from __future__ import print_function
import tensorflow as tf
import tensorflow.contrib.layers as layers
import tensorflow.contrib.crf as crf
import time
import codecs
import os
import cPickle as pickle
import numpy as np
from itertools import izip
INT_TYPE = np.int32
FLOAT_TYPE = np.float32
################################################################################
# Model #
################################################################################
class Model(object):
def __init__(self, scope, sess):
self.scope = scope
self.sess = sess
def build_input_graph(self, vocab_size, emb_size, word_vocab_size, word_emb_size, word_window_size):
"""
Gather embeddings from lookup tables.
"""
seq_ids = tf.placeholder(dtype=INT_TYPE, shape=[None, None], name='seq_ids')
seq_word_ids = [tf.placeholder(dtype=INT_TYPE, shape=[None, None], name='seq_feature_%d_ids' % i)
for i in range(word_window_size)]
embeddings = tf.get_variable('embeddings', [vocab_size, emb_size])
embedding_output = tf.nn.embedding_lookup([embeddings], seq_ids)
word_outputs = []
word_embeddings = tf.get_variable('word_embeddings', [word_vocab_size, word_emb_size])
for i in range(word_window_size):
word_outputs.append(tf.nn.embedding_lookup([word_embeddings], seq_word_ids[i]))
return seq_ids, seq_word_ids, tf.concat([embedding_output] + word_outputs, 2, 'inputs')
def build_tagging_graph(self, inputs, hidden_layers, channels, num_tags, use_crf, lamd, dropout_emb,
dropout_hidden, kernel_size, use_bn, use_wn, active_type):
"""
Build a deep neural model for sequence tagging.
"""
stag_ids = tf.placeholder(dtype=INT_TYPE, shape=[None, None], name='stag_ids')
seq_lengths = tf.placeholder(dtype=INT_TYPE, shape=[None], name='seq_lengths')
# Default is not train.
is_train = tf.placeholder(dtype=tf.bool, shape=[], name='is_train')
masks = tf.cast(tf.sequence_mask(seq_lengths), FLOAT_TYPE)
# Dropout on embedding output.
if dropout_emb:
inputs = tf.cond(is_train,
lambda: tf.nn.dropout(inputs, 1 - dropout_emb),
lambda: inputs)
hidden_output = inputs
pre_channels = inputs.get_shape()[-1].value
for i in xrange(hidden_layers):
k = kernel_size
cur_channels = channels[i]
filter_w = tf.get_variable('filter_w_%d' % i, shape=[k, pre_channels, cur_channels], dtype=FLOAT_TYPE)
filter_v = tf.get_variable('filter_v_%d' % i, shape=[k, pre_channels, cur_channels], dtype=FLOAT_TYPE)
bias_b = tf.get_variable('bias_b_%d' % i, shape=[cur_channels],
initializer=tf.zeros_initializer(dtype=FLOAT_TYPE))
bias_c = tf.get_variable('bias_c_%d' % i, shape=[cur_channels],
initializer=tf.zeros_initializer(dtype=FLOAT_TYPE))
# Weight normalization.
if use_wn:
epsilon = 1e-12
g_w = tf.get_variable('g_w_%d' % i, shape=[k, 1, cur_channels], dtype=FLOAT_TYPE)
g_v = tf.get_variable('g_v_%d' % i, shape=[k, 1, cur_channels], dtype=FLOAT_TYPE)
# Perform wn
filter_w = g_w * filter_w / (tf.sqrt(tf.reduce_sum(filter_w ** 2, 1, keep_dims=True)) + epsilon)
filter_v = g_v * filter_v / (tf.sqrt(tf.reduce_sum(filter_v ** 2, 1, keep_dims=True)) + epsilon)
w = tf.nn.conv1d(hidden_output, filter_w, 1, 'SAME') + bias_b
v = tf.nn.conv1d(hidden_output, filter_v, 1, 'SAME') + bias_c
if use_bn:
w = layers.batch_norm(inputs=v, decay=0.9, is_training=is_train, center=True, scale=True,
scope='BatchNorm_w_%d' % i)
v = layers.batch_norm(inputs=w, decay=0.9, is_training=is_train, center=True, scale=True,
scope='BatchNorm_v_%d' % i)
if active_type == 'glu':
hidden_output = w * tf.nn.sigmoid(v)
elif active_type == 'relu':
hidden_output = tf.nn.relu(w)
elif active_type == 'gtu':
hidden_output = tf.tanh(w) * tf.nn.sigmoid(v)
elif active_type == 'tanh':
hidden_output = tf.tanh(w)
elif active_type == 'linear':
hidden_output = w
elif active_type == 'bilinear':
hidden_output = w * v
# Mask paddings.
hidden_output = hidden_output * tf.expand_dims(masks, -1)
# Dropout on hidden output.
if dropout_hidden:
hidden_output = tf.cond(is_train,
lambda: tf.nn.dropout(hidden_output, 1 - dropout_hidden),
lambda: hidden_output
)
pre_channels = cur_channels
# Un-scaled log probabilities.
scores = layers.fully_connected(hidden_output, num_tags, tf.identity)
if use_crf:
cost, transitions = crf.crf_log_likelihood(inputs=scores, tag_indices=stag_ids,
sequence_lengths=seq_lengths)
cost = - tf.reduce_mean(cost)
else:
reshaped_scores = tf.reshape(scores, [-1, num_tags])
reshaped_stag_ids = tf.reshape(stag_ids, [-1])
real_distribution = layers.one_hot_encoding(reshaped_stag_ids, num_tags)
cost = tf.nn.softmax_cross_entropy_with_logits(reshaped_scores, real_distribution)
cost = tf.reduce_sum(tf.reshape(cost, tf.shape(stag_ids)) * masks) / tf.cast(tf.shape(inputs)[0],
FLOAT_TYPE)
# Calculate L2 penalty.
l2_penalty = 0
if lamd > 0:
for v in tf.trainable_variables():
if '/B:' not in v.name and '/biases:' not in v.name:
l2_penalty += lamd * tf.nn.l2_loss(v)
train_cost = cost + l2_penalty
# Summary cost.
tf.summary.scalar('cost', cost)
summaries = tf.summary.merge_all()
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
if update_ops:
updates = tf.group(*update_ops)
with tf.control_dependencies([updates]):
cost = tf.identity(cost)
return stag_ids, seq_lengths, is_train, cost, train_cost, scores, summaries
def build_graph(self):
parameters = self.parameters
with tf.variable_scope(name_or_scope=self.scope, initializer=tf.uniform_unit_scaling_initializer()):
seq_ids_pl, seq_other_ids_pls, inputs = self.build_input_graph(vocab_size=parameters['vocab_size'],
emb_size=parameters['emb_size'],
word_window_size=parameters['word_window_size'],
word_vocab_size=parameters['word_vocab_size'],
word_emb_size=parameters['word_emb_size'])
stag_ids_pl, seq_lengths_pl, is_train_pl, cost_op, train_cost_op, scores_op, summary_op = \
self.build_tagging_graph(inputs=inputs,
num_tags=parameters['num_tags'],
use_crf=parameters['use_crf'],
lamd=parameters['lamd'],
dropout_emb=parameters['dropout_emb'],
dropout_hidden=parameters['dropout_hidden'],
hidden_layers=parameters['hidden_layers'],
channels=parameters['channels'],
kernel_size=parameters['kernel_size'],
use_bn=parameters['use_bn'],
use_wn=parameters['use_wn'],
active_type=parameters['active_type'])
self.seq_ids_pl = seq_ids_pl
self.seq_other_ids_pls = seq_other_ids_pls
self.stag_ids_pl = stag_ids_pl
self.seq_lengths_pl = seq_lengths_pl
self.is_train_pl = is_train_pl
self.cost_op = cost_op
self.train_cost_op = train_cost_op
self.scores_op = scores_op
self.summary_op = summary_op
def inference(self, scores, sequence_lengths=None):
"""
Inference label sequence given scores.
If transitions is given, then perform veterbi search, else perform greedy search.
Args:
scores: A numpy array with shape (batch, max_length, num_tags).
sequence_lengths: A numpy array with shape (batch,).
Returns:
A numpy array with shape (batch, max_length).
"""
if not self.parameters['use_crf']:
return np.argmax(scores, 2)
else:
with tf.variable_scope(self.scope, reuse=True):
transitions = tf.get_variable('transitions').eval(session=self.sess)
paths = np.zeros(scores.shape[:2], dtype=INT_TYPE)
for i in xrange(scores.shape[0]):
tag_score, length = scores[i], sequence_lengths[i]
if length == 0:
continue
path, _ = crf.viterbi_decode(tag_score[:length], transitions)
paths[i, :length] = path
return paths
def train(self, train_data, dev_data, test_data, model_dir, log_dir, emb_size, word_emb_size, optimizer,
hidden_layers, channels, kernel_size, active_type, use_bn, use_wn, use_crf, lamd, dropout_emb,
dropout_hidden, evaluator, batch_size, eval_batch_size, pre_trained_emb_path, fix_word_emb,
reserve_all_word_emb, pre_trained_word_emb_path, max_epoches, print_freq):
"""
This function is the main function for preparing data and training the model.
"""
assert len(channels) == hidden_layers
# Parse optimization method and parameters.
optimizer = optimizer.split('_')
optimizer_name = optimizer[0]
optimizer_options = [eval(i) for i in optimizer[1:]]
optimizer = {
'sgd': tf.train.GradientDescentOptimizer,
'adadelta': tf.train.AdadeltaOptimizer,
'adam': tf.train.AdamOptimizer,
'mom': tf.train.MomentumOptimizer
}[optimizer_name](*optimizer_options)
print('Preparing data...', end='')
if not os.path.exists(model_dir):
os.makedirs(model_dir)
mappings_path = os.path.join(model_dir, 'mappings.pkl')
parameters_path = os.path.join(model_dir, 'parameters.pkl')
# Load character embeddings.
pre_trained = {}
if pre_trained_emb_path and os.path.isfile(pre_trained_emb_path):
for l in codecs.open(pre_trained_emb_path, 'r', 'utf8'):
we = l.split()
if len(we) == emb_size + 1:
w, e = we[0], np.array(map(float, we[1:]))
pre_trained[w] = e
# Load word embeddings.
pre_trained_word = {}
if pre_trained_word_emb_path and os.path.isfile(pre_trained_word_emb_path):
for l in codecs.open(pre_trained_word_emb_path, 'r', 'utf8', 'ignore'):
we = l.split()
if len(we) == word_emb_size + 1:
w, e = we[0], np.array(map(float, we[1:]))
pre_trained_word[w] = e
# Load or create mappings.
if os.path.isfile(mappings_path):
item2id, id2item, tag2id, id2tag, word2id, id2word = pickle.load(open(mappings_path, 'r'))
else:
item2id, id2item = create_mapping(create_dic(train_data[0], add_unk=True, add_pad=True))
tag2id, id2tag = create_mapping(create_dic(train_data[-1]))
words = []
for t in train_data[1:-1]:
words.extend(t)
for t in dev_data[1:-1]:
words.extend(t)
for t in test_data[1:-1]:
words.extend(t)
word_dic = create_dic(words, add_unk=True, add_pad=True)
for k in word_dic.keys():
if k not in pre_trained_word and k != '<UNK>' and k != '<PAD>':
word_dic.pop(k)
if reserve_all_word_emb:
for w in pre_trained_word:
if w not in word_dic:
word_dic[w] = 0
word2id, id2word = create_mapping(word_dic)
# Save the mappings to disk.
pickle.dump((item2id, id2item, tag2id, id2tag, word2id, id2word), open(mappings_path, 'w'))
# Hyper parameters.
word_window_size = len(train_data) - 2
parameters = {
'vocab_size': len(item2id),
'emb_size': emb_size,
'word_window_size': word_window_size,
'word_vocab_size': len(word2id),
'word_emb_size': word_emb_size,
'hidden_layers': hidden_layers,
'channels': channels,
'kernel_size': kernel_size,
'use_bn': use_bn,
'use_wn': use_wn,
'num_tags': len(tag2id),
'use_crf': use_crf,
'lamd': lamd,
'dropout_emb': dropout_emb,
'dropout_hidden': dropout_hidden,
'active_type': active_type
}
if os.path.isfile(parameters_path):
parameters_old = pickle.load(open(parameters_path, 'r'))
if parameters != parameters_old:
raise Exception('Network parameters are not consistent!')
else:
pickle.dump(parameters, open(parameters_path, 'w'))
self.item2id = item2id
self.id2item = id2item
self.tag2id = tag2id
self.id2tag = id2tag
self.word2id = word2id
self.id2word = id2word
self.parameters = parameters
# Convert data to corresponding ids.
train_data_ids = data_to_ids(
train_data, [item2id] + [word2id] * word_window_size + [tag2id]
)
print('Finished.')
print("Start building the network...", end='')
self.build_graph()
print('Finished.')
def summary(name, dtype=FLOAT_TYPE):
value = tf.placeholder(dtype, shape=[])
return value, tf.summary.scalar(name, value)
dev_f1_pl, dev_summary_op = summary('dev f1')
test_f1_pl, test_summary_op = summary('test f1')
# Clip gradients and apply.
grads_and_vars = optimizer.compute_gradients(loss=self.train_cost_op, var_list=tf.trainable_variables())
grads_and_vars = [(g, v) for g, v in grads_and_vars if g is not None]
# If use fixed word embeddings, remove the grad
if fix_word_emb:
grads_and_vars = [(g, v) for g, v in grads_and_vars if '/word_embeddings' not in v.name]
grads_summary_op = tf.summary.histogram('grads', tf.concat([tf.reshape(g, [-1]) for g, _ in grads_and_vars], 0))
grads_norm = tf.sqrt(sum([tf.reduce_sum(tf.pow(g, 2)) for g, _ in grads_and_vars]))
grads_and_vars = [(g / (tf.reduce_max([grads_norm, 5]) / 5), v) for g, v in grads_and_vars]
train_op = optimizer.apply_gradients(grads_and_vars)
# Variables for recording training procedure.
best_epoch = tf.get_variable('best_epoch', shape=[], initializer=tf.zeros_initializer(), trainable=False,
dtype=INT_TYPE)
best_step = tf.get_variable('best_step', shape=[], initializer=tf.zeros_initializer(), trainable=False,
dtype=INT_TYPE)
best_dev_score = tf.get_variable('best_dev_score', shape=[], initializer=tf.zeros_initializer(),
trainable=False, dtype=FLOAT_TYPE)
best_test_score = tf.get_variable('best_test_score', shape=[], initializer=tf.zeros_initializer(),
trainable=False, dtype=FLOAT_TYPE)
init_op = tf.global_variables_initializer()
saver = tf.train.Saver(tf.global_variables())
summary_writer = tf.summary.FileWriter(log_dir + '/summaries')
print('Finished.')
print('Start training the network...')
self.sess.run(init_op)
start_time_begin = time.time()
try:
checkpoint = tf.train.latest_checkpoint(model_dir)
saver.restore(self.sess, checkpoint)
print('Restore model from %s.' % checkpoint)
except (tf.errors.DataLossError, TypeError, Exception):
# Failed to restore model from disk. Load pre-trained embeddings.
# Load character embeddings.
with tf.variable_scope(self.scope, reuse=True):
embeddings = tf.get_variable('embeddings')
value = self.sess.run(embeddings)
count = 0
for item in item2id:
item_id = item2id[item]
if item in pre_trained:
value[item_id] = pre_trained[item]
count += 1
# Run assign op.
self.sess.run(embeddings.assign(value))
del (pre_trained)
print('%d of %d character embeddings were loaded from pre-trained.' % (count, len(item2id)))
# Load word embeddings.
with tf.variable_scope(self.scope, reuse=True):
word_embeddings = tf.get_variable('word_embeddings')
value = self.sess.run(word_embeddings)
count = 0
for item in word2id:
item_id = word2id[item]
if item in pre_trained_word:
value[item_id] = pre_trained_word[item]
count += 1
# Run assign op.
self.sess.run(word_embeddings.assign(value))
del (pre_trained_word)
print('%d of %d word embeddings were loaded from pre-trained.' % (count, len(word2id)))
start_epoch, global_step, best_dev_f1 = self.sess.run((best_epoch, best_step, best_dev_score))
for epoch in range(start_epoch + 1, max_epoches + 1):
print('Starting epoch %d...' % epoch)
start_time = time.time()
loss_ep = 0
n_step = 0
iterator = data_iterator(train_data_ids, batch_size, shuffle=True)
for batch in iterator:
batch = create_input(batch)
seq_ids, seq_other_ids_list, stag_ids, seq_lengths = batch[0], batch[1: -2], batch[-2], batch[-1]
feed_dict = {self.seq_ids_pl: seq_ids.astype(INT_TYPE),
self.stag_ids_pl: stag_ids.astype(INT_TYPE),
self.seq_lengths_pl: seq_lengths.astype(INT_TYPE),
self.is_train_pl: True}
assert len(self.seq_other_ids_pls) == len(seq_other_ids_list)
for pl, v in zip(self.seq_other_ids_pls, seq_other_ids_list):
feed_dict[pl] = v
# feed_dict.update(drop_feed_dict) # enable noise input
loss, summaries, grads_summaries, _ = self.sess.run(
[self.cost_op, self.summary_op, grads_summary_op, train_op],
feed_dict=feed_dict)
loss_ep += loss
n_step += 1
global_step += 1
summary_writer.add_summary(summaries, global_step)
summary_writer.add_summary(grads_summaries, global_step)
# Show training information.
if global_step % print_freq == 0:
print(' Step %d, current cost %.6f, average cost %.6f' % (global_step, loss, loss_ep / n_step))
loss_ep = loss_ep / n_step
print('Epoch %d finished. Time: %ds Cost: %.6f' % (epoch, time.time() - start_time, loss_ep))
# Evaluate precision, recall and f1 with an external script.
dev_pre, dev_rec, dev_f1 = \
evaluator((dev_data[0], dev_data[-1], self.tag_all(dev_data[:-1], eval_batch_size)[1]),
log_dir + '/dev', epoch)
test_pre, test_rec, test_f1 = \
evaluator((test_data[0], test_data[-1], self.tag_all(test_data[:-1], eval_batch_size)[1]),
log_dir + '/test', epoch)
# Summary dev and test F1 score.
summary_writer.add_summary(self.sess.run(dev_summary_op, {dev_f1_pl: dev_f1}), epoch)
summary_writer.add_summary(self.sess.run(test_summary_op, {test_f1_pl: test_f1}), epoch)
print("Dev precision / recall / f1 score: %.2f / %.2f / %.2f" %
(dev_pre * 100, dev_rec * 100, dev_f1 * 100))
print("Test precision / recall / f1 score: %.2f / %.2f / %.2f" %
(test_pre * 100, test_rec * 100, test_f1 * 100))
if dev_f1 > best_dev_f1:
best_dev_f1 = dev_f1
self.sess.run((tf.assign(best_epoch, epoch),
tf.assign(best_dev_score, dev_f1),
tf.assign(best_test_score, test_f1),
tf.assign(best_step, global_step)))
path = saver.save(self.sess, model_dir + '/model', epoch)
print('New best score on dev.')
print('Save model at %s.' % path)
print('Finished.')
print('Total training time: %fs.' % (time.time() - start_time_begin))
def load_model(self, model_dir):
mappings_path = os.path.join(model_dir, 'mappings.pkl')
parameters_path = os.path.join(model_dir, 'parameters.pkl')
item2id, id2item, tag2id, id2tag, word2id, id2word = \
pickle.load(open(mappings_path, 'r'))
parameters = pickle.load(open(parameters_path))
self.item2id = item2id
self.id2item = id2item
self.tag2id = tag2id
self.id2tag = id2tag
self.word2id = word2id
self.id2word = id2word
self.parameters = parameters
print(parameters)
print('Building input graph...', end='')
self.build_graph()
print('Finished.')
print('Initializing variables...', end='')
init_op = tf.initialize_all_variables()
self.sess.run(init_op)
print('Finished.')
print('Reloading parameters...', end='')
saver = tf.train.Saver(tf.global_variables())
checkpoint = tf.train.latest_checkpoint(model_dir)
saver.restore(self.sess, checkpoint)
print('Finished.')
def tag(self, data_iter):
"""A tagging function.
Args:
data_iter: A iterator for generate batches.
Returns:
A generator for tagging result.
"""
output = []
for data in data_iter:
batch = data_to_ids(data, [self.item2id] + [self.word2id] * self.parameters['word_window_size'])
batch = create_input(batch)
seq_ids, seq_other_ids_list, seq_lengths = batch[0], batch[1: -1], batch[-1]
feed_dict = {self.seq_ids_pl: seq_ids.astype(INT_TYPE),
self.seq_lengths_pl: seq_lengths.astype(INT_TYPE),
self.is_train_pl: False}
for pl, v in zip(self.seq_other_ids_pls, seq_other_ids_list):
feed_dict[pl] = v.astype(INT_TYPE)
scores = self.sess.run(self.scores_op, feed_dict)
stag_ids = self.inference(scores, seq_lengths)
for seq, stag_id, length in izip(data[0], stag_ids, seq_lengths):
output.append((seq, [self.id2tag[t] for t in stag_id[:length]]))
yield zip(*output)
output = []
def tag_all(self, data, batch_size):
data_iter = data_iterator(data, batch_size=batch_size, shuffle=False)
output = []
for b in self.tag(data_iter):
output.extend(zip(*b))
return zip(*output)
################################################################################
# DATA UTILS #
################################################################################
def create_dic(item_list, add_unk=False, add_pad=False):
"""
Create a dictionary of items from a list of list of items.
"""
assert type(item_list) in (list, tuple)
dic = {}
for items in item_list:
for item in items:
if item not in dic:
dic[item] = 1
else:
dic[item] += 1
# Make sure that <PAD> have a id 0.
if add_pad:
dic['<PAD>'] = 1e20
# If specified, add a special item <UNK>.
if add_unk:
dic['<UNK>'] = 1e10
return dic
def create_mapping(items):
"""
Create a mapping (item to ID / ID to item) from a dictionary.
Items are ordered by decreasing frequency.
"""
if type(items) is dict:
sorted_items = sorted(items.items(), key=lambda x: (-x[1], x[0]))
id2item = {i: v[0] for i, v in enumerate(sorted_items)}
item2id = {v: k for k, v in id2item.items()}
return item2id, id2item
elif type(items) is list:
id2item = {i: v for i, v in enumerate(items)}
item2id = {v: k for k, v in id2item.items()}
return item2id, id2item
def create_input(batch):
"""
Take each sentence data in batch and return an input for
the training or the evaluation function.
"""
assert len(batch) > 0
lengths = [len(seq) for seq in batch[0]]
max_len = max(2, max(lengths))
ret = []
for d in batch:
dd = []
for seq_id, pos in izip(d, lengths):
assert len(seq_id) == pos
pad = [0] * (max_len - pos)
dd.append(seq_id + pad)
ret.append(np.array(dd))
ret.append(np.array(lengths))
return ret
def data_to_ids(data, mappings):
"""
Map text data to ids.
"""
def strQ2B(ustring):
rstring = ""
for uchar in ustring:
inside_code = ord(uchar)
if inside_code == 12288:
inside_code = 32
elif 65281 <= inside_code <= 65374:
inside_code -= 65248
rstring += unichr(inside_code)
return rstring
def strB2Q(ustring):
rstring = ""
for uchar in ustring:
inside_code = ord(uchar)
if inside_code == 32:
inside_code = 12288
elif 32 <= inside_code <= 126:
inside_code += 65248
rstring += unichr(inside_code)
return rstring
def map(item, mapping):
if item in mapping:
return mapping[item]
item = strB2Q(item)
if item in mapping:
return mapping[item]
item = strQ2B(item)
if item in mapping:
return mapping[item]
return mapping['<UNK>']
def map_seq(seqs, mapping):
return [[map(item, mapping) for item in seq] for seq in seqs]
ret = []
for d, m in izip(data, mappings):
ret.append(map_seq(d, m))
return tuple(ret)
def data_iterator(inputs, batch_size, shuffle=True, max_length=200):
"""
A simple iterator for generating dynamic mini batches.
"""
assert len(inputs) > 0
assert all([len(item) == len(inputs[0]) for item in inputs])
inputs = zip(*inputs)
if shuffle:
np.random.shuffle(inputs)
batch = []
bs = batch_size
for d in inputs:
if len(d[0]) > max_length:
bs = max(1, min(batch_size * max_length / len(d[0]), bs))
if len(batch) < bs:
batch.append(d)
else:
yield zip(*batch)
batch = [d]
if len(d[0]) < max_length:
bs = batch_size
else:
bs = max(1, batch_size * max_length / len(d[0]))
if batch:
yield zip(*batch)
