/
main_lasagne_span.py
407 lines (288 loc) · 16.2 KB
/
main_lasagne_span.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
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
import sys
import os
import time
import numpy as np
import theano
import theano.tensor as T
from datetime import datetime
import cPickle
from scipy.stats import pearsonr
from progressbar import ProgressBar
sys.path.insert(0, os.path.abspath('../Lasagne'))
from lasagne.layers import InputLayer, LSTMLayer, NonlinearityLayer, SliceLayer, FlattenLayer, EmbeddingLayer,\
ElemwiseMergeLayer, ReshapeLayer, get_output, get_all_params, get_all_param_values, set_all_param_values, \
get_output_shape, DropoutLayer,DenseLayer,ElemwiseSumLayer,Conv2DLayer, Conv1DLayer, CustomRecurrentLayer, \
AbsSubLayer,ConcatLayer, Pool1DLayer, FeaturePoolLayer,count_params,MaxPool2DLayer,MaxPool1DLayer,DimshuffleLayer
from lasagne.regularization import regularize_layer_params_weighted, l2, l1,regularize_layer_params,\
regularize_network_params
from lasagne.nonlinearities import tanh, sigmoid, softmax, rectify
from lasagne.objectives import categorical_crossentropy, squared_error, categorical_accuracy, binary_crossentropy,\
binary_accuracy
from lasagne.updates import sgd, adagrad, adadelta, nesterov_momentum, rmsprop, adam
from lasagne.init import GlorotUniform
from utils import read_sequence_dataset_onehot, iterate_minibatches_,loadWord2VecMap, generateTestInput
def event_span_classifier(args, input_var, target_var, wordEmbeddings, seqlen, num_feats):
print("Building model with 1D Convolution")
vocab_size = wordEmbeddings.shape[1]
wordDim = wordEmbeddings.shape[0]
kw = 2
num_filters = seqlen-kw+1
stride = 1
#important context words as channels
#CNN_sentence config
filter_size=wordDim
pool_size=seqlen-filter_size+1
input = InputLayer((None, seqlen, num_feats),input_var=input_var)
batchsize, _, _ = input.input_var.shape
emb = EmbeddingLayer(input, input_size=vocab_size, output_size=wordDim, W=wordEmbeddings.T)
#emb.params[emb.W].remove('trainable') #(batchsize, seqlen, wordDim)
#print get_output_shape(emb)
reshape = ReshapeLayer(emb, (batchsize, seqlen, num_feats*wordDim))
#print get_output_shape(reshape)
conv1d = Conv1DLayer(reshape, num_filters=num_filters, filter_size=wordDim, stride=1,
nonlinearity=tanh,W=GlorotUniform()) #nOutputFrame = num_flters,
#nOutputFrameSize = (num_feats*wordDim-filter_size)/stride +1
#print get_output_shape(conv1d)
conv1d = DimshuffleLayer(conv1d, (0,2,1))
#print get_output_shape(conv1d)
pool_size=num_filters
maxpool = MaxPool1DLayer(conv1d, pool_size=pool_size)
#print get_output_shape(maxpool)
#forward = FlattenLayer(maxpool)
#print get_output_shape(forward)
hid = DenseLayer(maxpool, num_units=args.hiddenDim, nonlinearity=sigmoid)
network = DenseLayer(hid, num_units=2, nonlinearity=softmax)
prediction = get_output(network)
loss = T.mean(binary_crossentropy(prediction,target_var))
lambda_val = 0.5 * 1e-4
layers = {emb:lambda_val, conv1d:lambda_val, hid:lambda_val, network:lambda_val}
penalty = regularize_layer_params_weighted(layers, l2)
loss = loss + penalty
params = get_all_params(network, trainable=True)
if args.optimizer == "sgd":
updates = sgd(loss, params, learning_rate=args.step)
elif args.optimizer == "adagrad":
updates = adagrad(loss, params, learning_rate=args.step)
elif args.optimizer == "adadelta":
updates = adadelta(loss, params, learning_rate=args.step)
elif args.optimizer == "nesterov":
updates = nesterov_momentum(loss, params, learning_rate=args.step)
elif args.optimizer == "rms":
updates = rmsprop(loss, params, learning_rate=args.step)
elif args.optimizer == "adam":
updates = adam(loss, params, learning_rate=args.step)
else:
raise "Need set optimizer correctly"
test_prediction = get_output(network, deterministic=True)
test_loss = T.mean(binary_crossentropy(test_prediction,target_var))
train_fn = theano.function([input_var, target_var],
loss, updates=updates, allow_input_downcast=True)
test_acc = T.mean(binary_accuracy(test_prediction, target_var))
val_fn = theano.function([input_var, target_var], [test_loss, test_acc], allow_input_downcast=True)
return train_fn, val_fn, network
def save_network(filename, param_values):
with open(filename, 'wb') as f:
cPickle.dump(param_values, f, protocol=cPickle.HIGHEST_PROTOCOL)
def load_network(filename):
with open(filename, 'rb') as f:
param_values = cPickle.load(f)
return param_values
if __name__ == '__main__':
import argparse
parser = argparse.ArgumentParser(description="Usage")
parser.add_argument("--minibatch",dest="minibatch",type=int,default=30)
parser.add_argument("--optimizer",dest="optimizer",type=str,default="adagrad")
parser.add_argument("--epochs",dest="epochs",type=int,default=2)
parser.add_argument("--step",dest="step",type=float,default=0.01)
parser.add_argument("--hiddenDim",dest="hiddenDim",type=int,default=50)
parser.add_argument("--mode",dest="mode",type=str,default='train')
args = parser.parse_args()
# Load the dataset
base_dir = os.path.dirname(os.path.realpath(__file__))
data_dir = os.path.join(base_dir, 'data')
model_dir = os.path.join(base_dir, 'models_single')
if not os.path.exists(model_dir):
os.makedirs(model_dir)
fileIdx = 1
while True:
model_save_path = os.path.join(model_dir,
'model-'+str(args.minibatch)+'-'+args.optimizer+'-'+str(args.epochs)+'-'+str(args.step)+'-'+str(fileIdx))
model_save_pre_path = os.path.join(model_dir,
'model-'+str(args.minibatch)+'-'+args.optimizer+'-'+str(args.epochs)+'-'+str(args.step)+'-'+str(fileIdx-1))
if not os.path.exists(model_save_path+".span"):
break
fileIdx += 1
input_var = T.itensor3('inputs')
target_var = T.fmatrix('targets')
wordEmbeddings = loadWord2VecMap(os.path.join(data_dir, 'word2vec.bin'))
wordEmbeddings = wordEmbeddings.astype(np.float32)
if args.mode == "train":
print("Loading training data...")
X_train, Y_labels_train, seqlen, num_feats = read_sequence_dataset_onehot(data_dir, "train")
X_dev, Y_labels_dev,_,_ = read_sequence_dataset_onehot(data_dir, "dev")
print "window_size is %d"%((seqlen-1)/2)
print "number features is %d"%num_feats
train_fn, val_fn, network = event_span_classifier(args, input_var, target_var, wordEmbeddings, seqlen, num_feats)
print("Starting training modality model...")
best_val_acc = 0
maxlen_train = 0
for x in range(0, len(X_train) - args.minibatch + 1, args.minibatch):
maxlen_train += 1
for epoch in range(args.epochs):
train_loss = 0
train_batches = 0
start_time = time.time()
pbar = ProgressBar(maxval=maxlen_train).start()
for i, batch in enumerate(iterate_minibatches_((X_train, Y_labels_train), args.minibatch, shuffle=True)):
time.sleep(0.01)
pbar.update(i + 1)
inputs, labels= batch
train_loss += train_fn(inputs, labels[:,0:2])
train_batches += 1
pbar.finish()
val_loss = 0
val_acc = 0
val_batches = 0
maxlen_dev = 0
for x in range(0, len(X_dev) - args.minibatch + 1, args.minibatch):
maxlen_dev += 1
pbar = ProgressBar(maxval=maxlen_dev).start()
#important, when the size of dev is big, need use minibatch instead of the whole dev, unless GpuDnnPool:error
for i, batch in enumerate(iterate_minibatches_((X_dev, Y_labels_dev), args.minibatch, shuffle=True)):
time.sleep(0.01)
pbar.update(i + 1)
inputs, labels= batch
loss, acc = val_fn(inputs, labels[:,0:2])
val_acc += acc
val_loss += loss
val_batches += 1
pbar.finish()
print("Epoch {} of {} took {:.3f}s".format(
epoch + 1, args.epochs, time.time() - start_time))
print("training loss:\t\t{:.6f}".format(train_loss / train_batches))
val_score = val_acc / val_batches * 100
print("validation accuracy:\t\t{:.2f} %".format(val_score))
if best_val_acc < val_score:
best_val_acc = val_score
print "Saving model......"
save_network(model_save_path+".span",get_all_param_values(network))
# get real performance
pred_fn = theano.function([input_var], T.argmax(get_output(network, deterministic=True), axis=1))
ann_dir = os.path.join(base_dir, 'annotation/coloncancer')
plain_dir = os.path.join(base_dir, 'original')
output_dir = os.path.join(base_dir, 'uta-output-span-validate')
window_size = (seqlen-1)/2
input_text_dev_dir = os.path.join(plain_dir, "dev")
totalPredEvents = 0
totalCorrEvents = 0
for dir_path, dir_names, file_names in os.walk(input_text_dev_dir):
pbar = ProgressBar(maxval=len(file_names)).start()
for i, fn in enumerate(file_names):
time.sleep(0.01)
pbar.update(i + 1)
spans, features = generateTestInput(data_dir, input_text_dev_dir, fn, window_size, num_feats)
totalPredEvents += len(spans)
predict = pred_fn(features)
dn = os.path.join(output_dir, fn)
if not os.path.exists(dn):
os.makedirs(dn)
outputAnn_path = os.path.join(dn, fn+"."+"Temporal-Relation.system.complete.xml")
with open(outputAnn_path, 'w') as f:
f.write("<?xml version=\"1.0\" encoding=\"UTF-8\"?>\n\n\n")
f.write("<data>\n")
f.write("<info>\n")
f.write(" <savetime>"+datetime.now().strftime('%H:%M:%S %d-%m-%Y')+"</savetime>\n")
f.write(" <progress>completed</progress>\n")
f.write("</info>"+"\n\n\n")
f.write("<schema path=\"./\" protocal=\"file\">temporal-schema.xml</schema>\n\n\n")
f.write("<annotations>\n\n\n")
count=0
for idx, span_label in enumerate(predict):
if span_label == 1:
totalCorrEvents += 1
f.write("\t<entity>\n")
f.write("\t\t<id>"+str(count)+"@"+fn+"@system"+"</id>\n")
f.write("\t\t<span>"+str(spans[idx][0])+","+str(spans[idx][1])+"</span>\n")
f.write("\t\t<type>EVENT</type>\n")
f.write("\t\t<parentsType></parentsType>\n")
f.write("\t\t<properties>\n")
f.write("\t\t\t<DocTimeRel>BEFORE</DocTimeRel>\n")
f.write("\t\t\t<Type>N/A</Type>\n")
f.write("\t\t\t<Degree>N/A</Degree>\n")
f.write("\t\t\t<Polarity>"+"POS"+"</Polarity>\n")
f.write("\t\t\t<ContextualModality>"+"ACTUAL"+"</ContextualModality>\n")
f.write("\t\t\t<ContextualAspect>N/A</ContextualAspect>\n")
f.write("\t\t\t<Permanence>UNDETERMINED</Permanence>\n")
f.write("\t\t</properties>\n")
f.write("\t</entity>\n\n")
count += 1
f.write("\n\n</annotations>\n")
f.write("</data>")
pbar.finish()
print "Total pred events is %d"%totalPredEvents
print "Total corr events is %d"%totalCorrEvents
os.system("python -m anafora.evaluate -r annotation/coloncancer/Dev/ -p uta-output-span-validate/")
elif args.mode == "test":
print("Starting testing...")
print("Loading model...")
_, _,seqlen, num_feats = read_sequence_dataset_onehot(data_dir, "dev")
print "window_size is %d"%((seqlen-1)/2)
print "number features is %d"%num_feats
_, _, network = event_span_classifier(args, input_var, target_var, wordEmbeddings, seqlen, num_feats)
print model_save_pre_path
saved_params = load_network(model_save_pre_path+".span")
set_all_param_values(network, saved_params)
pred_fn = theano.function([input_var], T.argmax(get_output(network, deterministic=True), axis=1))
ann_dir = os.path.join(base_dir, 'annotation/coloncancer')
plain_dir = os.path.join(base_dir, 'original')
output_dir = os.path.join(base_dir, 'uta-output-span')
input_text_test_dir = os.path.join(plain_dir, "test")
window_size = (seqlen-1)/2
totalPredEvents = 0
totalCorrEvents = 0
for dir_path, dir_names, file_names in os.walk(input_text_test_dir):
pbar = ProgressBar(maxval=len(file_names)).start()
for i, fn in enumerate(file_names):
time.sleep(0.01)
pbar.update(i + 1)
spans, features = generateTestInput(data_dir, input_text_test_dir, fn, window_size, num_feats)
totalPredEvents += len(spans)
predict = pred_fn(features)
dn = os.path.join(output_dir, fn)
if not os.path.exists(dn):
os.makedirs(dn)
outputAnn_path = os.path.join(dn, fn+"."+"Temporal-Relation.system.complete.xml")
with open(outputAnn_path, 'w') as f:
f.write("<?xml version=\"1.0\" encoding=\"UTF-8\"?>\n\n\n")
f.write("<data>\n")
f.write("<info>\n")
f.write(" <savetime>"+datetime.now().strftime('%H:%M:%S %d-%m-%Y')+"</savetime>\n")
f.write(" <progress>completed</progress>\n")
f.write("</info>"+"\n\n\n")
f.write("<schema path=\"./\" protocal=\"file\">temporal-schema.xml</schema>\n\n\n")
f.write("<annotations>\n\n\n")
count=0
for i, span_label in enumerate(predict):
if span_label == 1:
totalCorrEvents += 1
f.write("\t<entity>\n")
f.write("\t\t<id>"+str(count)+"@"+fn+"@system"+"</id>\n")
f.write("\t\t<span>"+str(spans[i][0])+","+str(spans[i][1])+"</span>\n")
f.write("\t\t<type>EVENT</type>\n")
f.write("\t\t<parentsType></parentsType>\n")
f.write("\t\t<properties>\n")
f.write("\t\t\t<DocTimeRel>BEFORE</DocTimeRel>\n")
f.write("\t\t\t<Type>N/A</Type>\n")
f.write("\t\t\t<Degree>N/A</Degree>\n")
f.write("\t\t\t<Polarity>"+"POS"+"</Polarity>\n")
f.write("\t\t\t<ContextualModality>"+"ACTUAL"+"</ContextualModality>\n")
f.write("\t\t\t<ContextualAspect>N/A</ContextualAspect>\n")
f.write("\t\t\t<Permanence>UNDETERMINED</Permanence>\n")
f.write("\t\t</properties>\n")
f.write("\t</entity>\n\n")
count += 1
f.write("\n\n</annotations>\n")
f.write("</data>")
pbar.finish()
print "Total pred events is %d"%totalPredEvents
print "Total corr events is %d"%totalCorrEvents