本文首发于微信公众号：NewBeeNLP
写在前面
这是NLP保姆级教程的第二篇----基于RNN的文本分类实现（Text RNN）
参考的的论文是来自2016年复旦大学IJCAI上的发表的关于循环神经网络在多任务文本分类上的应用：Recurrent Neural Network for Text Classification with Multi-Task Learning[1]
论文概览
在先前的许多工作中，模型的学习都是基于单任务，对于复杂的问题，也可以分解为简单且相互独立的子问题来单独解决，然后再合并结果，得到最初复杂问题的结果。这样做看似合理，其实是不正确的，因为现实世界中很多问题不能分解为一个一个独立的子问题，即使可以分解，各个子问题之间也是相互关联的，通过一些共享因素或「共享表示（share representation）」 联系在一起。把现实问题当做一个个独立的单任务处理，往往会忽略了问题之间所富含的丰富的关联信息。
上面的问题引出了本文的重点——「多任务学习（Multi-task learning）」，把多个相关（related）的任务（task）放在一起学习。多个任务之间共享一些因素，它们可以在学习过程中，共享它们所学到的信息，这是单任务学习没有具备的。相关联的多任务学习比单任务学习能去的更好的泛化（generalization）效果。本文基于 RNN 循环神经网络，提出三种不同的信息共享机制，整体网络是基于所有的任务共同学习得到。
下面具体介绍一下文章中的三个模型。
Model I: Uniform-Layer Architecture
其中等号右侧第一项和第二项分别表示该任务「特有」的word embedding和该模型中「共享」的word embedding，两者做一个concatenation。
LSTM网络层是所有任务所共享的，对于任务m的最后sequence representation为LSTM的输出：
Model II: Coupled-Layer Architecture
为了更好地控制在不同LSTM layer之间的信息流动，作者提出了一个global gating unit，使得模型具有决定信息流动程度的能力。
为此，他们改写了LSTM中的表达式：
其中，
Model III: Shared-Layer Architecture
模型表现
论文作者在4个数据集上对上述模型做了评价，并和其他state-of-the-art的网络模型进行了对比，均显示最好的效果。
代码实现
RNN的代码框架和上一篇介绍的CNN类似，首先定义一个RNN类来实现论文中的模型
class RNN(BaseModel):
"""
A RNN class for sentence classification
With an embedding layer + Bi-LSTM layer + FC layer + softmax
"""
def __init__(self, sequence_length, num_classes, vocab_size,
embed_size, learning_rate, decay_steps, decay_rate,
hidden_size, is_training, l2_lambda, grad_clip,
initializer=tf.random_normal_initializer(stddev=0.1)):
这里的模型包括了一层embedding，一层双向LSTM，一层全连接层最后接上一个softmax分类函数。
然后依次定义模型，训练，损失等函数在后续调用。
def inference(self):
"""
1. embedding layer
2. Bi-LSTM layer
3. concat Bi-LSTM output
4. FC(full connected) layer
5. softmax layer
"""
# embedding layer
with tf.name_scope('embedding'):
self.embedded_words = tf.nn.embedding_lookup(self.Embedding, self.input_x)
# Bi-LSTM layer
with tf.name_scope('Bi-LSTM'):
lstm_fw_cell = rnn.BasicLSTMCell(self.hidden_size)
lstm_bw_cell = rnn.BasicLSTMCell(self.hidden_size)
if self.dropout_keep_prob is not None:
lstm_fw_cell = rnn.DropoutWrapper(lstm_fw_cell, output_keep_prob=self.dropout_keep_prob)
lstm_bw_cell = rnn.DropoutWrapper(lstm_bw_cell, output_keep_prob=self.dropout_keep_prob)
outputs, output_states = tf.nn.bidirectional_dynamic_rnn(lstm_fw_cell, lstm_bw_cell,
self.embedded_words,
dtype=tf.float32)
output = tf.concat(outputs, axis=2)
output_last = tf.reduce_mean(output, axis=1)
# FC layer
with tf.name_scope('output'):
self.score = tf.matmul(output_last, self.W_projection) + self.b_projection
return self.score
def loss(self):
# loss
with tf.name_scope('loss'):
losses = tf.nn.softmax_cross_entropy_with_logits(labels=self.input_y, logits=self.score)
data_loss = tf.reduce_mean(losses)
l2_loss = tf.add_n([tf.nn.l2_loss(cand_v) for cand_v in tf.trainable_variables()
if 'bias' not in cand_v.name]) * self.l2_lambda
data_loss += l2_loss
return data_loss
def train(self):
learning_rate = tf.train.exponential_decay(self.learning_rate, self.global_step,
self.decay_steps, self.decay_rate, staircase=True)
optimizer = tf.train.AdamOptimizer(learning_rate)
grads_and_vars = optimizer.compute_gradients(self.loss_val)
grads_and_vars = [(tf.clip_by_norm(grad, self.grad_clip), val) for grad, val in grads_and_vars]
train_op = optimizer.apply_gradients(grads_and_vars, global_step=self.global_step)
return train_op
训练部分的数据集这里就直接采用CNN那篇文章相同的数据集（懒...），预处理的方式与函数等都是一样的，，，
def train(x_train, y_train, vocab_processor, x_dev, y_dev):
with tf.Graph().as_default():
session_conf = tf.ConfigProto(
# allows TensorFlow to fall back on a device with a certain operation implemented
allow_soft_placement= FLAGS.allow_soft_placement,
# allows TensorFlow log on which devices (CPU or GPU) it places operations
log_device_placement=FLAGS.log_device_placement
)
sess = tf.Session(config=session_conf)
with sess.as_default():
# initialize cnn
rnn = RNN(sequence_length=x_train.shape[1],
num_classes=y_train.shape[1],
vocab_size=len(vocab_processor.vocabulary_),
embed_size=FLAGS.embed_size,
l2_lambda=FLAGS.l2_reg_lambda,
is_training=True,
grad_clip=FLAGS.grad_clip,
learning_rate=FLAGS.learning_rate,
decay_steps=FLAGS.decay_steps,
decay_rate=FLAGS.decay_rate,
hidden_size=FLAGS.hidden_size
)
# output dir for models and summaries
timestamp = str(time.time())
out_dir = os.path.abspath(os.path.join(os.path.curdir, 'run', timestamp))
if not os.path.exists(out_dir):
os.makedirs(out_dir)
print('Writing to {} \n'.format(out_dir))
# checkpoint dir. checkpointing – saving the parameters of your model to restore them later on.
checkpoint_dir = os.path.abspath(os.path.join(out_dir, FLAGS.ckpt_dir))
checkpoint_prefix = os.path.join(checkpoint_dir, 'model')
if not os.path.exists(checkpoint_dir):
os.makedirs(checkpoint_dir)
saver = tf.train.Saver(tf.global_variables(), max_to_keep=FLAGS.num_checkpoints)
# Write vocabulary
vocab_processor.save(os.path.join(out_dir, 'vocab'))
# Initialize all
sess.run(tf.global_variables_initializer())
def train_step(x_batch, y_batch):
"""
A single training step
:param x_batch:
:param y_batch:
:return:
"""
feed_dict = {
rnn.input_x: x_batch,
rnn.input_y: y_batch,
rnn.dropout_keep_prob: FLAGS.dropout_keep_prob
}
_, step, loss, accuracy = sess.run(
[rnn.train_op, rnn.global_step, rnn.loss_val, rnn.accuracy],
feed_dict=feed_dict
)
time_str = datetime.datetime.now().isoformat()
print("{}: step {}, loss {:g}, acc {:g}".format(time_str, step, loss, accuracy))
def dev_step(x_batch, y_batch):
"""
Evaluate model on a dev set
Disable dropout
:param x_batch:
:param y_batch:
:param writer:
:return:
"""
feed_dict = {
rnn.input_x: x_batch,
rnn.input_y: y_batch,
rnn.dropout_keep_prob: 1.0
}
step, loss, accuracy = sess.run(
[rnn.global_step, rnn.loss_val, rnn.accuracy],
feed_dict=feed_dict
)
time_str = datetime.datetime.now().isoformat()
print("dev results:{}: step {}, loss {:g}, acc {:g}".format(time_str, step, loss, accuracy))
# generate batches
batches = data_process.batch_iter(list(zip(x_train, y_train)), FLAGS.batch_size, FLAGS.num_epochs)
# training loop
for batch in batches:
x_batch, y_batch = zip(*batch)
train_step(x_batch, y_batch)
current_step = tf.train.global_step(sess, rnn.global_step)
if current_step % FLAGS.validate_every == 0:
print('\n Evaluation:')
dev_step(x_dev, y_dev)
print('')
path = saver.save(sess, checkpoint_prefix, global_step=current_step)
print('Save model checkpoint to {} \n'.format(path))
def main(argv=None):
x_train, y_train, vocab_processor, x_dev, y_dev = prepocess()
train(x_train, y_train, vocab_processor, x_dev, y_dev)
if __name__ == '__main__':
tf.app.run()「完整代码可以在公众号后台回复"RNN2016"获取。」
本文参考资料
[1]Recurrent Neural Network for Text Classification with Multi-Task Learning: https://arxiv.org/abs/1605.05101
- END -
Transformers Assemble（PART III）
本文首发于微信公众号：NewBeeNLP
