-
Notifications
You must be signed in to change notification settings - Fork 3
/
Copy pathencoder_decoder_iter.py
193 lines (150 loc) · 8.09 KB
/
encoder_decoder_iter.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
import mxnet as mx
import numpy as np
from collections import namedtuple
from sklearn.cluster import KMeans
EncDecBucketKey = namedtuple('EncDecBucketKey', ['enc_len', 'dec_len'])
class EncoderDecoderBatch(object):
def __init__(self, all_data, all_mask, all_label, init_states, bucket_key):
# provide data, essential assignment
# need to optimize this list creation!!!
self.data = [ mx.nd.array(all_data), mx.nd.array(all_mask) ] + [ mx.nd.zeros(x[1]) for x in init_states ]
# provide label, essential assignment
self.label = [ mx.nd.array(all_label) ]
self.init_states = init_states
# bucket_key is essential for this databatch
self.bucket_key = bucket_key
#all_data.shape is (x,y,z)
self.batch_size = all_data.shape[0]
# this two properties are essential too!
@property
def provide_data(self):
return [
('data', (self.batch_size, self.bucket_key.enc_len + self.bucket_key.dec_len)),
('mask', (self.batch_size, self.bucket_key.enc_len + self.bucket_key.dec_len))
] + self.init_states
@property
def provide_label(self):
return [('label', (self.batch_size, self.bucket_key.dec_len))]
def synchronize_batch_size(train_iter, test_iter):
batch_size = min(train_iter.batch_size, test_iter.batch_size)
train_iter.batch_size = batch_size
test_iter.batch_size = batch_size
train_iter.generate_init_states()
test_iter.generate_init_states()
# now define the bucketing, padding and batching SequenceIterator...
class EncoderDecoderIter(mx.io.DataIter):
def __init__(self, data_label, word2idx, idx2word, num_hidden, num_layers,
init_states_function, batch_size=1, num_buckets=10, shuffle=False, rev=False):
super(EncoderDecoderIter, self).__init__() # calling DataIter.__init__()
# data is a numpy array of 3 dimensions, (#, timesteps, vector_dim)
# let's say you have 2 sequences, #1 has len 5, dimension 10
self.data_label = data_label #numpy multi-dimensional array
# data_label[i] is the ith sequence
self.word2idx = word2idx
self.idx2word = idx2word
self.num_hidden = num_hidden
self.num_layers = num_layers
self.num_buckets = num_buckets
# arrange the data so that
# now we need to find the buckets based on the input data...
self.buckets, self.buckets_count, self.assignments = self.generate_buckets()
# buckets are a tuple of the encoder/decoder length
self.batch_size = min(np.min(self.buckets_count), batch_size)
self.init_states_function = init_states_function
self.pad_label = word2idx['<PAD>']
self.shuffle = shuffle
self.rev = rev # reverse the encoder input
self.reset()
self.generate_init_states()
def generate_init_states(self):
self.init_states = self.init_states_function(self.num_layers, self.num_hidden, self.batch_size)
def generate_buckets(self):
enc_dec_data = []
for data, label in self.data_label:
enc_len = len(data) - 1 # minue one because of the <EOS>
dec_len = len(label)
enc_dec_data.append((enc_len, dec_len))
enc_dec_data = np.array(enc_dec_data)
kmeans = KMeans(n_clusters = self.num_buckets, random_state = 1) # use clustering to decide the buckets
assignments = kmeans.fit_predict(enc_dec_data) # get the assignments
# get the max of every cluster
buckets = np.array([np.max( enc_dec_data[assignments==i], axis=0 ) for i in range(self.num_buckets)])
# get # of sequences in each bucket... then assign the batch size as the minimum(minimum(bucketsize), batchsize)
buckets_count = np.array( [ enc_dec_data[assignments==i].shape[0] for i in range(self.num_buckets) ] )
return buckets, buckets_count, assignments
@property
def default_bucket_key(self):
enc_len, dec_len = np.max(self.buckets, axis=0)
return EncDecBucketKey(enc_len = enc_len, dec_len = dec_len)
@property
def provide_data(self): # this is necessary when specifying custom DataIter
# length of data variable is length of encoder + length of decoder
enc_dec_bucket_key = self.default_bucket_key
return [
('data', (self.batch_size, enc_dec_bucket_key.enc_len + enc_dec_bucket_key.dec_len)),
('mask', (self.batch_size, enc_dec_bucket_key.enc_len + enc_dec_bucket_key.dec_len))
] + self.init_states
#
@property
def provide_label(self): # this is necessary when specifying custom DataIter
# length of label variable is only the length of decoder
enc_dec_bucket_key = self.default_bucket_key
return [('label', (self.batch_size, enc_dec_bucket_key.dec_len))]
# for custom DataIter, we must implement this class as an iterable and return a DataBatch
def __iter__(self): # this is necessary to convert this class into an iterable
return self
def __next__(self):
if self.iter_next():
# suppose to get self.cursor:self.cursor + self.batch_size
batch = self.data_label[self.assignments == self.cur_permute_bucket]\
[ self.in_bucket_permutation[self.cursor:self.cursor+self.batch_size] ]
# get size of this bucket
enc_len, dec_len = self.buckets[self.cur_permute_bucket] # this enc_len already deducted the <EOS>
# total length of rnn sequence is enc_len+dec_len
all_data = np.full((self.batch_size, enc_len+dec_len), self.pad_label, dtype=float)
all_label = np.full((self.batch_size, dec_len), self.pad_label, dtype=float)
all_mask = np.zeros((self.batch_size, enc_len+dec_len), dtype=float)
for i, (data, label) in enumerate(batch):
if self.rev:
# reverse the input except for the <EOS> at end of input
# according to Ilya Sutskever et al. Sequence to Sequence Learning with Neural Networks
data[:-1] = np.flipud(data[:-1])
enc_input = np.concatenate((data, label[:-1])) # data <EOS> label
z = enc_len - data.shape[0] + 1
all_data[i, z:enc_len + label.shape[0]] = enc_input
all_mask[i, z:enc_len + label.shape[0]] = 1.0
all_label[i, :label.shape[0]] = label
return EncoderDecoderBatch(all_data, all_mask, all_label, self.init_states, EncDecBucketKey(enc_len=enc_len, dec_len=dec_len))
else:
raise StopIteration
def iter_next(self):
self.cursor += self.batch_size
if self.cursor < self.buckets_count[self.cur_permute_bucket]:
if self.cursor + self.batch_size > self.buckets_count[self.cur_permute_bucket]:
# it is going to overflow the bucket
self.cursor -= self.cursor + self.batch_size - self.buckets_count[self.cur_permute_bucket]
return True
else:
self.cur_bucket += 1
if self.cur_bucket < self.num_buckets:
self.cursor = 0
self.cur_permute_bucket = self.bucket_permutation[self.cur_bucket]
if self.shuffle:
self.in_bucket_permutation = np.random.permutation(self.buckets_count[self.cur_permute_bucket])
else:
self.in_bucket_permutation = np.array(range(self.buckets_count[self.cur_permute_bucket]))
return True
else:
return False
def reset(self): # for iterable
self.cursor = -self.batch_size
self.cur_bucket = 0
if self.shuffle:
self.bucket_permutation = np.random.permutation(self.num_buckets)
else:
self.bucket_permutation = np.array(range(self.num_buckets))
self.cur_permute_bucket = self.bucket_permutation[self.cur_bucket]
if self.shuffle:
self.in_bucket_permutation = np.random.permutation(self.buckets_count[self.cur_permute_bucket])
else:
self.in_bucket_permutation = np.array(range(self.buckets_count[self.cur_permute_bucket]))