首页 > 代码库 > TensorFlow (RNN)深度学习 双向LSTM(BiLSTM)+CRF 实现 sequence labeling 序列标注问题 源码下载
TensorFlow (RNN)深度学习 双向LSTM(BiLSTM)+CRF 实现 sequence labeling 序列标注问题 源码下载
http://blog.csdn.net/scotfield_msn/article/details/60339415
在TensorFlow (RNN)深度学习下 双向LSTM(BiLSTM)+CRF 实现 sequence labeling
双向LSTM+CRF跑序列标注问题
源码下载
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.考虑将特定的人工提取的规则融入到NN模型中去。
好了,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
注明:codes参考的是chilynn
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] = "<PAD>"
- id2label[0] = "<PAD>"
- char2id["<PAD>"] = 0
- label2id["<PAD>"] = 0
- id2char[len(chars) + 1] = "<NEW>"
- char2id["<NEW>"] = 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["<NEW>"]
- 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
TensorFlow (RNN)深度学习 双向LSTM(BiLSTM)+CRF 实现 sequence labeling 序列标注问题 源码下载