-
Notifications
You must be signed in to change notification settings - Fork 0
/
NARM.py
520 lines (389 loc) · 17.1 KB
/
NARM.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
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
'''
Build NARM model
'''
from __future__ import print_function
import pickle
from collections import OrderedDict
import sys
import time
import numpy as np
import theano
from theano import config
import theano.tensor as T
from theano.sandbox.rng_mrg import MRG_RandomStreams as RandomStreams
import data_process
datasets = {'rsc2015': (data_process.load_data, data_process.prepare_data)}
# Set the random number generators' seeds for consistency
SEED = 42
np.random.seed(SEED)
def numpy_floatX(data):
return np.asarray(data, dtype=config.floatX)
def get_minibatches_idx(n, minibatch_size, shuffle=False):
"""
Used to shuffle the dataset at each iteration.
"""
idx_list = np.arange(n, dtype="int32")
if shuffle:
np.random.shuffle(idx_list)
minibatches = []
minibatch_start = 0
for i in range(n // minibatch_size):
minibatches.append(idx_list[minibatch_start:
minibatch_start + minibatch_size])
minibatch_start += minibatch_size
if minibatch_start != n:
# Make a minibatch out of what is left
minibatches.append(idx_list[minibatch_start:])
return zip(range(len(minibatches)), minibatches)
def get_dataset(name):
return datasets[name][0], datasets[name][1]
def zipp(params, tparams):
"""
When we reload the model. Needed for the GPU stuff.
"""
for kk, vv in params.items():
tparams[kk].set_value(vv)
def unzip(zipped):
"""
When we pickle the model. Needed for the GPU stuff.
"""
new_params = OrderedDict()
for kk, vv in zipped.items():
new_params[kk] = vv.get_value()
return new_params
def dropout_layer(state_before, use_noise, trng, drop_p=0.5):
retain = 1. - drop_p
proj = T.switch(use_noise, (state_before * trng.binomial(state_before.shape,
p=retain, n=1,
dtype=state_before.dtype)), state_before * retain)
return proj
def _p(pp, name):
return '%s_%s' % (pp, name)
def init_params(options):
"""
Global (not GRU) parameter. For the embeding and the classifier.
"""
params = OrderedDict()
# embedding
params['Wemb'] = init_weights((options['n_items'], options['dim_proj']))
params = get_layer(options['encoder'])[0](options,
params,
prefix=options['encoder'])
# attention
params['W_encoder'] = init_weights((options['hidden_units'], options['hidden_units']))
params['W_decoder'] = init_weights((options['hidden_units'], options['hidden_units']))
params['bl_vector'] = init_weights((1, options['hidden_units']))
# classifier
# params['U'] = init_weights((2*options['hidden_units'], options['n_items']))
# params['b'] = np.zeros((options['n_items'],)).astype(config.floatX)
params['bili'] = init_weights((options['dim_proj'], 2 * options['hidden_units']))
return params
def load_params(path, params):
pp = np.load(path)
for kk, vv in params.items():
if kk not in pp:
raise Warning('%s is not in the archive' % kk)
params[kk] = pp[kk]
return params
def init_tparams(params):
tparams = OrderedDict()
for kk, pp in params.items():
tparams[kk] = theano.shared(params[kk], name=kk)
return tparams
def get_layer(name):
fns = layers[name]
return fns
def init_weights(shape):
sigma = np.sqrt(2. / shape[0])
return numpy_floatX(np.random.randn(*shape) * sigma)
def ortho_weight(ndim):
W = np.random.randn(ndim, ndim)
u, s, v = np.linalg.svd(W)
return u.astype(config.floatX)
def param_init_gru(options, params, prefix='gru'):
"""
Init the GRU parameter:
:see: init_params
"""
Wxrz = np.concatenate([init_weights((options['dim_proj'], options['hidden_units'])),
init_weights((options['dim_proj'], options['hidden_units'])),
init_weights((options['dim_proj'], options['hidden_units']))], axis=1)
params[_p(prefix, 'Wxrz')] = Wxrz
Urz = np.concatenate([ortho_weight(options['hidden_units']),
ortho_weight(options['hidden_units'])], axis=1)
params[_p(prefix, 'Urz')] = Urz
Uh = ortho_weight(options['hidden_units'])
params[_p(prefix, 'Uh')] = Uh
b = np.zeros((3 * options['hidden_units'],))
params[_p(prefix, 'b')] = b.astype(config.floatX)
return params
def gru_layer(tparams, state_below, options, prefix='gru', mask=None):
nsteps = state_below.shape[0]
if state_below.ndim == 3:
n_samples = state_below.shape[1]
else:
n_samples = 1
assert mask is not None
def _slice(_x, n, dim):
if _x.ndim == 3:
return _x[:, :, n * dim:(n + 1) * dim]
return _x[:, n * dim:(n + 1) * dim]
def _step(m_, x_, h_):
preact = T.dot(h_, tparams[_p(prefix, 'Urz')])
preact += x_[:, 0:2 * options['hidden_units']]
z = T.nnet.hard_sigmoid(_slice(preact, 0, options['hidden_units']))
r = T.nnet.hard_sigmoid(_slice(preact, 1, options['hidden_units']))
h = T.tanh(T.dot((h_ * r), tparams[_p(prefix, 'Uh')]) + _slice(x_, 2, options['hidden_units']))
h = (1.0 - z) * h_ + z * h
h = m_[:, None] * h + (1. - m_)[:, None] * h_
return h
state_below = (T.dot(state_below, tparams[_p(prefix, 'Wxrz')]) +
tparams[_p(prefix, 'b')])
hidden_units = options['hidden_units']
rval, updates = theano.scan(_step,
sequences=[mask, state_below],
outputs_info=T.alloc(numpy_floatX(0.), n_samples, hidden_units),
name=_p(prefix, '_layers'),
n_steps=nsteps)
return rval
layers = {'gru': (param_init_gru, gru_layer)}
def adam(loss, all_params, learning_rate=0.001, b1=0.9, b2=0.999, e=1e-8, gamma=1-1e-8):
"""
ADAM update rules
Default values are taken from [Kingma2014]
References:
[Kingma2014] Kingma, Diederik, and Jimmy Ba.
"Adam: A Method for Stochastic Optimization."
arXiv preprint arXiv:1412.6980 (2014).
http://arxiv.org/pdf/1412.6980v4.pdf
"""
updates = OrderedDict()
all_grads = theano.grad(loss, all_params)
alpha = learning_rate
t = theano.shared(np.float32(1))
b1_t = b1*gamma**(t-1) #(Decay the first moment running average coefficient)
for theta_previous, g in zip(all_params, all_grads):
m_previous = theano.shared(np.zeros(theta_previous.get_value().shape, dtype=config.floatX))
v_previous = theano.shared(np.zeros(theta_previous.get_value().shape, dtype=config.floatX))
m = b1_t*m_previous + (1 - b1_t)*g # (Update biased first moment estimate)
v = b2*v_previous + (1 - b2)*g**2 # (Update biased second raw moment estimate)
m_hat = m / (1-b1**t) # (Compute bias-corrected first moment estimate)
v_hat = v / (1-b2**t) # (Compute bias-corrected second raw moment estimate)
theta = theta_previous - (alpha * m_hat) / (T.sqrt(v_hat) + e) #(Update parameters)
updates[m_previous] = m
updates[v_previous] = v
updates[theta_previous] = theta
updates[t] = t + 1.
return updates
def build_model(tparams, options):
trng = RandomStreams(SEED)
# Used for dropout.
use_noise = theano.shared(numpy_floatX(0.))
x = T.matrix('x', dtype='int64')
mask = T.matrix('mask', dtype=config.floatX)
y = T.vector('y', dtype='int64')
n_timesteps = x.shape[0]
n_samples = x.shape[1]
emb = tparams['Wemb'][x.flatten()].reshape([n_timesteps,
n_samples,
options['dim_proj']])
if options['use_dropout']:
emb = dropout_layer(emb, use_noise, trng, drop_p=0.25)
proj = get_layer(options['encoder'])[1](tparams, emb, options,
prefix=options['encoder'],
mask=mask)
def compute_alpha(state1, state2):
tmp = T.nnet.hard_sigmoid(T.dot(tparams['W_encoder'], state1.T) + T.dot(tparams['W_decoder'], state2.T))
alpha = T.dot(tparams['bl_vector'], tmp)
res = T.sum(alpha, axis=0)
return res
last_h = proj[-1]
sim_matrix, _ = theano.scan(
fn=compute_alpha,
sequences=proj,
non_sequences=proj[-1]
)
att = T.nnet.softmax(sim_matrix.T * mask.T) * mask.T
p = att.sum(axis=1)[:, None]
weight = att / p
atttention_proj = (proj * weight.T[:, :, None]).sum(axis=0)
proj = T.concatenate([atttention_proj, last_h], axis=1)
if options['use_dropout']:
proj = dropout_layer(proj, use_noise, trng, drop_p=0.5)
ytem = T.dot(tparams['Wemb'], tparams['bili'])
pred = T.nnet.softmax(T.dot(proj, ytem.T))
# pred = T.nnet.softmax(T.dot(proj, tparams['U']) + tparams['b'])
f_pred_prob = theano.function([x, mask], pred, name='f_pred_prob')
# f_weight = theano.function([x, mask], weight, name='f_weight')
off = 1e-8
if pred.dtype == 'float16':
off = 1e-6
cost = -T.log(pred[T.arange(n_samples), y] + off).mean()
return use_noise, x, mask, y, f_pred_prob, cost
def pred_evaluation(f_pred_prob, prepare_data, data, iterator):
"""
Compute recall@20 and mrr@20
f_pred_prob: Theano fct computing the prediction
prepare_data: usual prepare_data for that dataset.
"""
recall = 0.0
mrr = 0.0
evalutation_point_count = 0
# pred_res = []
# att = []
for _, valid_index in iterator:
x, mask, y = prepare_data([data[0][t] for t in valid_index],
np.array(data[1])[valid_index])
preds = f_pred_prob(x, mask)
# weights = f_weight(x, mask)
targets = y
ranks = (preds.T > np.diag(preds.T[targets])).sum(axis=0) + 1
rank_ok = (ranks <= 20)
# pred_res += list(rank_ok)
recall += rank_ok.sum()
mrr += (1.0 / ranks[rank_ok]).sum()
evalutation_point_count += len(ranks)
# att.append(weights)
recall = numpy_floatX(recall) / evalutation_point_count
mrr = numpy_floatX(mrr) / evalutation_point_count
eval_score = (recall, mrr)
# ff = open('/storage/lijing/mydataset/res_attention_correct.pkl', 'wb')
# pickle.dump(pred_res, ff)
# ff.close()
# ff2 = open('/storage/lijing/mydataset/attention_weights.pkl', 'wb')
# pickle.dump(att, ff2)
# ff2.close()
return eval_score
def train_gru(
dim_proj=50, # word embeding dimension
hidden_units=100, # GRU number of hidden units.
patience=100, # Number of epoch to wait before early stop if no progress
max_epochs=30, # The maximum number of epoch to run
dispFreq=100, # Display to stdout the training progress every N updates
lrate=0.001, # Learning rate
n_items=37484, # Vocabulary size
encoder='gru', # TODO: can be removed must be gru.
saveto='gru_model.npz', # The best model will be saved there
is_valid=True, # Compute the validation error after this number of update.
is_save=False, # Save the parameters after every saveFreq updates
batch_size=512, # The batch size during training.
valid_batch_size=512, # The batch size used for validation/test set.
dataset='rsc2015',
# Parameter for extra option
use_dropout=True, # if False slightly faster, but worst test error
# This frequently need a bigger model.
reload_model=None, # Path to a saved model we want to start from.
test_size=-1, # If >0, we keep only this number of test example.
):
# Model options
model_options = locals().copy()
print("model options", model_options)
load_data, prepare_data = get_dataset(dataset)
print('Loading data')
train, valid, test = load_data()
print('Building model')
# This create the initial parameters as numpy ndarrays.
# Dict name (string) -> numpy ndarray
params = init_params(model_options)
if reload_model:
load_params('gru_model.npz', params)
# This create Theano Shared Variable from the parameters.
# Dict name (string) -> Theano Tensor Shared Variable
# params and tparams have different copy of the weights.
tparams = init_tparams(params)
# use_noise is for dropout
(use_noise, x, mask,
y, f_pred_prob, cost) = build_model(tparams, model_options)
all_params = list(tparams.values())
updates = adam(cost, all_params, lrate)
train_function = theano.function(inputs=[x, mask, y], outputs=cost, updates=updates)
print('Optimization')
kf_valid = get_minibatches_idx(len(valid[0]), valid_batch_size)
kf_test = get_minibatches_idx(len(test[0]), valid_batch_size)
print("%d train examples" % len(train[0]))
print("%d valid examples" % len(valid[0]))
print("%d test examples" % len(test[0]))
history_errs = []
history_vali = []
best_p = None
bad_count = 0
uidx = 0 # the number of update done
estop = False # early stop
try:
for eidx in range(max_epochs):
start_time = time.time()
n_samples = 0
epoch_loss = []
# Get new shuffled index for the training set.
kf = get_minibatches_idx(len(train[0]), batch_size, shuffle=True)
for _, train_index in kf:
uidx += 1
use_noise.set_value(1.)
# Select the random examples for this minibatch
y = [train[1][t] for t in train_index]
x = [train[0][t]for t in train_index]
# Get the data in numpy.ndarray format
# This swap the axis!
# Return something of shape (minibatch maxlen, n samples)
x, mask, y = prepare_data(x, y)
n_samples += x.shape[1]
loss = train_function(x, mask, y)
epoch_loss.append(loss)
if np.isnan(loss) or np.isinf(loss):
print('bad loss detected: ', loss)
return 1., 1., 1.
if np.mod(uidx, dispFreq) == 0:
print('Epoch ', eidx, 'Update ', uidx, 'Loss ', np.mean(epoch_loss))
if saveto and is_save:
print('Saving...')
if best_p is not None:
params = best_p
else:
params = unzip(tparams)
np.savez(saveto, history_errs=history_errs, **params)
print('Saving done')
if is_valid:
use_noise.set_value(0.)
valid_evaluation = pred_evaluation(f_pred_prob, prepare_data, valid, kf_valid)
test_evaluation = pred_evaluation(f_pred_prob, prepare_data, test, kf_test)
history_errs.append([valid_evaluation, test_evaluation])
if best_p is None or valid_evaluation[0] >= np.array(history_vali).max():
best_p = unzip(tparams)
print('Best perfomance updated!')
bad_count = 0
print('Valid Recall@20:', valid_evaluation[0], ' Valid Mrr@20:', valid_evaluation[1],
'\nTest Recall@20', test_evaluation[0], ' Test Mrr@20:', test_evaluation[1])
if len(history_vali) > 10 and valid_evaluation[0] <= np.array(history_vali).max():
bad_count += 1
print('===========================>Bad counter: ' + str(bad_count))
print('current validation recall: ' + str(valid_evaluation[0]) +
' history max recall:' + str(np.array(history_vali).max()))
if bad_count > patience:
print('Early Stop!')
estop = True
history_vali.append(valid_evaluation[0])
end_time = time.time()
print('Seen %d samples' % n_samples)
print(('This epoch took %.1fs' % (end_time - start_time)), file=sys.stderr)
if estop:
break
except KeyboardInterrupt:
print("Training interupted")
if best_p is not None:
zipp(best_p, tparams)
else:
best_p = unzip(tparams)
use_noise.set_value(0.)
valid_evaluation = pred_evaluation(f_pred_prob, prepare_data, valid, kf_valid)
test_evaluation = pred_evaluation(f_pred_prob, prepare_data, test, kf_test)
print('=================Best performance=================')
print('Valid Recall@20:', valid_evaluation[0], ' Valid Mrr@20:', valid_evaluation[1],
'\nTest Recall@20', test_evaluation[0], ' Test Mrr@20:', test_evaluation[1])
print('==================================================')
if saveto and is_save:
np.savez('Best_performance', valid_evaluation=valid_evaluation, test_evaluation=test_evaluation, history_errs=history_errs,
**best_p)
return valid_evaluation, test_evaluation
if __name__ == '__main__':
# See function train for all possible parameter and there definition.
eval_valid, eval_test = train_gru(max_epochs=30, test_size=-1)