forked from marekrei/sequence-labeler
-
Notifications
You must be signed in to change notification settings - Fork 0
/
sequence_labeler.py
171 lines (144 loc) · 10.5 KB
/
sequence_labeler.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
import sys
import theano
import numpy
import collections
import lasagne
import crf
import recurrence
try:
import cPickle as pickle
except:
import pickle
sys.setrecursionlimit(50000)
floatX=theano.config.floatX
class SequenceLabeler(object):
def __init__(self, config):
self.config = config
self.params = collections.OrderedDict()
self.rng = numpy.random.RandomState(config["random_seed"])
word_ids = theano.tensor.imatrix('word_ids')
char_ids = theano.tensor.itensor3('char_ids')
char_mask = theano.tensor.ftensor3('char_mask')
label_ids = theano.tensor.imatrix('label_ids')
learningrate = theano.tensor.fscalar('learningrate')
is_training = theano.tensor.iscalar('is_training')
cost = 0.0
input_tensor = None
input_vector_size = 0
self.word_embeddings = self.create_parameter_matrix('word_embeddings', (config["n_words"], config["word_embedding_size"]))
input_tensor = self.word_embeddings[word_ids]
input_vector_size = config["word_embedding_size"]
char_embeddings = self.create_parameter_matrix('char_embeddings', (config["n_chars"], config["char_embedding_size"]))
char_input_tensor = char_embeddings[char_ids].reshape((char_ids.shape[0]*char_ids.shape[1],char_ids.shape[2],config["char_embedding_size"]))
char_mask_reshaped = char_mask.reshape((char_ids.shape[0]*char_ids.shape[1],char_ids.shape[2]))
char_output_tensor = recurrence.create_birnn(char_input_tensor, config["char_embedding_size"], char_mask_reshaped, config["char_recurrent_size"], return_combined=True, fn_create_parameter_matrix=self.create_parameter_matrix, name="char_birnn")
char_output_tensor = recurrence.create_feedforward(char_output_tensor, config["char_recurrent_size"]*2, config["word_embedding_size"], "tanh", fn_create_parameter_matrix=self.create_parameter_matrix, name="char_ff")
char_output_tensor = char_output_tensor.reshape((char_ids.shape[0],char_ids.shape[1],config["word_embedding_size"]))
if config["char_integration_method"] == "input":
input_tensor = theano.tensor.concatenate([input_tensor, char_output_tensor], axis=2)
input_vector_size += config["word_embedding_size"]
elif config["char_integration_method"] == "attention":
static_input_tensor = theano.gradient.disconnected_grad(input_tensor)
is_unk = theano.tensor.eq(word_ids, config["unk_token_id"])
is_unk_tensor = is_unk.dimshuffle(0,1,'x')
char_output_tensor_normalised = char_output_tensor / char_output_tensor.norm(2, axis=2)[:, :, numpy.newaxis]
static_input_tensor_normalised = static_input_tensor / static_input_tensor.norm(2, axis=2)[:, :, numpy.newaxis]
cosine_cost = 1.0 - (char_output_tensor_normalised * static_input_tensor_normalised).sum(axis=2)
cost += theano.tensor.switch(is_unk, 0.0, cosine_cost).sum()
attention_evidence_tensor = theano.tensor.concatenate([input_tensor, char_output_tensor], axis=2)
attention_output = recurrence.create_feedforward(attention_evidence_tensor, config["word_embedding_size"]*2, config["word_embedding_size"], "tanh", self.create_parameter_matrix, "attention_tanh")
attention_output = recurrence.create_feedforward(attention_output, config["word_embedding_size"], config["word_embedding_size"], "sigmoid", self.create_parameter_matrix, "attention_sigmoid")
input_tensor = input_tensor * attention_output + char_output_tensor * (1.0 - attention_output)
if config["dropout_input"] > 0.0:
p = config["dropout_input"]
trng = theano.tensor.shared_randomstreams.RandomStreams(seed=1)
dropout_mask = trng.binomial(n=1, p=1-p, size=input_tensor.shape, dtype=floatX)
input_train = dropout_mask * input_tensor
input_test = (1 - p) * input_tensor
input_tensor = theano.tensor.switch(theano.tensor.neq(is_training, 0), input_train, input_test)
recurrent_forward = recurrence.create_lstm(input_tensor.dimshuffle(1,0,2), input_vector_size, None,
config["word_recurrent_size"], False, go_backwards=False,
fn_create_parameter_matrix=self.create_parameter_matrix, name="word_birnn_forward").dimshuffle(1,0,2)
recurrent_backward = recurrence.create_lstm(input_tensor.dimshuffle(1,0,2), input_vector_size, None,
config["word_recurrent_size"], False, go_backwards=True,
fn_create_parameter_matrix=self.create_parameter_matrix, name="word_birnn_backward").dimshuffle(1,0,2)
if config["lmcost_gamma"] > 0.0:
lmcost_max_vocab_size = min(self.config["n_words"], self.config["lmcost_max_vocab_size"])
cost += config["lmcost_gamma"] * self.construct_lmcost(recurrent_forward[:,:-1,:], config["word_recurrent_size"], word_ids[:,1:], self.config["lmcost_layer_size"], lmcost_max_vocab_size, "lmcost_forward")
cost += config["lmcost_gamma"] * self.construct_lmcost(recurrent_backward[:,1:,:], config["word_recurrent_size"], word_ids[:,:-1], self.config["lmcost_layer_size"], lmcost_max_vocab_size, "lmcost_backward")
processed_tensor = theano.tensor.concatenate([recurrent_forward, recurrent_backward], axis=2)
processed_tensor_size = config["word_recurrent_size"] * 2
if config["narrow_layer_size"] > 0:
processed_tensor = recurrence.create_feedforward(processed_tensor, processed_tensor_size, config["narrow_layer_size"], "tanh", fn_create_parameter_matrix=self.create_parameter_matrix, name="narrow_ff")
processed_tensor_size = config["narrow_layer_size"]
W_output = self.create_parameter_matrix('W_output', (processed_tensor_size, config["n_labels"]))
bias_output = self.create_parameter_matrix('bias_output', (config["n_labels"],))
output = theano.tensor.dot(processed_tensor, W_output) + bias_output
output = output[:,1:-1,:] # removing <s> and </s>
if config["crf_on_top"] == True:
all_paths_scores, real_paths_scores, best_sequence, scores = crf.construct("crf", output, config["n_labels"], label_ids, self.create_parameter_matrix)
predicted_labels = best_sequence
output_probs = scores
cost += - (real_paths_scores - all_paths_scores).sum()
else:
output_probs_ = theano.tensor.nnet.softmax(output.reshape((word_ids.shape[0]*(word_ids.shape[1]-2), config["n_labels"])))
output_probs = output_probs_.reshape((word_ids.shape[0], (word_ids.shape[1]-2), config["n_labels"]))
predicted_labels = theano.tensor.argmax(output_probs, axis=2)
cost += theano.tensor.nnet.categorical_crossentropy(output_probs_, label_ids.reshape((-1,))).sum()
gradients = theano.tensor.grad(cost, list(self.params.values()), disconnected_inputs='ignore')
if config["opt_strategy"] == "adadelta":
updates = lasagne.updates.adadelta(gradients, self.params.values(), learningrate)
elif config["opt_strategy"] == "adam":
updates = lasagne.updates.adam(gradients, self.params.values(), learningrate)
elif config["opt_strategy"] == "sgd":
updates = lasagne.updates.sgd(gradients, self.params.values(), learningrate)
else:
raise ValueError("Unknown optimisation strategy: " + str(config["opt_strategy"]))
input_vars_train = [word_ids, char_ids, char_mask, label_ids, learningrate]
input_vars_test = [word_ids, char_ids, char_mask, label_ids]
output_vars = [cost, predicted_labels]
self.train = theano.function(input_vars_train, output_vars, updates=updates, on_unused_input='ignore', allow_input_downcast = True, givens=({is_training: numpy.cast['int32'](1)}))
self.test = theano.function(input_vars_test, output_vars, on_unused_input='ignore', allow_input_downcast = True, givens=({is_training: numpy.cast['int32'](0)}))
self.test_return_probs = theano.function(input_vars_test, output_vars + [output_probs,], on_unused_input='ignore', allow_input_downcast = True, givens=({is_training: numpy.cast['int32'](0)}))
def construct_lmcost(self, hidden_layer, hidden_layer_size, word_id_targets, lmcost_layer_size, max_vocab_size, name):
lmcost_layer = recurrence.create_feedforward(hidden_layer, hidden_layer_size, lmcost_layer_size, "tanh", fn_create_parameter_matrix=self.create_parameter_matrix, name=name+"_tanh")
lmcost_output = recurrence.create_feedforward(lmcost_layer, lmcost_layer_size, max_vocab_size+1, "linear", fn_create_parameter_matrix=self.create_parameter_matrix, name=name+"_output")
lmcost_output = theano.tensor.nnet.softmax(lmcost_output.reshape((lmcost_layer.shape[0]*lmcost_layer.shape[1], max_vocab_size+1)))
word_id_targets = theano.tensor.switch(theano.tensor.ge(word_id_targets, max_vocab_size), max_vocab_size, word_id_targets)
lmcost = theano.tensor.nnet.categorical_crossentropy(lmcost_output, word_id_targets.reshape((-1,))).sum()
return lmcost
def create_parameter_matrix(self, name, size):
param_vals = numpy.asarray(self.rng.normal(loc=0.0, scale=0.1, size=size), dtype=floatX)
param_shared = theano.shared(param_vals, name)
self.params[name] = param_shared
return param_shared
def get_parameter_count(self):
total = 0
for key, val in self.params.items():
total += val.get_value().size
return total
def get_parameter_count_without_word_embeddings(self):
total = 0
for key, val in self.params.items():
if val == self.word_embeddings:
continue
total += val.get_value().size
return total
def save(self, filename):
dump = {}
dump["config"] = self.config
dump["params"] = {}
for param_name in self.params:
dump["params"][param_name] = self.params[param_name].get_value()
with open(filename, 'wb') as f:
pickle.dump(dump, f, protocol=pickle.HIGHEST_PROTOCOL)
@staticmethod
def load(filename):
sequencelabeler = None
with open(filename, 'rb') as f:
dump = pickle.load(f)
sequencelabeler = SequenceLabeler(dump["config"])
for param_name in sequencelabeler.params:
assert(param_name in dump["params"])
sequencelabeler.params[param_name].set_value(dump["params"][param_name])
return sequencelabeler