tensorflow笔记3：CRF函数：tf.contrib.crf.crf_log_likelihood()

在分析训练代码的时候，遇到了，tf.contrib.crf.crf_log_likelihood，这个函数，于是想简单理解下：

函数的目的：使用crf 来计算损失，里面用到的优化方法是：最大似然估计

使用方法：

tf.contrib.crf.crf_log_likelihood(inputs, tag_indices, sequence_lengths, transition_params=None)
See the guide: CRF (contrib)

Computes the log-likelihood of tag sequences in a CRF.

Args:
inputs: A [batch_size, max_seq_len, num_tags] tensor of unary potentials to use as input to the CRF layer.
tag_indices: A [batch_size, max_seq_len] matrix of tag indices for which we compute the log-likelihood.
sequence_lengths: A [batch_size] vector of true sequence lengths.
transition_params: A [num_tags, num_tags] transition matrix, if available. Returns:
log_likelihood: A scalar containing the log-likelihood of the given sequence of tag indices.
transition_params: A [num_tags, num_tags] transition matrix. This is either provided by the caller or created in this function.

函数讲解：

1、tf.contrib.crf.crf_log_likelihood

crf_log_likelihood(inputs,tag_indices,sequence_lengths,transition_params=None)

在一个条件随机场里面计算标签序列的log-likelihood

参数:

inputs: 一个形状为[batch_size, max_seq_len, num_tags] 的tensor,一般使用BILSTM处理之后输出转换为他要求的形状作为CRF层的输入.
tag_indices: 一个形状为[batch_size, max_seq_len] 的矩阵,其实就是真实标签.
sequence_lengths: 一个形状为 [batch_size] 的向量,表示每个序列的长度.
transition_params: 形状为[num_tags, num_tags] 的转移矩阵

返回：

log_likelihood: 标量,log-likelihood
transition_params: 形状为[num_tags, num_tags] 的转移矩阵

2、tf.contrib.crf.viterbi_decode

viterbi_decode(score,transition_params) 
通俗一点,作用就是返回最好的标签序列.这个函数只能够在测试时使用,在tensorflow外部解码

参数:

score: 一个形状为[seq_len, num_tags] matrix of unary potentials.
transition_params: 形状为[num_tags, num_tags] 的转移矩阵

返回：

viterbi: 一个形状为[seq_len] 显示了最高分的标签索引的列表.
viterbi_score: A float containing the score for the Viterbi sequence.

3、tf.contrib.crf.crf_decode

crf_decode(potentials,transition_params,sequence_length) 
在tensorflow内解码

参数:

potentials: 一个形状为[batch_size, max_seq_len, num_tags] 的tensor,
transition_params: 一个形状为[num_tags, num_tags] 的转移矩阵
sequence_length: 一个形状为[batch_size] 的 ,表示batch中每个序列的长度

返回：

decode_tags:一个形状为[batch_size, max_seq_len] 的tensor,类型是tf.int32.表示最好的序列标记.
best_score: 有个形状为[batch_size] 的tensor, 包含每个序列解码标签的分数.

转载来自知乎：

如果你需要预测的是个序列，那么可以选择用crf_log_likelihood作为损失函数

crf_log_likelihood(
inputs,
tag_indices,
sequence_lengths,
transition_params=None
)

输入：

inputs：unary potentials，也就是每个标签的预测概率值，这个值根据实际情况选择计算方法，CNN,RNN...都可以

tag_indices，这个就是真实的标签序列了

sequence_lengths，这是一个样本真实的序列长度，因为为了对齐长度会做些padding，但是可以把真实的长度放到这个参数里

transition_params，转移概率，可以没有，没有的话这个函数也会算出来

输出：

log_likelihood,

transition_params,转移概率，如果输入没输，它就自己算个给返回

作者：知乎用户
链接：https://www.zhihu.com/question/57666556/answer/326803900
来源：知乎
著作权归作者所有。商业转载请联系作者获得授权，非商业转载请注明出处。

官方的示例代码：如何使用crf来计算：

# !/home/wcg/tools/local/anaconda3/bin/python
# coding=utf8
import numpy as np
import tensorflow as tf

#data settings
num_examples = 10
num_words = 20
num_features = 100
num_tags = 5 

# 5 tags
#x shape = [10,20,100]
#random features.
x = np.random.rand(num_examples,num_words,num_features).astype(np.float32)

#y shape = [10,20]

#Random tag indices representing the gold sequence.
y = np.random.randint(num_tags,size = [num_examples,num_words]).astype(np.int32)

# 序列的长度
#sequence_lengths = [19,19,19,19,19,19,19,19,19,19]
sequence_lengths = np.full(num_examples,num_words - 1,dtype=np.int32)

#Train and evaluate the model.
with tf.Graph().as_default():
    with tf.Session() as session:
         # Add the data to the TensorFlow gtaph.
         x_t = tf.constant(x) #观测序列
         y_t = tf.constant(y) # 标记序列
         sequence_lengths_t = tf.constant(sequence_lengths)

         # Compute unary scores from a linear layer.
         # weights shape = [100,5]
         weights = tf.get_variable("weights", [num_features, num_tags])

         # matricized_x_t shape = [200,100]
         matricized_x_t = tf.reshape(x_t, [-1, num_features])

         # compute                           [200,100]      [100,5]   get [200,5]
         # 计算结果
         matricized_unary_scores = tf.matmul(matricized_x_t, weights)

         #  unary_scores shape = [10,20,5]                  [10,20,5]
         unary_scores = tf.reshape(matricized_unary_scores, [num_examples, num_words, num_tags])
         # compute the log-likelihood of the gold sequences and keep the transition
         # params for inference at test time.
         #                                                shape      shape   [10,20,5]   [10,20]   [10]
         log_likelihood,transition_params = tf.contrib.crf.crf_log_likelihood(unary_scores,y_t,sequence_lengths_t)

         viterbi_sequence, viterbi_score = tf.contrib.crf.crf_decode(unary_scores, transition_params, sequence_lengths_t)
         # add a training op to tune the parameters.
         loss = tf.reduce_mean(-log_likelihood)

         # 定义梯度下降算法的优化器
         #learning_rate 0.01
         train_op = tf.train.GradientDescentOptimizer(0.01).minimize(loss)

         #train for a fixed number of iterations.
         session.run(tf.global_variables_initializer())

         '''
        #eg:
        In [61]: m_20
        Out[61]: array([[ 3,  4,  5,  6,  7,  8,  9, 10, 11, 12]])

        In [62]: n_20
        Out[62]: array([[0, 1, 2, 3, 4, 5, 6, 7, 8, 9]])

        In [59]: n_20<m_20
        Out[59]: array([[ True,  True,  True,  True,  True,  True,  True,  True,  True, True]], dtype=bool)

         '''
         #这里用mask过滤掉不符合的结果
         mask = (np.expand_dims(np.arange(num_words), axis=0) < np.expand_dims(sequence_lengths, axis=1))

         ###mask = array([[ True,  True,  True,  True,  True,  True,  True,  True,  True, True]], dtype=bool)
         #序列的长度
         total_labels = np.sum(sequence_lengths)

         print ("mask:",mask)

         print ("total_labels:",total_labels)
         for i in range(1000):
             #tf_unary_scores,tf_transition_params,_ = session.run([unary_scores,transition_params,train_op])
             tf_viterbi_sequence,_=session.run([viterbi_sequence,train_op])
             if i%100 == 0:
                '''
                false*false = false  false*true= false ture*true = true
                '''
                #序列中预测对的个数
                correct_labels = np.sum((y==tf_viterbi_sequence) *  mask)
                accuracy = 100.0*correct_labels/float(total_labels)
                print ("Accuracy: %.2f%%" %accuracy)