-
Notifications
You must be signed in to change notification settings - Fork 0
/
qrnn.py
434 lines (396 loc) · 20.2 KB
/
qrnn.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
'''QRNN class and functions for 2l-dr: headline generation (take 2)
implements all the layer functions and operations
for a Quasi-RNN https://arxiv.org/pdf/1611.01576.pdf
Also implements the seq2seq function for tf.nn.model_with_buckets()
There are some eval_* variants of some functions, which were written to run
the decode step during training.
'''
import tensorflow as tf
from tensorflow.python.util import nest
class QRNN(object):
def _init_vars(self):
''' initialize tf vars. i feel like this is an incorrect use of
scoping but i couldn't really figure out how else to do it '''
for i in xrange(self.num_layers):
input_shape = self.embedding_size if i == 0 else \
self.num_convs
with tf.variable_scope("QRNN/"+self.name +
"/Variable/Convolution/"+str(i),
reuse=False):
filter_shape = self._get_filter_shape(input_shape)
tf.get_variable('W', filter_shape,
initializer=self.initializer, dtype=tf.float32)
tf.get_variable('b', [self.num_convs*3],
initializer=self.initializer, dtype=tf.float32)
with tf.variable_scope("QRNN/"+self.name +
"/Variable/Conv_w_enc_out/"+str(i),
reuse=False):
v_shape = (self.num_convs, self.num_convs*3)
tf.get_variable('V', v_shape,
initializer=self.initializer, dtype=tf.float32)
tf.get_variable('b', [self.num_convs*3],
initializer=self.initializer, dtype=tf.float32)
filter_shape = self._get_filter_shape(input_shape)
tf.get_variable('W', filter_shape,
initializer=self.initializer, dtype=tf.float32)
with tf.variable_scope('QRNN/'+self.name +
'/Conv_with_attention/', reuse=False):
attn_weight_shape = [self.num_convs, self.num_convs]
tf.get_variable('W_k', attn_weight_shape,
initializer=self.initializer, dtype=tf.float32)
tf.get_variable('W_c', attn_weight_shape,
initializer=self.initializer, dtype=tf.float32)
tf.get_variable('b_o', [self.num_convs],
initializer=self.initializer, dtype=tf.float32)
def __init__(self, num_symbols, seq_length,
embedding_size, num_layers, conv_size, num_convs,
output_projection=None, name=''):
''' init qrnn class '''
self.num_symbols = num_symbols
self.seq_length = seq_length
self.embedding_size = embedding_size
self.num_layers = num_layers
self.conv_size = conv_size
self.num_convs = num_convs
self.output_projection = output_projection
self.initializer = tf.random_normal_initializer()
self.name = name
self._init_vars()
def get_embeddings(self, embeddings, word_ids):
''' get word embeddings '''
if word_ids is None:
return None
return tf.nn.embedding_lookup(embeddings, word_ids)
def fo_pool(self, Z, F, O, seq_len=None, c_prev=None):
''' fo-pooling function defined in Bradbury et al. on QRNNs
very reminiscent of LSTM gates'''
if seq_len is None:
seq_len = self.seq_length
# Z, F, O dims: [batch_size, sequence_length, num_convs]
H = [tf.fill(tf.pack([tf.shape(Z)[0], tf.shape(Z)[2]]), 0.0)]
if c_prev is not None:
C = [c_prev]
else:
C = [tf.fill(tf.pack([tf.shape(Z)[0], tf.shape(Z)[2]]), 0.0)]
# recurrent definition, must be computed one timestep at a time
for i in range(1, seq_len):
c_i = tf.mul(F[:, i, :], C[-1]) + \
tf.mul(1-F[:, i, :], Z[:, i, :])
# C[:, i, :] = c_i
C.append(c_i)
h_i = tf.mul(O[:, i, :], c_i)
# H[:, i, :] = h_i
H.append(tf.squeeze(h_i))
# i think we want output [batch, seq_len, num_convs]
return tf.reshape(tf.pack(H), tf.shape(Z)), C[-1]
def eval_fo_pool(self, Z, F, O, seq_len, c_prev=None):
''' fo-pool variant for use during evaluation '''
# Z, F, O dims: [batch_size, sequence_length, num_convs]
H = []
C = [c_prev]
for i in range(0, seq_len):
c_i = tf.mul(F[:, i, :], C[-1]) + \
tf.mul(1-F[:, i, :], Z[:, i, :])
# C[:, i, :] = c_i
C.append(c_i)
h_i = tf.mul(O[:, i, :], c_i)
# H[:, i, :] = h_i
H.append(tf.squeeze(h_i))
# i think we want output [batch, seq_len, num_convs]
return tf.reshape(tf.pack(H), tf.shape(Z)), C[-1]
# def f_pool(self, Z, F, sequence_length):
# # Z, F dims: [batch_size, sequence_length, num_convs]
# H = tf.fill(tf.shape(Z), 0)
# for i in range(1, self.seq_length):
# H[:, i, :] = tf.mul(F[:, i, :], H[:, i-1, :]) + \
# tf.mul(1-F[:, i, :])
# return np.array(H)
def _get_filter_shape(self, input_shape):
''' set up dimensions for convolution filter '''
return [self.conv_size, input_shape, 1, self.num_convs*3]
# convolution dimension results maths
# out_height = ceil(float(in_height - filter_height + 1) /
# float(strides[1])) = sequence_length
# out_width = ceil(float(in_width - filter_width + 1) /
# float(strides[2])) = 1
# in_height = sequence_length + filter_height - 1
# filter_height = conv_size
# in_width = embedding_size
# filter_width = embedding_size
def conv_layer(self, layer_id, inputs, input_shape, center_conv=False):
''' execute a convolution over inputs. default is to used a masked
convolution. '''
with tf.variable_scope("QRNN/"+self.name +
"/Variable/Convolution/"+str(layer_id),
reuse=True):
filter_shape = self._get_filter_shape(input_shape)
W = tf.get_variable('W', filter_shape,
initializer=self.initializer, dtype=tf.float32)
b = tf.get_variable('b', [self.num_convs*3],
initializer=self.initializer, dtype=tf.float32)
if not center_conv:
num_pads = self.conv_size - 1
# input dims ~should~ now be [batch_size, sequence_length,
# embedding_size, 1]
padded_input = tf.pad(tf.expand_dims(inputs, -1),
[[0, 0], [num_pads, 0],
[0, 0], [0, 0]],
"CONSTANT")
else:
assert self.conv_size % 2 == 1
num_pads = (self.conv_size - 1) / 2
padded_input = tf.pad(tf.expand_dims(inputs, -1),
[[0, 0], [num_pads, num_pads],
[0, 0], [0, 0]],
"CONSTANT")
conv = tf.nn.conv2d(
padded_input,
W,
strides=[1, 1, 1, 1],
padding="VALID",
name="conv") + b
# conv dims: [batch_size, sequence_length,
# 1, num_convs*3]
# squeeze out 3rd D
# split 4th (now 3rd) dim into 3
Z, F, O = tf.split(2, 3, tf.squeeze(conv, [2]))
return self.fo_pool((tf.tanh(Z)), tf.sigmoid(F), tf.sigmoid(O))
def conv_with_encode_output(self, layer_id, h_t, inputs,
input_shape, pool=True,
seq_len=None):
''' execute a convolution, also feeding in a previous
output before the pooling step.
option to disable pooling: used in conv_with_attention'''
if seq_len is None:
seq_len = self.seq_length
pooling = self.fo_pool if pool else lambda x, y, z, seq_len: (x, y, z)
with tf.variable_scope("QRNN/"+self.name +
"/Variable/Conv_w_enc_out/"+str(layer_id),
reuse=True):
v_shape = (self.num_convs, self.num_convs*3)
V = tf.get_variable('V', v_shape,
initializer=self.initializer, dtype=tf.float32)
b = tf.get_variable('b', [self.num_convs*3],
initializer=self.initializer, dtype=tf.float32)
filter_shape = self._get_filter_shape(input_shape)
W = tf.get_variable('W', filter_shape,
initializer=self.initializer, dtype=tf.float32)
num_pads = self.conv_size - 1
h_tV = tf.matmul(h_t, V)
Z_v, F_v, O_v = tf.split(1, 3, h_tV)
# input dims ~should~ now be [batch_size, sequence_length,
# embedding_size, 1]
padded_input = tf.pad(tf.expand_dims(inputs, -1),
[[0, 0], [num_pads, 0],
[0, 0], [0, 0]],
"CONSTANT")
conv = tf.nn.conv2d(
padded_input,
W,
strides=[1, 1, 1, 1],
padding="VALID",
name="conv") + b
# conv dims: [batch_size, sequence_length,
# 1, num_convs*3]
# squeeze out 3rd D
# split 4th (now 3rd) dim into 3
Z_conv, F_conv, O_conv = tf.split(2, 3, tf.squeeze(conv))
Z = Z_conv + tf.expand_dims(Z_v, 1)
F = F_conv + tf.expand_dims(F_v, 1)
O = O_conv + tf.expand_dims(O_v, 1)
return pooling(tf.tanh(Z), tf.sigmoid(F), tf.sigmoid(O), seq_len)
def conv_with_attention(self, layer_id, encode_outputs, inputs,
input_shape, seq_len=None):
''' perform a convolution step with soft attention '''
if seq_len is None:
seq_len = self.seq_length
h_t = tf.squeeze(encode_outputs[-1][:, -1, :])
Z, F, O = self.conv_with_encode_output(layer_id, h_t, inputs,
input_shape, pool=False)
# input dim [batch, seq_len, num_convs]
with tf.variable_scope('QRNN/'+self.name +
'/Conv_with_attention/', reuse=True):
attn_weight_shape = [self.num_convs, self.num_convs]
W_k = tf.get_variable('W_k', attn_weight_shape,
initializer=self.initializer,
dtype=tf.float32)
W_c = tf.get_variable('W_c', attn_weight_shape,
initializer=self.initializer,
dtype=tf.float32)
b_o = tf.get_variable('b_o', [self.num_convs],
initializer=self.initializer,
dtype=tf.float32)
# calculate attention
enc_final_state = encode_outputs[-1]
H = [tf.fill(tf.pack([tf.shape(Z)[0], tf.shape(Z)[2]]), 0.0)]
C = [tf.fill(tf.pack([tf.shape(Z)[0], tf.shape(Z)[2]]), 0.0)]
for i in range(1, seq_len):
c_i = tf.mul(F[:, i, :], C[-1]) + \
tf.mul(1-F[:, i, :], Z[:, i, :])
C.append(c_i)
# C_i dim [batch, num_convs]
# enc_final_state dim [batch, seq_len, num_convs]
c_dot_h = tf.reduce_sum(tf.mul(tf.expand_dims(c_i, 1),
enc_final_state), axis=2)
# alpha dim [batch, seq_len]
alpha = tf.nn.softmax(c_dot_h)
k_t = tf.mul(tf.expand_dims(alpha, -1), enc_final_state)
x = tf.matmul(tf.reshape(k_t, [-1, self.num_convs]), W_k)
x2 = tf.reduce_sum(tf.reshape(x, tf.shape(k_t)), axis=1)
y = tf.matmul(c_i, W_c)+b_o
h_i = tf.mul(O[:, i, :], x2+y)
H.append(tf.squeeze(h_i))
return tf.reshape(tf.pack(H), tf.shape(Z)), C[-1]
# def transform_output(self, inputs):
# # input dim list of [batch, num_convs]
# shape = (self.num_convs, self.num_symbols)
# with tf.variable_scope('QRNN/'+self.name+'/Transform_output'):
# W = tf.get_variable('W', shape,
# initializer=self.initializer,
# dtype=tf.float32)
# b = tf.get_variable('b', [self.num_symbols],
# initializer=self.initializer, d
# type=tf.float32)
# # TODO: do efficiently
# result = []
# for i in inputs:
# result.append(tf.nn.xw_plus_b(i, W, b))
# return result
def eval_conv_with_encode_output(self, layer_id, h_t, inputs,
input_shape, c_prev, pool=True):
seq_len = self.conv_size
pooling = self.eval_fo_pool if pool else \
lambda v, w, x, y, z: (v, w, x)
with tf.variable_scope("QRNN/"+self.name +
"/Variable/Conv_w_enc_out/"+str(layer_id),
reuse=True):
v_shape = (self.num_convs, self.num_convs*3)
V = tf.get_variable('V', v_shape,
initializer=self.initializer, dtype=tf.float32)
b = tf.get_variable('b', [self.num_convs*3],
initializer=self.initializer, dtype=tf.float32)
filter_shape = self._get_filter_shape(input_shape)
W = tf.get_variable('W', filter_shape,
initializer=self.initializer, dtype=tf.float32)
h_tV = tf.matmul(h_t, V)
Z_v, F_v, O_v = tf.split(1, 3, h_tV)
conv = tf.nn.conv2d(
tf.expand_dims(inputs, -1),
W,
strides=[1, 1, 1, 1],
padding="VALID",
name="conv") + b
# conv dims: [batch_size, sequence_length,
# 1, num_convs*3]
# squeeze out 3rd D
# split 4th (now 3rd) dim into 3
Z_conv, F_conv, O_conv = tf.split(2, 3, tf.squeeze(conv, [2]))
Z = Z_conv + tf.expand_dims(Z_v, 1)
F = F_conv + tf.expand_dims(F_v, 1)
O = O_conv + tf.expand_dims(O_v, 1)
return pooling(tf.tanh(Z), tf.sigmoid(F),
tf.sigmoid(O), seq_len-self.conv_size+1, c_prev)
def eval_conv_with_attention(self, layer_id, encode_outputs, inputs,
input_shape, c_prev):
seq_len = self.conv_size
h_t = tf.squeeze(encode_outputs[-1][:, -1, :])
Z, F, O = self.eval_conv_with_encode_output(layer_id, h_t, inputs,
input_shape, c_prev,
pool=False)
# input dim [batch, seq_len, num_convs]
with tf.variable_scope('QRNN/'+self.name +
'/Conv_with_attention/', reuse=True):
attn_weight_shape = [self.num_convs, self.num_convs]
W_k = tf.get_variable('W_k', attn_weight_shape,
initializer=self.initializer,
dtype=tf.float32)
W_c = tf.get_variable('W_c', attn_weight_shape,
initializer=self.initializer,
dtype=tf.float32)
b_o = tf.get_variable('b_o', [self.num_convs],
initializer=self.initializer,
dtype=tf.float32)
# calculate attention
enc_final_state = encode_outputs[-1]
H = []
C = [c_prev]
for i in range(0, seq_len-self.conv_size+1):
c_i = tf.mul(F[:, i, :], C[-1]) + \
tf.mul(1-F[:, i, :], Z[:, i, :])
C.append(c_i)
# C_i dim [batch, num_convs]
# enc_final_state dim [batch, seq_len, num_convs]
c_dot_h = tf.reduce_sum(tf.mul(tf.expand_dims(c_i, 1),
enc_final_state), axis=2)
# alpha dim [batch, seq_len]
alpha = tf.nn.softmax(c_dot_h)
k_t = tf.mul(tf.expand_dims(alpha, -1), enc_final_state)
x = tf.matmul(tf.reshape(k_t, [-1, self.num_convs]), W_k)
x2 = tf.reduce_sum(tf.reshape(x, tf.shape(k_t)), axis=1)
y = tf.matmul(c_i, W_c)+b_o
h_i = tf.mul(O[:, i, :], x2+y)
H.append(tf.squeeze(h_i))
return tf.reshape(tf.pack(H), tf.shape(Z)), C[-1]
def init_encoder_and_decoder(num_encoder_symbols, num_decoder_symbols,
enc_seq_length, dec_seq_length,
embedding_size, num_layers, conv_size, num_convs,
output_projection):
encoder = QRNN(num_encoder_symbols, enc_seq_length,
embedding_size, num_layers, conv_size, num_convs, 'enc')
decoder = QRNN(num_decoder_symbols, dec_seq_length,
embedding_size, num_layers, conv_size, num_convs,
output_projection, 'dec')
return encoder, decoder
def seq2seq_f(encoder, decoder, encoder_inputs, decoder_inputs,
feed_prev, embeddings, cell, center_conv=False):
''' runs an encode-decode step for an QRNNenc + RNNdec model '''
# inputs are lists of placeholders, each one is shape [None]
# self.enc_input_size = len(encoder_inputs)
# self.dec_input_size = len(decoder_inputs)
encode_outputs = []
# pack inputs to be shape [sequence_length, batch_size]
encoder_inputs = tf.transpose(tf.pack(encoder_inputs))
# embed to be shape [batch_size, sequence_length, embed_size]
embedded_enc_inputs = encoder.get_embeddings(embeddings, encoder_inputs)
# encode with qrnn
for i in range(encoder.num_layers):
inputs = embedded_enc_inputs if i == 0 else encode_outputs[-1]
input_shape = encoder.embedding_size if i == 0 else encoder.num_convs
encode_outputs.append(encoder.conv_layer(i, inputs, input_shape,
center_conv)[0])
encoder_state = tuple([encode_outputs[i][:, -1, :]
for i in range(encoder.num_layers)])
encode_outputs = tf.concat(1, [tf.reverse(e, [False, True, False])
for e in encode_outputs])
# decode with rnn
def decode(feed_prev_bool):
reuse = None if feed_prev_bool else True
with tf.variable_scope(tf.get_variable_scope(), reuse=reuse):
loop_function = tf.nn.seq2seq._extract_argmax_and_embed(
embeddings,
decoder.output_projection,
True) if feed_prev_bool else None
embedded_dec_inputs = [tf.nn.embedding_lookup(embeddings, i)
for i in decoder_inputs]
outputs, state = tf.nn.seq2seq.attention_decoder(
embedded_dec_inputs,
encoder_state,
encode_outputs,
cell,
loop_function=loop_function)
state_list = [state]
if nest.is_sequence(state):
state_list = nest.flatten(state)
# tf.cond has to return a single value
return outputs + state_list
# we want to feed previous input in during testing
outputs_and_state = tf.cond(feed_prev,
lambda: decode(True),
lambda: decode(False))
outputs_len = len(decoder_inputs)
state_list = outputs_and_state[outputs_len:]
state = state_list[0]
if nest.is_sequence(encoder_state):
state = nest.pack_sequence_as(structure=encoder_state,
flat_sequence=state_list)
return outputs_and_state[:outputs_len], state