + self.weights['bo']
g = np.dot(self.hidden, self.weights['HMg']) + np.dot(e, self.weights['IMg']) + self.weights['bg']
i = sigmoid(i)
f = sigmoid(f)
o = sigmoid(o)
g = tanh(g)
self.cell = np.multiply(self.cell, f) + np.multiply(g, i)
# new hidden transformation matrix
self.hidden = np.multiply(tanh(self.cell), o)
def project(self, hidden, vocab=None):
# output word representation
if self.share_embedding:
if self.config['D_softmax']:
temp = np.dot(hidden, self.weights["PM"])
y = []
col_s = 0
for i, (size, s, e) in enumerate(self.config['embedding_seg']):
if vocab:
if e is None:
e = sys.maxsize
sub_vocab_lookup = [v - s for v in vocab if v >= s and v < e]
y.append(np.dot(temp[:, col_s: col_s + size], self.blocks[i][sub_vocab_lookup].T))
else:
y.append(np.dot(temp[:, col_s: col_s + size], self.blocks[i].T))
col_s += size
if vocab:
y = np.concatenate(y, axis=1) + self.weights['b2'][vocab]
else:
y = np.concatenate(y, axis=1) + self.weights['b2']
elif self.config['V_table']:
temp = np.dot(hidden, self.weights["PM"])
y = []
for i, (size, s, e) in enumerate(self.config['embedding_seg']):
if vocab:
if e is None:
e = sys.maxsize
sub_vocab_lookup = [v - s for v in vocab if v >= s and v < e]
if i != 0:
y.append(np.dot(np.dot(temp, self.v_tables[i].T), self.blocks[i][sub_vocab_lookup].T))
else:
y.append(np.dot(temp, self.blocks[i][sub_vocab_lookup].T))
else:
if i != 0:
y.append(np.dot(np.dot(temp, self.v_tables[i].T), self.blocks[i].T))
else:
y.append(np.dot(temp, self.blocks[i].T))
if vocab:
y = np.concatenate(y, axis=1) + self.weights['b2'][vocab]
else:
y = np.concatenate(y, axis=1) + self.weights['b2']
else:
if vocab:
y = np.dot(np.dot(hidden, self.weights["PM"]), self.weights["LM"][vocab].T) + self.weights['b2'][vocab]
else:
y = np.dot(np.dot(hidden, self.weights["PM"]), self.weights["LM"].T) + self.weights['b2']
else:
if vocab:
y = np.dot(hidden, self.weights['UM'][vocab]) + self.weights['b2'][vocab]
else:
y = np.dot(hidden, self.weights['UM']) + self.weights['b2']
return y
def predict_with_context(self, index, hidden, cell, vocab=None):
self.hidden = hidden
self.cell = cell
return self.predict(index, vocab), self.hidden, self.cell
def evaluate(self, start, inputs):
probs = []
pred = self.predict([start], vocab=None, reset=True)
for input in inputs:
probs.append(pred[0, input])
pred = self.predict([input])
return [-np.log(p) for p in probs]
def show_prob(inputs):
results = model.evaluate(w2i['<eos>'], [w2i[word] for word in inputs])
print(results)
print(sum([np.log(prob) for prob in results]))
if __name__ == "__main__":
# test the model
experiment_id = 22
config = json.loads(open(os.path.join(experiment_path, str(experiment_id), "config.json"), "rt").read())
vocab = Vocab(config['vocab_size'], char_based=config['char_rnn'])
i2w = vocab.i2w
w2i = vocab.w2i
model = LSTM_Model(experiment_id=experiment_id)
starting_text = '<eos>'
result = []
step = 0
#print(sum(model.evaluate(w2i[starting_text], [w2i[x] for x in 'ことも無'])))
#print(sum(model.evaluate(w2i[starting_text], [w2i[x] for x in 'こともむ'])))
while True:
result.append(starting_text)
pred = model.predict([w2i[starting_text]])
pred = pred[0].tolist()
next_idx = sample(pred)
starting_text = i2w[next_idx]
step = step + 1
if step == 100:
break
print('--- generated sentence')
print(' '.join([x.split('/')[0] for x in result]))
show_prob(result)
print('--- random sentence by same collection of words, check the difference to see if the model is correct')
shuffle(result)
print(' '.join([x.split('/')[0] for x in result]))
show_prob(result)