-
Notifications
You must be signed in to change notification settings - Fork 0
/
rnn_pronouns_new_keras.py
343 lines (242 loc) · 12.5 KB
/
rnn_pronouns_new_keras.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
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
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
from keras.models import Sequential, Graph, Model
from keras.layers import Dense, Dropout, Activation, Merge, Input, merge
from keras.layers.recurrent import LSTM, GRU
from keras.optimizers import SGD
from keras.datasets import reuters
from keras.callbacks import Callback,ModelCheckpoint
import conllutil
from sklearn.feature_extraction import DictVectorizer
from sklearn.metrics import f1_score, classification_report
import codecs
import lwvlib
import numpy as np
import gzip
import sys
from svm_pronouns import iter_data
import json
import copy
class CustomCallback(Callback):
def __init__(self, dev_left,dev_right,dev_labels,index2label):
self.dev_left=dev_left
self.dev_right=dev_right
self.dev_labels=dev_labels
self.index2label=index2label
self.dev_labels_text=[]
for l in self.dev_labels:
self.dev_labels_text.append(index2label[np.argmax(l)])
def on_epoch_end(self, epoch, logs={}):
print logs
corr=0
tot=0
preds = self.model.predict([self.dev_left, self.dev_right], verbose=1)
preds_text=[]
for l in preds:
preds_text.append(self.index2label[np.argmax(l)])
print "Micro f-score:", f1_score(self.dev_labels_text,preds_text,average=u"micro")
print "Macro f-score:", f1_score(self.dev_labels_text,preds_text,average=u"macro")
print classification_report(self.dev_labels_text, preds_text)
for i in xrange(len(self.dev_labels)):
# next_index = sample(preds[i])
next_index = np.argmax(preds[i])
# print preds[i],next_index,index2label[next_index]
l = self.index2label[next_index]
# print "correct:", index2label[np.argmax(dev_labels[i])], "predicted:",l
if self.index2label[np.argmax(self.dev_labels[i])]==l:
corr+=1
tot+=1
print corr,"/",tot
vocab=set()
vocab2index = None
index2vocab = None
dist_labels= None
label2index = None
index2label = None
window=10
def data(f, batch=50):
global vocab2index, dist_labels, label2index, vocab, window, index2label
contexts_left=[]
contexts_right=[]
labels=[]
for sent,replace_tokens in iter_data(f, max_examples=batch): # label, category, source sent, target sent, alignment, doc id
label=sent[0]
target=sent[3].strip().split(u" ")
for l,replace in zip(label.split(u" "),replace_tokens):
# l = correct label
# replace = index of the token in target sentence
context_left=[]
context_right=[]
# print l,replace,target[replace]
# left sided context
if replace<window:
context_left=target[:replace]
context_left.insert(0,u"<BOS>")
else:
context_left=target[replace-window:replace]
# right sided context
if replace>len(target)-window:
context_right=target[replace:]
context_right.append(u"<EOS>")
else:
context_right=target[replace:replace+window]
contexts_right.append(u" ".join(t.lower() for t in context_right)) # make it a string, why...?
contexts_left.append(u" ".join(t.lower() for t in context_left))
labels.append(l)
text=u" ".join(l for l in contexts_left+contexts_right)
# create index dictionaries, these must then be saven as json
if not vocab2index:
vocab=set(text.split(u" "))
print >> sys.stderr, "Vocabulary size:", len(vocab)
print >> sys.stderr, "Classes:", len(set(labels))
vocab2index = dict((c, i) for i, c in enumerate(vocab))
# index2vocab = dict((i, c) for i, c in enumerate(vocab))
dist_labels=set(labels)
label2index = dict((c, i) for i, c in enumerate(dist_labels))
index2label = dict((i, c) for i, c in enumerate(dist_labels))
# vectorize the data
vectorized_left=np.zeros((len(contexts_left),window,len(vocab)))
vectorized_labels=np.zeros((len(labels),len(dist_labels)))
for i,context in enumerate(contexts_left): # context is a string
for j,token in enumerate(context.split(u" ")): # hah, now I'm again splitting it! Should do something...
if token not in vocab2index:
continue
vectorized_left[i,j,vocab2index[token]]=1.0
vectorized_labels[i,label2index[labels[i]]]=1.0
vectorized_right=np.zeros((len(contexts_right),window,len(vocab)))
for i,context in enumerate(contexts_right):
for j,token in enumerate(context.split(u" ")):
if token not in vocab2index:
continue
vectorized_right[i,j,vocab2index[token]]=1.0
#I'm so sorry you have to see this T_T
revec_left = np.zeros((vectorized_left.shape[:2]), dtype=np.int32)
revec_right = np.zeros((vectorized_right.shape[:2]), dtype=np.int32)
for a, example in enumerate(vectorized_left):
for b, sequence in enumerate(example):
revec_left[a][b] = np.argmax(sequence)
for a, example in enumerate(vectorized_right):
for b, sequence in enumerate(example):
revec_right[a][b] = np.argmax(sequence)
return revec_left, revec_right, vectorized_labels #vectorized_left,vectorized_right,vectorized_labels
#labels_v=np.array([labels2index[i] for i in next_chars ])
#from keras.utils import np_utils, generic_utils
#labels_v = np_utils.to_categorical(next_chars_v, len(chars)) # http://stackoverflow.com/questions/31997366/python-keras-shape-mismatch-error
def sample(a, temperature=1.0):
# helper function to sample an index from a probability array
a = np.log(a) / temperature
a = np.exp(a) / np.sum(np.exp(a))
return np.argmax(np.random.multinomial(1, a, 1))
#f=codecs.getreader(u"utf-8")(gzip.open(u"train_data/all.en-fr.filtered.withids"))
# GET DATA
with codecs.open(u"train_data/all.en-fr.filtered.withids", u"rt", u"utf-8") as f:
train_left,train_right,train_labels=data(f,batch=10000)
print "Training examples:",len(train_labels)
#dev data
with codecs.open(u"dev_data/TEDdev.en-fr.data.filtered.withids",u"rt",u"utf-8") as f:
dev_left,dev_right,dev_labels=data(f,batch=1000)
print "Devel examples:",len(dev_labels)
vec_size = 128
#Let's build a fancy functional model a'la new keras
print 'Build model...'
#First the inputs
left_target = Input(shape=(window, ), name='left_target', dtype='int32')
right_target = Input(shape=(window, ), name='right_target', dtype='int32')
#Future inputs
'''
left_target_pos = Input(shape=(window, ), name='left_target_pos', dtype='int32')
right_target_pos = Input(shape=(window, ), name='right_target_pos', dtype='int32')
left_source = Input(shape=(window, ), name='left_source', dtype='int32')
right_source = Input(shape=(window, ), name='right_source', dtype='int32')
'''
#Future embeddings and vectors
'''
shared_emb_pos = Embedding(len(vocab_pos), vec_size, input_length=window, mask_zero=True)
shared_emb_source = Embedding(len(vocab_source), vec_size, input_length=window, mask_zero=True)
vector_left_target_pos = shared_emb_pos(left_target_pos)
vector_right_target_pos = shared_emb_pos(right_target_pos)
vector_left_source = shared_emb_source(left_source)
vector_right_source = shared_emb_source(right_source)
'''
#Then the embeddings
from keras.layers.embeddings import Embedding
shared_emb = Embedding(len(vocab), vec_size, input_length=window, mask_zero=True)
vector_left_target = shared_emb(left_target)
vector_right_target = shared_emb(right_target)
#The lstms
right_target_lstm = LSTM(128)
left_target_lstm = LSTM(128)
left_target_lstm_out = left_target_lstm(vector_left_target)
right_target_lstm_out = right_target_lstm(vector_right_target)
merged_vector = merge([left_target_lstm_out, right_target_lstm_out], mode='concat', concat_axis=-1)
#The prediction layer
dense_out = Dense(128, activation='relu')(merged_vector)
predictions = Dense(len(dist_labels), activation='softmax')(dense_out)
model = Model(input=[left_target, right_target], output=predictions)
model.compile(optimizer='rmsprop', loss='categorical_crossentropy', metrics=['accuracy'])
#Make callbacks
evalcb=CustomCallback(dev_left,dev_right,dev_labels,index2label) # evaluate after each epoch
savecb=ModelCheckpoint(u"rnn_model.model", monitor='val_acc', verbose=1, save_best_only=True, mode='auto') # save model (I have not tested this!)
model.fit([train_left, train_right], train_labels, batch_size=10, nb_epoch=20, callbacks=[evalcb, savecb])
'''
model = Graph()
model.add_input(name='left', input_shape=(window, ), dtype='int')
model.add_input(name='right', input_shape=(window, ), dtype='int')
#For future
#model.add_input(name='left_s', input_shape=(window, ))
#model.add_input(name='right_s', input_shape=(window, ))
vec_size=128
#Hack Embeddings in
from keras.layers.embeddings import Embedding
#Oh dear! What a hacky way to do this!
shared_emb = Embedding(len(vocab), vec_size, input_length=window, mask_zero=True)
#right_emb = copy.copy(left_emb)#Embedding(len(vocab), vec_size, input_length=window, mask_zero=True)
#right_emb.params = left_emb.params
#right_emb.W = left_emb.W
model.add_node(left_emb, name='emb_left', input='left')
model.add_node(right_emb, name='emb_right', input='right')
#Left & Right LSTM
#model.add_node(LSTM(128, return_sequences=False, dropout_W=0.1, dropout_U=0.1, input_shape=(window, vec_size)), name='left_lstm', input='emb_left')
#model.add_node(LSTM(128, return_sequences=False, dropout_W=0.1, dropout_U=0.1, input_shape=(window, vec_size)), name='right_lstm', input='emb_right')
model.add_node(LSTM(128, return_sequences=True, input_shape=(window, vec_size)), name='left_lstm_p', input='emb_left')
model.add_node(LSTM(128, return_sequences=False, input_shape=(window, vec_size)), name='left_lstm', input='left_lstm_p')
model.add_node(LSTM(128, return_sequences=True, input_shape=(window, vec_size)), name='right_lstm_p', input='emb_right')
model.add_node(LSTM(128, return_sequences=False, input_shape=(window, vec_size)), name='right_lstm', input='right_lstm_p')
#Time to Predict
model.add_node(Dense(128, activation='relu'), name='dense_1', inputs=['left_lstm', 'right_lstm'], merge_mode='concat')
model.add_node(Dense(len(dist_labels), activation='softmax'), name='softmax_1', input='dense_1') # why I have two dense layers here?
model.add_output(name='output', input='softmax_1')
model.compile(loss={'output':'categorical_crossentropy'}, optimizer='adam')
#dev data
with codecs.open(u"dev_data/TEDdev.en-fr.data.filtered.withids",u"rt",u"utf-8") as f:
dev_left,dev_right,dev_labels=data(f,batch=1000)
print "Devel examples:",len(dev_labels)
evalcb=CustomCallback(dev_left,dev_right,dev_labels,index2label) # evaluate after each epoch
savecb=ModelCheckpoint(u"rnn_model.model", monitor='val_acc', verbose=1, save_best_only=True, mode='auto') # save model (I have not tested this!)
#for x in range(100):
model.fit({'left':train_left, 'right':train_right, 'output':train_labels}, batch_size=10 ,nb_epoch=5, verbose=1, show_accuracy=True, validation_data={'left':dev_left, 'right':dev_right, 'output':dev_labels}, callbacks=[evalcb])
#print model.evaluate({'left':dev_left, 'right':dev_right, 'output':dev_labels}, show_accuracy=True)
import pdb;pdb.set_trace()
'''
'''
# BUILD MODEL: 2 concatenated LSTMs
print 'Build model...'
model_left = Sequential()
model_left.add()
model_right = Sequential()
model_right.add(LSTM(128, return_sequences=False, input_shape=(window, len(vocab))))
#model_right.add(Dropout(0.2))
decoder = Sequential()
decoder.add(Merge([model_left, model_right], mode='concat'))
decoder.add(Dense(128, activation='relu'))
decoder.add(Dense(len(dist_labels), activation='softmax')) # why I have two dense layers here?
decoder.compile(loss='categorical_crossentropy', optimizer='rmsprop', class_mode='categorical')
#model.add(LSTM(128, return_sequences=False))
#model.add(Dropout(0.2))
#model.add(Dense(len(dist_labels)))
#model.add(Activation('softmax'))
#model.compile(loss='categorical_crossentropy', optimizer='rmsprop')
#np.random.seed(1337) # for reproducibility
'''
#Seems cool, but I've got no time to get them working!
#evalcb=CustomCallback(dev_left,dev_right,dev_labels,index2label) # evaluate after each epoch
#savecb=ModelCheckpoint(u"rnn_model.model", monitor='val_acc', verbose=1, save_best_only=True, mode='auto') # save model (I have not tested this!)
#decoder.fit([train_left,train_right], train_labels, batch_size=10, nb_epoch=100, verbose=1, show_accuracy=True, validation_data=([dev_left,dev_right],dev_labels), callbacks=[evalcb,savecb])