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
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)
