-
Notifications
You must be signed in to change notification settings - Fork 1
/
predict.py
464 lines (435 loc) · 21.5 KB
/
predict.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
from utils import *
import copy
import numpy as np
class Predict(object):
def build_sampler(self, layers, tparams, options, use_noise, trng, word2vec_Wemb = None):
debug_print = []
#debug_print.append( theano.printing.Print('input_p.shapa')(input_p.shape))
# context: #annotations x dim
ctx0 = T.matrix('ctx_sampler', dtype='float32')
ctx_mask = T.vector('ctx_mask', dtype='float32')
ctx_ = ctx0
counts = ctx_mask.sum(-1)
ctx_mean = ctx_.sum(0)/counts
ctx_max = ctx_.max(0)
ctx_ = ctx0.dimshuffle('x',0,1)
ctx_input = ctx_mean
# initial state/cell
tu_init_state = [T.alloc(0., options['rnn_word_dim'])]
tu_init_memory = [T.alloc(0., options['rnn_word_dim'])]
mu_init_state = [T.alloc(0., options['rnn_cond_wv_dim'])]
mu_init_memory = [T.alloc(0., options['rnn_cond_wv_dim'])]
'''
if options['smoothing'] :
a_init_state = [T.alloc(0., options['latent_size_a'])]
#a_init_memory = [T.alloc(0., options['latent_size_a'])]
else :
a_init_state = None
'''
z_init_state = [T.alloc(0., options['latent_size_z'])]
mu_p_init = [T.alloc(0., options['latent_size_z'])]
print 'Building f_init...',
'''
f_init = theano.function([ctx0, ctx_mask], [ctx0]+tu_init_state+tu_init_memory+
mu_init_state+mu_init_memory+
a_init_state+a_init_memory+
z_init_state+
mu_p_init, name='f_init',
on_unused_input='ignore',
profile=False)
'''
f_init = theano.function([ctx0, ctx_mask], [ctx0]+tu_init_state+tu_init_memory+
mu_init_state+mu_init_memory+
z_init_state+
mu_p_init, name='f_init',
on_unused_input='ignore',
profile=False)
print 'Done'
x = T.vector('x_sampler', dtype='int64')
tu_init_state = [T.matrix('tu_init_state', dtype='float32')]
tu_init_memory = [T.matrix('tu_init_memory', dtype='float32')]
mu_init_state = [T.matrix('mu_init_state', dtype='float32')]
mu_init_memory = [T.matrix('mu_init_memory', dtype='float32')]
'''
if options['smoothing'] :
a_init_state = [T.matrix('a_init_state', dtype='float32')]
#a_init_memory = [T.matrix('a_init_memory', dtype='float32')]
'''
# if it's the first word, emb should be all zero
if options['word2vec']:
emb = T.switch(x[:, None] < 0,
T.alloc(0., 1, word2vec_Wemb.shape[1]), word2vec_Wemb[x])
else :
emb = T.switch(x[:, None] < 0,
T.alloc(0., 1, tparams['Wemb'].shape[1]), tparams['Wemb'][x])
# emb ff
emb_ff1 = layers.get_layer('ff')[1](tparams, emb,activ=options['nonlin_decoder'],
prefix="emb_ff1")
#emb_ff2 = layers.get_layer('ff')[1](tparams, emb_ff1,activ=options['nonlin_decoder'],
# prefix='emb_ff2')
emb_drop = layers.dropout_layer(emb_ff1, use_noise, trng)
tu_gru = layers.get_layer('lstm')[1](options,tparams, emb, one_step=True,
init_state=tu_init_state[0],
init_memory=tu_init_memory[0],
prefix='tu_rnn')
#debug_print.append( theano.printing.Print('mu_init_state.shapa')(mu_init_state.shape))
if options['att_fun'] == None:
mu_gru = layers.get_layer('lstm_cond')[1](options, tparams, tu_gru[0],
mask=None, context=ctx_input,
one_step=True,
init_state=mu_init_state[0],
init_memory=mu_init_memory[0],
trng=trng,
use_noise=use_noise,
prefix='mu_rnn')
else :
mu_gru = layers.get_layer('lstm_att')[1](options, tparams, tu_gru[0],
mask=None, context=ctx_,
one_step=True,
init_state=mu_init_state[0],
init_memory=mu_init_memory[0],
trng=trng,
use_noise=use_noise,
prefix='mu_rnn')
tu_next_state = [tu_gru[0]]
tu_next_memory = [tu_gru[1]]
mu_next_state = [mu_gru[0]]
mu_next_memory = [mu_gru[1]]
proj_h = mu_gru[0]
d_layer = proj_h
if options['use_dropout']:
d_drop_layer = layers.dropout_layer(d_layer, use_noise, trng)
'''
input_a_layer = T.concatenate([d_drop_layer, emb_drop], axis=1)
if options['smoothing']:
a_layer = layers.get_layer('gru_cond')[1](options, tparams, input_a_layer,one_step=True,
init_state=a_init_state[0],context=ctx_input,
prefix='a_rnn')
#a_layer = layers.get_layer('lstm')[1](options, tparams, input_a_layer,one_step=True,
# init_state=a_init_state[0],init_memory=a_init_memory[0]
# prefix='a_rnn')
#a_layer = a_layer[:, ::-1]
a_next_state = [a_layer[0]]
#a_next_memory = [a_layer[1]]
input_a = a_layer[0]
else:
temp_a = layers.get_layer('ff')[1](options, tparams, input_a_layer,
prefix='a_layer_0')
for i in range(options['flat_mlp_num'] - 1):
temp_a = layers.get_layer('ff')[1](options, tparams, temp_a,
prefix='a_layer_' + str(i + 1))
a_layer = temp_a
input_a = a_layer
#debug_print.append( theano.printing.Print('a_layer.shapa')(a_layer.shape))
'''
#################
###stochastic parts####
#################
# Define shared variables for quantities to be updated across batches (truncated BPTT)
z_init = [T.matrix('z', dtype='float32')]
mu_p_init = [T.matrix('mu_p_init', dtype='float32')]
stochastic_layer = layers.stochastic_layer_onestep_noq(options,tparams,
input_p=d_drop_layer,#input_q=input_a,
z_init=z_init[0],mu_p_init=mu_p_init[0],
num_units=options['latent_size_z'],
unroll_scan=options['unroll_scan'],
use_mu_residual_q=options['use_mu_residual_q']
)
z_layer = [stochastic_layer[0]]
mean_prior_layer = [stochastic_layer[1]]
log_var_prior_layer = stochastic_layer[2]
'''
mean_q_layer = stochastic_layer[3]
log_var_q_layer = stochastic_layer[4]
'''
z_dropout_layer = layers.dropout_layer(z_layer[0], use_noise, trng)
'''
z_layer_shp = z_dropout_layer.shape
z_layer_reshaped = z_dropout_layer.reshape([z_layer_shp[0]*z_layer_shp[1],
z_layer_shp[2]])
d_layer_shp = d_drop_layer.shape
d_layer_reshaped = d_drop_layer.reshape([d_layer_shp[0]*d_layer_shp[1],
d_layer_shp[2]])
'''
input_gen_ff = T.concatenate([d_drop_layer, z_dropout_layer], axis=1)
'''
gen_word_emb_ff = layers.get_layer('ff')[1](tparams, input_gen_ff, activ=options['nonlin_decoder'],
prefix='gen_word_emb_ff')
'''
logit = layers.get_layer('ff')[1](tparams, input_gen_ff,
prefix='ff_logit_zd', activ='linear')
# logit_shp = logit.shape
next_probs = T.nnet.softmax(logit)
next_sample = trng.multinomial(pvals=next_probs).argmax(1)
# next word probability
print 'building f_next...'
'''
f_next = theano.function([x, ctx0, ctx_mask]+
tu_init_state+tu_init_memory+
mu_init_state+mu_init_memory+
a_init_state+a_init_memory
z_init+
mu_p_init,
[next_probs, next_sample]+
tu_next_state+tu_next_memory+
mu_next_state+mu_next_memory+
a_next_state+a_next_memory+
z_layer+
mean_prior_layer,
name='f_next', profile=False,
on_unused_input='ignore')
'''
f_next = theano.function([x, ctx0, ctx_mask]+
tu_init_state+tu_init_memory+
mu_init_state+mu_init_memory+
z_init+
mu_p_init,
[next_probs, next_sample]+
tu_next_state+tu_next_memory+
mu_next_state+mu_next_memory+
z_layer+
mean_prior_layer,
name='f_next', profile=False,
on_unused_input='ignore')
print 'Done'
return f_init, f_next
def gen_sample(self, tparams, f_init, f_next, ctx0, ctx_mask,
trng=None, k=1, maxlen=30, stochastic=False):
'''
ctx0: (26,1024)
ctx_mask: (26,)
'''
if k > 1:
assert not stochastic, 'Beam search does not support stochastic sampling'
sample = []
z_res = []
p_mean = []
sample_score = []
if stochastic:
sample_score = 0
live_k = 1
dead_k = 0
hyp_samples = [[]] * live_k
hyp_z_res = [[]]* live_k
hyp_p_meam= [[]]* live_k
hyp_scores = np.zeros(live_k).astype('float32')
# [(26,1024),(512,),(512,)]
rval = f_init(ctx0, ctx_mask)
ctx0 = rval[0]
# next gru and stacked gru state and memory
next_states = []
next_memorys = []
n_layers_rnn = 2
n_rnn_return = 2
for lidx in xrange(n_layers_rnn):
next_states.append([])
next_memorys.append([])
next_states[lidx].append(rval[n_rnn_return*lidx+1])
next_states[lidx][-1] = next_states[lidx][-1].reshape([live_k, next_states[lidx][-1].shape[0]])
next_memorys[lidx].append(rval[n_rnn_return*lidx+2])
next_memorys[lidx][-1] = next_memorys[lidx][-1].reshape([live_k, next_memorys[lidx][-1].shape[0]])
#print "init gru state shape is ",len(next_states),',',len(next_states[0])
'''
next_a_state = []
next_a_state.append([])
next_a_state[0].append(rval[-3])
next_a_state = []
next_a_state.append([])
next_a_state[0].append(rval[-4])
next_a_memory = []
next_a_memory.append([])
next_a_memory[0].append(rval[-3])
'''
next_z = []
next_z.append([])
next_z[0].append(rval[-2])
next_mu_p = []
next_mu_p.append([])
next_mu_p[0].append(rval[-1])
#print "init next_mu_p shape is ",len(next_mu_p),',',len(next_mu_p[0]),','
next_w = -1 * np.ones((1,)).astype('int64')
# next_state: [(1,512)]
# next_memory: [(1,512)]
for ii in xrange(maxlen):
# return [(1, 50000), (1,), (1, 512), (1, 512)]
# next_w: vector
# ctx: matrix
# ctx_mask: vector
# next_state: [matrix]
# next_memory: [matrix]
#print "next_states ", len(next_states),',',len(next_states[1]),',',len(next_states[1][0]),',',len(next_states[1][0][0])
rval = f_next(*([next_w, ctx0, ctx_mask] +
next_states[0] + next_memorys[0] +
next_states[1] + next_memorys[1] +
next_z +
next_mu_p))
next_p = rval[0]
next_w = rval[1] # already argmax sorted
next_states = []
next_memorys = []
for lidx in xrange(n_layers_rnn):
next_states.append([])
next_memorys.append([])
next_states[lidx].append(rval[n_rnn_return*lidx+2])
next_memorys[lidx].append(rval[n_rnn_return*lidx+3])
#print "gru state is ", len(next_states),',',len(next_states[0]),',',len(next_states[0][0])
'''
next_a_state = [rval[-3]]
next_a_state = [rval[-4]]
next_a_memory = [rval[-3]]
'''
next_z = [rval[-2]]
next_mu_p = [rval[-1]]
#print "init next_a shape is ",len(next_a),',',len(next_a[0]),','
#print "init next_mu_p shape is ",len(next_mu_p),',',len(next_mu_p[0]),','
if stochastic:
sample.append(next_w[0]) # take the most likely one
sample_score += next_p[0,next_w[0]]
z_layer.append(next_z[0])
if next_w[0] == 0:
break
else:
# the first run is (1,50000)
cand_scores = hyp_scores[:,None] - np.log(next_p)
cand_flat = cand_scores.flatten()
ranks_flat = cand_flat.argsort()[:(k-dead_k)]
voc_size = next_p.shape[1]
trans_indices = ranks_flat / voc_size # index of row
word_indices = ranks_flat % voc_size # index of col
costs = cand_flat[ranks_flat]
new_hyp_samples = []
new_hyp_scores = np.zeros(k-dead_k).astype('float32')
new_hyp_z_res = []
new_hyp_p_mean=[]
new_hyp_states = []
new_hyp_memories = []
#new_hyp_a_state = []
#new_hyp_a_state.append([])
#new_hyp_a_memory = []
#new_hyp_a_memory.append([])
new_hyp_z = []
new_hyp_z.append([])
new_hyp_mu_p = []
new_hyp_mu_p.append([])
for lidx in xrange(n_layers_rnn):
new_hyp_states.append([])
new_hyp_memories.append([])
for idx, [ti, wi] in enumerate(zip(trans_indices, word_indices)):
new_hyp_samples.append(hyp_samples[ti]+[wi])
new_hyp_z_res.append(hyp_z_res[ti]+[next_z[0][ti]])
new_hyp_p_mean.append(hyp_p_meam[ti]+[next_mu_p[0][ti]])
new_hyp_scores[idx] = copy.copy(costs[idx])
for lidx in np.arange(n_layers_rnn):
new_hyp_states[lidx].append(copy.copy(next_states[lidx][0][ti]))
new_hyp_memories[lidx].append(copy.copy(next_memorys[lidx][0][ti]))
#new_hyp_a_state[0].append( copy.copy(next_a_state[0][ti]))
#new_hyp_a_memory[0].append( copy.copy(next_a_memory[0][ti]))
new_hyp_z[0].append(copy.copy(next_z[0][ti]))
new_hyp_mu_p[0].append(copy.copy(next_mu_p[0][ti]))
#print "init new_hyp_states shape is ",len(new_hyp_states),',',len(new_hyp_states[0]),','
#print "init new_hyp_mu_p shape is ",len(new_hyp_mu_p),',',len(new_hyp_mu_p[0]),','
# check the finished samples
new_live_k = 0
hyp_samples = []
hyp_z_res = []
hyp_p_meam = []
hyp_scores = []
hyp_states = []
hyp_a_state = []
hyp_a_state.append([])
hyp_a_memory = []
hyp_a_memory.append([])
hyp_z = []
hyp_z.append([])
hyp_mu_p = []
hyp_mu_p.append([])
hyp_memories = []
for lidx in xrange(n_layers_rnn):
hyp_states.append([])
hyp_memories.append([])
for idx in xrange(len(new_hyp_samples)):
if new_hyp_samples[idx][-1] == 0:
sample.append(new_hyp_samples[idx])
sample_score.append(new_hyp_scores[idx])
z_res.append(new_hyp_z_res[idx])
p_mean.append(new_hyp_p_mean[idx])
dead_k += 1
else:
new_live_k += 1
hyp_samples.append(new_hyp_samples[idx])
hyp_z_res.append(new_hyp_z_res[idx])
hyp_p_meam.append(new_hyp_p_mean[idx])
hyp_scores.append(new_hyp_scores[idx])
for lidx in xrange(n_layers_rnn):
hyp_states[lidx].append(new_hyp_states[lidx][idx])
hyp_memories[lidx].append(new_hyp_memories[lidx][idx])
#hyp_a_state[0].append(new_hyp_a_state[0][idx])
#hyp_a_memory[0].append(new_hyp_a_memory[0][idx])
hyp_z[0].append(new_hyp_z[0][idx])
hyp_mu_p[0].append(new_hyp_mu_p[0][idx])
#print "init hyp_states shape is ",len(hyp_states),',',len(hyp_states[0]),','
#print "init hyp_mu_p shape is ",len(hyp_mu_p),',',len(hyp_mu_p[0]),','
hyp_scores = np.array(hyp_scores)
live_k = new_live_k
if new_live_k < 1:
break
if dead_k >= k:
break
next_w = np.array([w[-1] for w in hyp_samples])
next_states = []
next_memorys = []
for lidx in xrange(n_layers_rnn):
next_states.append([])
next_memorys.append([])
next_states[lidx].append(np.array(hyp_states[lidx]))
next_memorys[lidx].append(np.array(hyp_memories[lidx]))
#next_a_state=hyp_a_state
#next_a_memory=hyp_a_memory
next_z = hyp_z
#z_layer.append(next_z)
next_mu_p = hyp_mu_p
#print "init next_states shape is ",len(next_states),',',len(next_states[0]),',',len(next_states[0][0])
#print "init next_mu_p shape is ",len(next_mu_p),',',len(next_mu_p[0]),','
if not stochastic:
# dump every remaining one
if live_k > 0:
for idx in xrange(live_k):
sample.append(hyp_samples[idx])
sample_score.append(hyp_scores[idx])
z_res.append([hyp_z[0][idx]])
p_mean.append([hyp_mu_p[0][idx]])
'''
for i in np.arange(len(sample)):
length = len(sample[i])
sample_score[i] = sample_score[i]/(1.0 * length)
'''
return sample, sample_score, next_states, z_res,p_mean
def sample_execute(self, engine, options, tparams, f_init, f_next, x, ctx, mask_ctx, trng):
stochastic = False
for jj in xrange(np.minimum(10, x.shape[1])):
sample, score, _,_,_ = self.gen_sample(tparams, f_init, f_next, ctx[jj], mask_ctx[jj],
trng=trng, k=5, maxlen=30, stochastic=stochastic)
if not stochastic:
best_one = np.argmin(score)
sample = sample[best_one]
else:
sample = sample
print 'Truth ', jj, ': ',
for vv in x[:, jj]:
if vv == 0:
break
if vv in engine.ix_word:
print engine.ix_word[vv],
else:
print 'UNK',
print
for kk, ss in enumerate([sample]):
print 'Sample (', jj, ') ', ': ',
for vv in ss:
if vv == 0:
break
if vv in engine.ix_word:
print engine.ix_word[vv],
else:
print 'UNK',
print