TensorFlow RNN深度学习 BiLSTM+CRF 实现 sequence labeling 序列标注 源码

在TensorFlow RNN 深度学习下 BiLSTM+CRF 实现 sequence labeling 

双向LSTM+CRF 序列标注问题

源码

去年底样子一直在做NLP相关task,是个关于序列标注问题。这 sequence labeling属于NLP的经典问题了，开始尝试用HMM，哦不，用CRF做baseline，by the way, 用的CRF++。

关于CRF的理论就不再啰嗦了，街货。顺便提下，CRF比HMM在理论上以及实际效果上都要好不少。但我要说的是CRF跑我这task还是不太乐观。P值0.6样子，R低的离谱，所以F1很不乐观。mentor告诉我说是特征不足，师兄说是这个task本身就比较难做，F1低算是正常了。

CRF做完baseline后，一直在着手用BiLSTM+CRF跑 sequence labeling，奈何项目繁多，没有多余的精力去按照正常的计划做出来。后来还是一点一点的，按照大牛们的步骤以及参考现有的代码，把 BiLSTM+CRF的实现拿下了。后来发现，跑出来的效果也不太理想……可能是这个task确实变态……抑或模型还要加强吧~

这里对比下CRF与LSTM的cell，先说RNN吧，RNN其实是比CNN更适合做序列问题的模型，RNN隐层当前时刻的输入有一部分是前一时刻的隐层输出，这使得他能通过循环反馈连接看到前面的信息，将一段序列的前面的context capture 过来参与此刻的计算，并且还具备非线性的拟合能力，这都是CRF无法超越的地方。而LSTM的cell很好的将RNN的梯度弥散问题优化解决了，他对门卫gate说：老兄，有的不太重要的信息，你该忘掉就忘掉吧，免得占用现在的资源。而双向LSTM就更厉害了，不仅看得到过去，还能将未来的序列考虑进来，使得上下文信息充分被利用。而CRF，他不像LSTM能够考虑长远的上下文信息，它更多地考虑整个句子的局部特征的线性加权组合(通过特征模板扫描整个句子)，特别的一点，他计算的是联合概率，优化了整个序列，而不是拼接每个时刻的最优值。那么，将BILSTM与CRF一起就构成了还比较不错的组合，这目前也是学术界的流行做法~

另外针对目前的跑通结果提几个改进点：

1.+CNN，通过CNN的卷积操作去提取英文单词的字母细节。

2.+char representation,作用与上相似，提取更细粒度的细节。

3.more joint model to go.

fine，叨了不少。codes time:

完整代码以及相关预处理的数据请移步github: scofiled's github/bilstm+crf

requirements:

ubuntu14

python2.7

tensorflow 0.8

numpy

pandas0.15

BILSTM_CRF.py

import math
import helper
import numpy as np
import tensorflow as tf
from tensorflow.models.rnn import rnn, rnn_cell

class BILSTM_CRF(object):
    
    def __init__(self, num_chars, num_classes, num_steps=200, num_epochs=100, embedding_matrix=None, is_training=True, is_crf=True, weight=False):
        # Parameter
        self.max_f1 = 0
        self.learning_rate = 0.002
        self.dropout_rate = 0.5
        self.batch_size = 128
        self.num_layers = 1   
        self.emb_dim = 100
        self.hidden_dim = 100
        self.num_epochs = num_epochs
        self.num_steps = num_steps
        self.num_chars = num_chars
        self.num_classes = num_classes
        
        # placeholder of x, y and weight
        self.inputs = tf.placeholder(tf.int32, [None, self.num_steps])
        self.targets = tf.placeholder(tf.int32, [None, self.num_steps])
        self.targets_weight = tf.placeholder(tf.float32, [None, self.num_steps])
        self.targets_transition = tf.placeholder(tf.int32, [None])
        
        # char embedding
        if embedding_matrix != None:
            self.embedding = tf.Variable(embedding_matrix, trainable=False, name="emb", dtype=tf.float32)
        else:
            self.embedding = tf.get_variable("emb", [self.num_chars, self.emb_dim])
        self.inputs_emb = tf.nn.embedding_lookup(self.embedding, self.inputs)
        self.inputs_emb = tf.transpose(self.inputs_emb, [1, 0, 2])
        self.inputs_emb = tf.reshape(self.inputs_emb, [-1, self.emb_dim])
        self.inputs_emb = tf.split(0, self.num_steps, self.inputs_emb)

        # lstm cell
        lstm_cell_fw = tf.nn.rnn_cell.BasicLSTMCell(self.hidden_dim)
        lstm_cell_bw = tf.nn.rnn_cell.BasicLSTMCell(self.hidden_dim)

        # dropout
        if is_training:
            lstm_cell_fw = tf.nn.rnn_cell.DropoutWrapper(lstm_cell_fw, output_keep_prob=(1 - self.dropout_rate))
            lstm_cell_bw = tf.nn.rnn_cell.DropoutWrapper(lstm_cell_bw, output_keep_prob=(1 - self.dropout_rate))

        lstm_cell_fw = tf.nn.rnn_cell.MultiRNNCell([lstm_cell_fw] * self.num_layers)
        lstm_cell_bw = tf.nn.rnn_cell.MultiRNNCell([lstm_cell_bw] * self.num_layers)

        # get the length of each sample
        self.length = tf.reduce_sum(tf.sign(self.inputs), reduction_indices=1)
        self.length = tf.cast(self.length, tf.int32)  
        
        # forward and backward
        self.outputs, _, _ = rnn.bidirectional_rnn(
            lstm_cell_fw, 
            lstm_cell_bw,
            self.inputs_emb, 
            dtype=tf.float32,
            sequence_length=self.length
        )
        
        # softmax
        self.outputs = tf.reshape(tf.concat(1, self.outputs), [-1, self.hidden_dim * 2])
        self.softmax_w = tf.get_variable("softmax_w", [self.hidden_dim * 2, self.num_classes])
        self.softmax_b = tf.get_variable("softmax_b", [self.num_classes])
        self.logits = tf.matmul(self.outputs, self.softmax_w) + self.softmax_b

        if not is_crf:
            pass
        else:
            self.tags_scores = tf.reshape(self.logits, [self.batch_size, self.num_steps, self.num_classes])
            self.transitions = tf.get_variable("transitions", [self.num_classes + 1, self.num_classes + 1])
            
            dummy_val = -1000
            class_pad = tf.Variable(dummy_val * np.ones((self.batch_size, self.num_steps, 1)), dtype=tf.float32)
            self.observations = tf.concat(2, [self.tags_scores, class_pad])

            begin_vec = tf.Variable(np.array([[dummy_val] * self.num_classes + [0] for _ in range(self.batch_size)]), trainable=False, dtype=tf.float32)
            end_vec = tf.Variable(np.array([[0] + [dummy_val] * self.num_classes for _ in range(self.batch_size)]), trainable=False, dtype=tf.float32) 
            begin_vec = tf.reshape(begin_vec, [self.batch_size, 1, self.num_classes + 1])
            end_vec = tf.reshape(end_vec, [self.batch_size, 1, self.num_classes + 1])

            self.observations = tf.concat(1, [begin_vec, self.observations, end_vec])

            self.mask = tf.cast(tf.reshape(tf.sign(self.targets),[self.batch_size * self.num_steps]), tf.float32)
            
            # point score
            self.point_score = tf.gather(tf.reshape(self.tags_scores, [-1]), tf.range(0, self.batch_size * self.num_steps) * self.num_classes + tf.reshape(self.targets,[self.batch_size * self.num_steps]))
            self.point_score *= self.mask
            
            # transition score
            self.trans_score = tf.gather(tf.reshape(self.transitions, [-1]), self.targets_transition)
            
            # real score
            self.target_path_score = tf.reduce_sum(self.point_score) + tf.reduce_sum(self.trans_score)
            
            # all path score
            self.total_path_score, self.max_scores, self.max_scores_pre  = self.forward(self.observations, self.transitions, self.length)
            
            # loss
            self.loss = - (self.target_path_score - self.total_path_score)
        
        # summary
        self.train_summary = tf.scalar_summary("loss", self.loss)
        self.val_summary = tf.scalar_summary("loss", self.loss)        
        
        self.optimizer = tf.train.AdamOptimizer(learning_rate=self.learning_rate).minimize(self.loss) 

    def logsumexp(self, x, axis=None):
        x_max = tf.reduce_max(x, reduction_indices=axis, keep_dims=True)
        x_max_ = tf.reduce_max(x, reduction_indices=axis)
        return x_max_ + tf.log(tf.reduce_sum(tf.exp(x - x_max), reduction_indices=axis))

    def forward(self, observations, transitions, length, is_viterbi=True, return_best_seq=True):
        length = tf.reshape(length, [self.batch_size])
        transitions = tf.reshape(tf.concat(0, [transitions] * self.batch_size), [self.batch_size, 6, 6])
        observations = tf.reshape(observations, [self.batch_size, self.num_steps + 2, 6, 1])
        observations = tf.transpose(observations, [1, 0, 2, 3])
        previous = observations[0, :, :, :]
        max_scores = []
        max_scores_pre = []
        alphas = [previous]
        for t in range(1, self.num_steps + 2):
            previous = tf.reshape(previous, [self.batch_size, 6, 1])
            current = tf.reshape(observations[t, :, :, :], [self.batch_size, 1, 6])
            alpha_t = previous + current + transitions
            if is_viterbi:
                max_scores.append(tf.reduce_max(alpha_t, reduction_indices=1))
                max_scores_pre.append(tf.argmax(alpha_t, dimension=1))
            alpha_t = tf.reshape(self.logsumexp(alpha_t, axis=1), [self.batch_size, 6, 1])
            alphas.append(alpha_t)
            previous = alpha_t           
            
        alphas = tf.reshape(tf.concat(0, alphas), [self.num_steps + 2, self.batch_size, 6, 1])
        alphas = tf.transpose(alphas, [1, 0, 2, 3])
        alphas = tf.reshape(alphas, [self.batch_size * (self.num_steps + 2), 6, 1])

        last_alphas = tf.gather(alphas, tf.range(0, self.batch_size) * (self.num_steps + 2) + length)
        last_alphas = tf.reshape(last_alphas, [self.batch_size, 6, 1])

        max_scores = tf.reshape(tf.concat(0, max_scores), (self.num_steps + 1, self.batch_size, 6))
        max_scores_pre = tf.reshape(tf.concat(0, max_scores_pre), (self.num_steps + 1, self.batch_size, 6))
        max_scores = tf.transpose(max_scores, [1, 0, 2])
        max_scores_pre = tf.transpose(max_scores_pre, [1, 0, 2])

        return tf.reduce_sum(self.logsumexp(last_alphas, axis=1)), max_scores, max_scores_pre        

    def train(self, sess, save_file, X_train, y_train, X_val, y_val):
        saver = tf.train.Saver()

        char2id, id2char = helper.loadMap("char2id")
        label2id, id2label = helper.loadMap("label2id")

        merged = tf.merge_all_summaries()
        summary_writer_train = tf.train.SummaryWriter('loss_log/train_loss', sess.graph)  
        summary_writer_val = tf.train.SummaryWriter('loss_log/val_loss', sess.graph)     
        
        num_iterations = int(math.ceil(1.0 * len(X_train) / self.batch_size))

        cnt = 0
        for epoch in range(self.num_epochs):
            # shuffle train in each epoch
            sh_index = np.arange(len(X_train))
            np.random.shuffle(sh_index)
            X_train = X_train[sh_index]
            y_train = y_train[sh_index]
            print "current epoch: %d" % (epoch)
            for iteration in range(num_iterations):
                # train
                X_train_batch, y_train_batch = helper.nextBatch(X_train, y_train, start_index=iteration * self.batch_size, batch_size=self.batch_size)
                y_train_weight_batch = 1 + np.array((y_train_batch == label2id['B']) | (y_train_batch == label2id['E']), float)
                transition_batch = helper.getTransition(y_train_batch)
                
                _, loss_train, max_scores, max_scores_pre, length, train_summary =\
                    sess.run([
                        self.optimizer, 
                        self.loss, 
                        self.max_scores, 
                        self.max_scores_pre, 
                        self.length,
                        self.train_summary
                    ], 
                    feed_dict={
                        self.targets_transition:transition_batch, 
                        self.inputs:X_train_batch, 
                        self.targets:y_train_batch, 
                        self.targets_weight:y_train_weight_batch
                    })

                predicts_train = self.viterbi(max_scores, max_scores_pre, length, predict_size=self.batch_size)
                if iteration % 10 == 0:
                    cnt += 1
                    precision_train, recall_train, f1_train = self.evaluate(X_train_batch, y_train_batch, predicts_train, id2char, id2label)
                    summary_writer_train.add_summary(train_summary, cnt)
                    print "iteration: %5d, train loss: %5d, train precision: %.5f, train recall: %.5f, train f1: %.5f" % (iteration, loss_train, precision_train, recall_train, f1_train)  
                    
                # validation
                if iteration % 100 == 0:
                    X_val_batch, y_val_batch = helper.nextRandomBatch(X_val, y_val, batch_size=self.batch_size)
                    y_val_weight_batch = 1 + np.array((y_val_batch == label2id['B']) | (y_val_batch == label2id['E']), float)
                    transition_batch = helper.getTransition(y_val_batch)
                    
                    loss_val, max_scores, max_scores_pre, length, val_summary =\
                        sess.run([
                            self.loss, 
                            self.max_scores, 
                            self.max_scores_pre, 
                            self.length,
                            self.val_summary
                        ], 
                        feed_dict={
                            self.targets_transition:transition_batch, 
                            self.inputs:X_val_batch, 
                            self.targets:y_val_batch, 
                            self.targets_weight:y_val_weight_batch
                        })
                    
                    predicts_val = self.viterbi(max_scores, max_scores_pre, length, predict_size=self.batch_size)
                    precision_val, recall_val, f1_val = self.evaluate(X_val_batch, y_val_batch, predicts_val, id2char, id2label)
                    summary_writer_val.add_summary(val_summary, cnt)
                    print "iteration: %5d, valid loss: %5d, valid precision: %.5f, valid recall: %.5f, valid f1: %.5f" % (iteration, loss_val, precision_val, recall_val, f1_val)

                    if f1_val > self.max_f1:
                        self.max_f1 = f1_val
                        save_path = saver.save(sess, save_file)
                        print "saved the best model with f1: %.5f" % (self.max_f1)

    def test(self, sess, X_test, X_test_str, output_path):
        char2id, id2char = helper.loadMap("char2id")
        label2id, id2label = helper.loadMap("label2id")
        num_iterations = int(math.ceil(1.0 * len(X_test) / self.batch_size))
        print "number of iteration: " + str(num_iterations)
        with open(output_path, "wb") as outfile:
            for i in range(num_iterations):
                print "iteration: " + str(i + 1)
                results = []
                X_test_batch = X_test[i * self.batch_size : (i + 1) * self.batch_size]
                X_test_str_batch = X_test_str[i * self.batch_size : (i + 1) * self.batch_size]
                if i == num_iterations - 1 and len(X_test_batch) < self.batch_size:
                    X_test_batch = list(X_test_batch)
                    X_test_str_batch = list(X_test_str_batch)
                    last_size = len(X_test_batch)
                    X_test_batch += [[0 for j in range(self.num_steps)] for i in range(self.batch_size - last_size)]
                    X_test_str_batch += [['x' for j in range(self.num_steps)] for i in range(self.batch_size - last_size)]
                    X_test_batch = np.array(X_test_batch)
                    X_test_str_batch = np.array(X_test_str_batch)
                    results = self.predictBatch(sess, X_test_batch, X_test_str_batch, id2label)
                    results = results[:last_size]
                else:
                    X_test_batch = np.array(X_test_batch)
                    results = self.predictBatch(sess, X_test_batch, X_test_str_batch, id2label)
                
                for i in range(len(results)):
                    doc = ''.join(X_test_str_batch[i])
                    outfile.write(doc + "<@>" +" ".join(results[i]).encode("utf-8") + "\n")

    def viterbi(self, max_scores, max_scores_pre, length, predict_size=128):
        best_paths = []
        for m in range(predict_size):
            path = []
            last_max_node = np.argmax(max_scores[m][length[m]])
            # last_max_node = 0
            for t in range(1, length[m] + 1)[::-1]:
                last_max_node = max_scores_pre[m][t][last_max_node]
                path.append(last_max_node)
            path = path[::-1]
            best_paths.append(path)
        return best_paths

    def predictBatch(self, sess, X, X_str, id2label):
        results = []
        length, max_scores, max_scores_pre = sess.run([self.length, self.max_scores, self.max_scores_pre], feed_dict={self.inputs:X})
        predicts = self.viterbi(max_scores, max_scores_pre, length, self.batch_size)
        for i in range(len(predicts)):
            x = ''.join(X_str[i]).decode("utf-8")
            y_pred = ''.join([id2label[val] for val in predicts[i] if val != 5 and val != 0])
            entitys = helper.extractEntity(x, y_pred)
            results.append(entitys)
        return results

    def evaluate(self, X, y_true, y_pred, id2char, id2label):
        precision = -1.0
        recall = -1.0
        f1 = -1.0
        hit_num = 0
        pred_num = 0
        true_num = 0
        for i in range(len(y_true)):
            x = ''.join([str(id2char[val].encode("utf-8")) for val in X[i]])
            y = ''.join([str(id2label[val].encode("utf-8")) for val in y_true[i]])
            y_hat = ''.join([id2label[val] for val in y_pred[i]  if val != 5])
            true_labels = helper.extractEntity(x, y)
            pred_labels = helper.extractEntity(x, y_hat)
            hit_num += len(set(true_labels) & set(pred_labels))
            pred_num += len(set(pred_labels))
            true_num += len(set(true_labels))
        if pred_num != 0:
            precision = 1.0 * hit_num / pred_num
        if true_num != 0:
            recall = 1.0 * hit_num / true_num
        if precision > 0 and recall > 0:
            f1 = 2.0 * (precision * recall) / (precision + recall)
        return precision, recall, f1  

util.py

#encoding:utf-8
import re
import os
import csv
import time
import pickle
import numpy as np
import pandas as pd

def getEmbedding(infile_path="embedding"):
	char2id, id_char = loadMap("char2id")
	row_index = 0
	with open(infile_path, "rb") as infile:
		for row in infile:
			row = row.strip()
			row_index += 1
			if row_index == 1:
				num_chars = int(row.split()[0])
				emb_dim = int(row.split()[1])
				emb_matrix = np.zeros((len(char2id.keys()), emb_dim))
				continue
			items = row.split()
			char = items[0]
			emb_vec = [float(val) for val in items[1:]]
			if char in char2id:
				emb_matrix[char2id[char]] = emb_vec
	return emb_matrix

def nextBatch(X, y, start_index, batch_size=128):
    last_index = start_index + batch_size
    X_batch = list(X[start_index:min(last_index, len(X))])
    y_batch = list(y[start_index:min(last_index, len(X))])
    if last_index > len(X):
        left_size = last_index - (len(X))
        for i in range(left_size):
            index = np.random.randint(len(X))
            X_batch.append(X[index])
            y_batch.append(y[index])
    X_batch = np.array(X_batch)
    y_batch = np.array(y_batch)
    return X_batch, y_batch

def nextRandomBatch(X, y, batch_size=128):
    X_batch = []
    y_batch = []
    for i in range(batch_size):
        index = np.random.randint(len(X))
        X_batch.append(X[index])
        y_batch.append(y[index])
    X_batch = np.array(X_batch)
    y_batch = np.array(y_batch)
    return X_batch, y_batch

# use "0" to padding the sentence
def padding(sample, seq_max_len):
	for i in range(len(sample)):
		if len(sample[i]) < seq_max_len:
			sample[i] += [0 for _ in range(seq_max_len - len(sample[i]))]
	return sample

def prepare(chars, labels, seq_max_len, is_padding=True):
	X = []
	y = []
	tmp_x = []
	tmp_y = []

	for record in zip(chars, labels):
		c = record[0]
		l = record[1]
		# empty line
		if c == -1:
			if len(tmp_x) <= seq_max_len:
				X.append(tmp_x)
				y.append(tmp_y)
			tmp_x = []
			tmp_y = []
		else:
			tmp_x.append(c)
			tmp_y.append(l)	
	if is_padding:
		X = np.array(padding(X, seq_max_len))
	else:
		X = np.array(X)
	y = np.array(padding(y, seq_max_len))

	return X, y

def extractEntity(sentence, labels):
    entitys = []
    re_entity = re.compile(r'BM*E')
    m = re_entity.search(labels)
    while m:
        entity_labels = m.group()
        start_index = labels.find(entity_labels)
        entity = sentence[start_index:start_index + len(entity_labels)]
        labels = list(labels)
        # replace the "BM*E" with "OO*O"
        labels[start_index: start_index + len(entity_labels)] = ['O' for i in range(len(entity_labels))] 
        entitys.append(entity)
        labels = ''.join(labels)
        m = re_entity.search(labels)
    return entitys

def loadMap(token2id_filepath):
	if not os.path.isfile(token2id_filepath):
		print "file not exist, building map"
		buildMap()

	token2id = {}
	id2token = {}
	with open(token2id_filepath) as infile:
		for row in infile:
			row = row.rstrip().decode("utf-8")
			token = row.split('\t')[0]
			token_id = int(row.split('\t')[1])
			token2id[token] = token_id
			id2token[token_id] = token
	return token2id, id2token

def saveMap(id2char, id2label):
	with open("char2id", "wb") as outfile:
		for idx in id2char:
			outfile.write(id2char[idx] + "\t" + str(idx)  + "\r\n")
	with open("label2id", "wb") as outfile:
		for idx in id2label:
			outfile.write(id2label[idx] + "\t" + str(idx) + "\r\n")
	print "saved map between token and id"

def buildMap(train_path="train.in"):
	df_train = pd.read_csv(train_path, delimiter='\t', quoting=csv.QUOTE_NONE, skip_blank_lines=False, header=None, names=["char", "label"])
	chars = list(set(df_train["char"][df_train["char"].notnull()]))
	labels = list(set(df_train["label"][df_train["label"].notnull()]))
	char2id = dict(zip(chars, range(1, len(chars) + 1)))
	label2id = dict(zip(labels, range(1, len(labels) + 1)))
	id2char = dict(zip(range(1, len(chars) + 1), chars))
	id2label =  dict(zip(range(1, len(labels) + 1), labels))
	id2char[0] = ""
	id2label[0] = ""
	char2id[""] = 0
	label2id[""] = 0
	id2char[len(chars) + 1] = ""
	char2id[""] = len(chars) + 1

	saveMap(id2char, id2label)
	
	return char2id, id2char, label2id, id2label

def getTrain(train_path, val_path, train_val_ratio=0.99, use_custom_val=False, seq_max_len=200):
	char2id, id2char, label2id, id2label = buildMap(train_path)
	df_train = pd.read_csv(train_path, delimiter='\t', quoting=csv.QUOTE_NONE, skip_blank_lines=False, header=None, names=["char", "label"])

	# map the char and label into id
	df_train["char_id"] = df_train.char.map(lambda x : -1 if str(x) == str(np.nan) else char2id[x])
	df_train["label_id"] = df_train.label.map(lambda x : -1 if str(x) == str(np.nan) else label2id[x])
	
	# convert the data in maxtrix
	X, y = prepare(df_train["char_id"], df_train["label_id"], seq_max_len)

	# shuffle the samples
	num_samples = len(X)
	indexs = np.arange(num_samples)
	np.random.shuffle(indexs)
	X = X[indexs]
	y = y[indexs]
	
	if val_path != None:
		X_train = X
		y_train = y	
		X_val, y_val = getTest(val_path, is_validation=True, seq_max_len=seq_max_len)
	else:
		# split the data into train and validation set
		X_train = X[:int(num_samples * train_val_ratio)]
		y_train = y[:int(num_samples * train_val_ratio)]
		X_val = X[int(num_samples * train_val_ratio):]
		y_val = y[int(num_samples * train_val_ratio):]

	print "train size: %d, validation size: %d" %(len(X_train), len(y_val))

	return X_train, y_train, X_val, y_val

def getTest(test_path="test.in", is_validation=False, seq_max_len=200):
	char2id, id2char = loadMap("char2id")
	label2id, id2label = loadMap("label2id")

	df_test = pd.read_csv(test_path, delimiter='\t', quoting=csv.QUOTE_NONE, skip_blank_lines=False, header=None, names=["char", "label"])
	
	def mapFunc(x, char2id):
		if str(x) == str(np.nan):
			return -1
		elif x.decode("utf-8") not in char2id:
			return char2id[""]
		else:
			return char2id[x.decode("utf-8")]

	df_test["char_id"] = df_test.char.map(lambda x:mapFunc(x, char2id))
	df_test["label_id"] = df_test.label.map(lambda x : -1 if str(x) == str(np.nan) else label2id[x])
	
	if is_validation:
		X_test, y_test = prepare(df_test["char_id"], df_test["label_id"], seq_max_len)
		return X_test, y_test
	else:
		df_test["char"] = df_test.char.map(lambda x : -1 if str(x) == str(np.nan) else x)
		X_test, _ = prepare(df_test["char_id"], df_test["char_id"], seq_max_len)
		X_test_str, _ = prepare(df_test["char"], df_test["char_id"], seq_max_len, is_padding=False)
		print "test size: %d" %(len(X_test))
		return X_test, X_test_str

def getTransition(y_train_batch):
	transition_batch = []
	for m in range(len(y_train_batch)):
		y = [5] + list(y_train_batch[m]) + [0]
		for t in range(len(y)):
			if t + 1 == len(y):
				continue
			i = y[t]
			j = y[t + 1]
			if i == 0:
				break
			transition_batch.append(i * 6 + j)
	transition_batch = np.array(transition_batch)
	return transition_batch

train.py

import time
import helper
import argparse
import numpy as np
import pandas as pd
import tensorflow as tf
from BILSTM_CRF import BILSTM_CRF

# python train.py train.in model -v validation.in -c char_emb -e 10 -g 2

parser = argparse.ArgumentParser()
parser.add_argument("train_path", help="the path of the train file")
parser.add_argument("save_path", help="the path of the saved model")
parser.add_argument("-v","--val_path", help="the path of the validation file", default=None)
parser.add_argument("-e","--epoch", help="the number of epoch", default=100, type=int)
parser.add_argument("-c","--char_emb", help="the char embedding file", default=None)
parser.add_argument("-g","--gpu", help="the id of gpu, the default is 0", default=0, type=int)

args = parser.parse_args()

train_path = args.train_path
save_path = args.save_path
val_path = args.val_path
num_epochs = args.epoch
emb_path = args.char_emb
gpu_config = "/cpu:0"
#gpu_config = "/gpu:"+str(args.gpu)
num_steps = 200 # it must consist with the test

start_time = time.time()
print "preparing train and validation data"
X_train, y_train, X_val, y_val = helper.getTrain(train_path=train_path, val_path=val_path, seq_max_len=num_steps)
char2id, id2char = helper.loadMap("char2id")
label2id, id2label = helper.loadMap("label2id")
num_chars = len(id2char.keys())
num_classes = len(id2label.keys())
if emb_path != None:
	embedding_matrix = helper.getEmbedding(emb_path)
else:
	embedding_matrix = None

print "building model"
config = tf.ConfigProto(allow_soft_placement=True)
with tf.Session(config=config) as sess:
	with tf.device(gpu_config):
		initializer = tf.random_uniform_initializer(-0.1, 0.1)
		with tf.variable_scope("model", reuse=None, initializer=initializer):
			model = BILSTM_CRF(num_chars=num_chars, num_classes=num_classes, num_steps=num_steps, num_epochs=num_epochs, embedding_matrix=embedding_matrix, is_training=True)

		print "training model"
		tf.initialize_all_variables().run()
		model.train(sess, save_path, X_train, y_train, X_val, y_val)

		print "final best f1 is: %f" % (model.max_f1)

		end_time = time.time()
		print "time used %f(hour)" % ((end_time - start_time) / 3600)

test.py

import time
import helper
import argparse
import numpy as np
import pandas as pd
import tensorflow as tf
from BILSTM_CRF import BILSTM_CRF

# python test.py model test.in test.out -c char_emb -g 2

parser = argparse.ArgumentParser()
parser.add_argument("model_path", help="the path of model file")
parser.add_argument("test_path", help="the path of test file")
parser.add_argument("output_path", help="the path of output file")
parser.add_argument("-c","--char_emb", help="the char embedding file", default=None)
parser.add_argument("-g","--gpu", help="the id of gpu, the default is 0", default=0, type=int)
args = parser.parse_args()

model_path = args.model_path
test_path = args.test_path
output_path = args.output_path
gpu_config = "/cpu:0"
emb_path = args.char_emb
num_steps = 200 # it must consist with the train

start_time = time.time()

print "preparing test data"
X_test, X_test_str = helper.getTest(test_path=test_path, seq_max_len=num_steps)
char2id, id2char = helper.loadMap("char2id")
label2id, id2label = helper.loadMap("label2id")
num_chars = len(id2char.keys())
num_classes = len(id2label.keys())
if emb_path != None:
	embedding_matrix = helper.getEmbedding(emb_path)
else:
	embedding_matrix = None

print "building model"
config = tf.ConfigProto(allow_soft_placement=True)
with tf.Session(config=config) as sess:
	with tf.device(gpu_config):
		initializer = tf.random_uniform_initializer(-0.1, 0.1)
		with tf.variable_scope("model", reuse=None, initializer=initializer):
			model = BILSTM_CRF(num_chars=num_chars, num_classes=num_classes, num_steps=num_steps, embedding_matrix=embedding_matrix, is_training=False)

		print "loading model parameter"
		saver = tf.train.Saver()
		saver.restore(sess, model_path)

		print "testing"
		model.test(sess, X_test, X_test_str, output_path)

		end_time = time.time()
		print "time used %f(hour)" % ((end_time - start_time) / 3600)

相关预处理的数据请参考github: scofiled's github/bilstm+crf