好久没有写博客了，这一次就将最近看的pytorch 教程中的lstm+crf的一些心得与困惑记录下来。
原文 PyTorch Tutorials
参考了很多其他大神的博客，https://blog.csdn.net/cuihuijun1hao/article/details/79405740
https://www.jianshu.com/p/97cb3b6db573
至于原理，非常建议读这篇英文博客，写的非常非常非常好！！！！！！值得打印出来细细品读！！！！！！！！！！！！！！！！！！！！！！！！！！！！！！！https://createmomo.github.io/2017/09/12/CRF_Layer_on_the_Top_of_BiLSTM_1/
在这位大神的基础上，根据自己的debug又添加了一些注释pytorch版的bilstm+crf实现sequence label
为了方便理解：
注意，
self.transitions = nn.Parameter(torch.randn(self.tagset_size, self.tagset_size)) 说明了转移矩阵是随机的！！！随机的！！！随机的！！！，而且放入了网络中，会更新的！！！会更新的！！！会更新的！！！
解释一下重点的函数功能：
def log_sum_exp(vec) 这个函数，是一个封装好的数学公式，里面先做减法的原因在于，减去最大值可以避免e的指数次，计算机上溢。
def _forward_alg(self, feats): 这个函数，只是根据 随机的transitions ，前向传播算出的一个score，用到了动态规划的思想，但是因为用的是随机的转移矩阵，算出的值很大 score>20
def _get_lstm_features(self, sentence): 可以看出，函数里经过了embedding，lstm，linear层，是根据LSTM算出的一个矩阵。这里是11x5的一个tensor，而这个11x5的tensor，就是发射矩阵！！！发射矩阵！！！发射矩阵！！！（emission matrix）
def _score_sentence(self, feats, tags):是根据真实的标签算出的一个score，这与上面的def _forward_alg(self, feats)有什么不同的地方嘛？共同之处在于，两者都是用的随机的转移矩阵算的score，但是不同地方在于，上面那个函数算了一个最大可能路径，但是实际上可能不是真实的 各个标签转移的值。例如说，真实的标签 是 N V V，但是因为transitions是随机的，所以上面的函数得到的其实是N N N这样，两者之间的score就有了差距。而后来的反向传播，就能够更新transitions，使得转移矩阵逼近真实的“转移矩阵”。（个人理解）
def _viterbi_decode(self, feats):维特比解码，实际上就是在预测的时候使用了，输出得分与路径值。
这个函数是重点：
def neg_log_likelihood(self, sentence, tags):
feats = self._get_lstm_features(sentence)#11*5 经过了LSTM+Linear矩阵后的输出，之后作为CRF的输入。
forward_score = self._forward_alg(feats) #0维的一个得分，20.*来着
gold_score = self._score_sentence(feats, tags)#tensor([ 4.5836])
return forward_score - gold_score #这是两者之间的差值，后来直接根据这个差值，反向传播。。。神奇！！！！！！
def forward(self, sentence):forward函数只是用来预测了，train的时候没用调用它，这让我感到很震惊，还有这种操作？
import torch
import torch.autograd as autograd
import torch.nn as nn
import torch.optim as optim
def to_scalar(var): #var是Variable,维度是１
# returns a python float
return var.view(-1).data.tolist()[0]
def argmax(vec):
# return the argmax as a python int
_, idx = torch.max(vec, 1)
return to_scalar(idx)
def prepare_sequence(seq, to_ix):
idxs = [to_ix[w] for w in seq]
tensor = torch.LongTensor(idxs)
return autograd.Variable(tensor)
# Compute log sum exp in a numerically stable way for the forward algorithm
def log_sum_exp(vec): #vec是1*5, type是Variable
max_score = vec[0, argmax(vec)]
#max_score维度是１， max_score.view(1,-1)维度是１＊１，max_score.view(1, -1).expand(1, vec.size()[1])的维度是１＊５
max_score_broadcast = max_score.view(1, -1).expand(1, vec.size()[1]) # vec.size()维度是1*5
return max_score + torch.log(torch.sum(torch.exp(vec - max_score_broadcast)))#为什么指数之后再求和，而后才log呢
class BiLSTM_CRF(nn.Module):
def __init__(self, vocab_size, tag_to_ix, embedding_dim, hidden_dim):
super(BiLSTM_CRF, self).__init__()
self.embedding_dim = embedding_dim
self.hidden_dim = hidden_dim
self.vocab_size = vocab_size
self.tag_to_ix = tag_to_ix
self.tagset_size = len(tag_to_ix)
self.word_embeds = nn.Embedding(vocab_size, embedding_dim)
self.lstm = nn.LSTM(embedding_dim, hidden_dim // 2, num_layers=1, bidirectional=True)
# Maps the output of the LSTM into tag space.
self.hidden2tag = nn.Linear(hidden_dim, self.tagset_size)
# Matrix of transition parameters. Entry i,j is the score of
# transitioning *to* i *from* j. 居然是随机初始化的！！！！！！！！！！！！！！！之后的使用也是用这随机初始化的值进行操作！！
self.transitions = nn.Parameter(torch.randn(self.tagset_size, self.tagset_size))
# These two statements enforce the constraint that we never transfer
# to the start tag and we never transfer from the stop tag
self.transitions.data[tag_to_ix[START_TAG], :] = -10000
self.transitions.data[:, tag_to_ix[STOP_TAG]] = -10000
self.hidden = self.init_hidden()
def init_hidden(self):
return (autograd.Variable(torch.randn(2, 1, self.hidden_dim // 2)),
autograd.Variable(torch.randn(2, 1, self.hidden_dim // 2)))
#预测序列的得分
def _forward_alg(self, feats):
# Do the forward algorithm to compute the partition function
init_alphas = torch.Tensor(1, self.tagset_size).fill_(-10000.) #1*5 而且全是-10000
# START_TAG has all of the score.
init_alphas[0][self.tag_to_ix[START_TAG]] = 0. #因为start tag是4，所以tensor([[-10000., -10000., -10000., 0., -10000.]])，将start的值为零，表示开始进行网络的传播，
# Wrap in a variable so that we will get automatic backprop
forward_var = autograd.Variable(init_alphas) #初始状态的forward_var，随着step t变化
# Iterate through the sentence 会迭代feats的行数次，
for feat in feats: #feat的维度是５ 依次把每一行取出来~
alphas_t = [] # The forward variables at this timestep
for next_tag in range(self.tagset_size):#next tag 就是简单 i，从0到len
# broadcast the emission score: it is the same regardless of
# the previous tag
emit_score = feat[next_tag].view(1, -1).expand(1, self.tagset_size) #维度是1*5 噢噢！原来，LSTM后的那个矩阵，就被当做是emit score了
# the ith entry of trans_score is the score of transitioning to
# next_tag from i
trans_score = self.transitions[next_tag].view(1, -1) #维度是１＊５
# The ith entry of next_tag_var is the value for the
# edge (i -> next_tag) before we do log-sum-exp
#第一次迭代时理解：
# trans_score所有其他标签到Ｂ标签的概率
# 由lstm运行进入隐层再到输出层得到标签Ｂ的概率，emit_score维度是１＊５，5个值是相同的
next_tag_var = forward_var + trans_score + emit_score
# The forward variable for this tag is log-sum-exp of all the
# scores.
alphas_t.append(log_sum_exp(next_tag_var).unsqueeze(0))
#此时的alphas t 是一个长度为5，例如<class 'list'>: [tensor(0.8259), tensor(2.1739), tensor(1.3526), tensor(-9999.7168), tensor(-0.7102)]
forward_var = torch.cat(alphas_t).view(1, -1)#到第（t-1）step时５个标签的各自分数
terminal_var = forward_var + self.transitions[self.tag_to_ix[STOP_TAG]] #最后只将最后一个单词的forward var与转移 stop tag的概率相加 tensor([[ 21.1036, 18.8673, 20.7906, -9982.2734, -9980.3135]])
alpha = log_sum_exp(terminal_var) #alpha是一个0维的tensor
return alpha
#得到feats
def _get_lstm_features(self, sentence):
self.hidden = self.init_hidden()
#embeds = self.word_embeds(sentence).view(len(sentence), 1, -1)
embeds = self.word_embeds(sentence)
embeds = embeds.unsqueeze(1)
lstm_out, self.hidden = self.lstm(embeds, self.hidden)#11*1*4
lstm_out = lstm_out.view(len(sentence), self.hidden_dim) #11*4
lstm_feats = self.hidden2tag(lstm_out)#11*5 is a linear layer
return lstm_feats
#得到gold_seq tag的score 即根据真实的label 来计算一个score，但是因为转移矩阵是随机生成的，故算出来的score不是最理想的值
def _score_sentence(self, feats, tags):
# Gives the score of a provided tag sequence #feats 11*5 tag 11 维
score = autograd.Variable(torch.Tensor([0]))
tags = torch.cat([torch.LongTensor([self.tag_to_ix[START_TAG]]), tags]) #将START_TAG的标签３拼接到tag序列最前面，这样tag就是12个了
for i, feat in enumerate(feats):
#self.transitions[tags[i + 1], tags[i]] 实际得到的是从标签i到标签i+1的转移概率
#feat[tags[i+1]], feat是step i 的输出结果，有５个值，对应B, I, E, START_TAG, END_TAG, 取对应标签的值
#transition【j,i】 就是从i ->j 的转移概率值
score = score + self.transitions[tags[i + 1], tags[i]] + feat[tags[i + 1]]
score = score + self.transitions[self.tag_to_ix[STOP_TAG], tags[-1]]
return score
#解码，得到预测的序列，以及预测序列的得分
def _viterbi_decode(self, feats):
backpointers = []
# Initialize the viterbi variables in log space
init_vvars = torch.Tensor(1, self.tagset_size).fill_(-10000.)
init_vvars[0][self.tag_to_ix[START_TAG]] = 0
# forward_var at step i holds the viterbi variables for step i-1
forward_var = autograd.Variable(init_vvars)
for feat in feats:
bptrs_t = [] # holds the backpointers for this step
viterbivars_t = [] # holds the viterbi variables for this step
for next_tag in range(self.tagset_size):
# next_tag_var[i] holds the viterbi variable for tag i at the
# previous step, plus the score of transitioning
# from tag i to next_tag.
# We don't include the emission scores here because the max
# does not depend on them (we add them in below)
next_tag_var = forward_var + self.transitions[next_tag] #其他标签（B,I,E,Start,End）到标签next_tag的概率
best_tag_id = argmax(next_tag_var)
bptrs_t.append(best_tag_id)
viterbivars_t.append(next_tag_var[0][best_tag_id].view(1))
# Now add in the emission scores, and assign forward_var to the set
# of viterbi variables we just computed
forward_var = (torch.cat(viterbivars_t) + feat).view(1, -1)#从step0到step(i-1)时5个序列中每个序列的最大score
backpointers.append(bptrs_t) #bptrs_t有５个元素
# Transition to STOP_TAG
terminal_var = forward_var + self.transitions[self.tag_to_ix[STOP_TAG]]#其他标签到STOP_TAG的转移概率
best_tag_id = argmax(terminal_var)
path_score = terminal_var[0][best_tag_id]
# Follow the back pointers to decode the best path.
best_path = [best_tag_id]
for bptrs_t in reversed(backpointers):#从后向前走，找到一个best路径
best_tag_id = bptrs_t[best_tag_id]
best_path.append(best_tag_id)
# Pop off the start tag (we dont want to return that to the caller)
start = best_path.pop()
assert start == self.tag_to_ix[START_TAG] # Sanity check
best_path.reverse()# 把从后向前的路径正过来
return path_score, best_path
def neg_log_likelihood(self, sentence, tags):
feats = self._get_lstm_features(sentence)#11*5 经过了LSTM+Linear矩阵后的输出，之后作为CRF的输入。
forward_score = self._forward_alg(feats) #0维的一个得分，20.*来着
gold_score = self._score_sentence(feats, tags)#tensor([ 4.5836])
return forward_score - gold_score
def forward(self, sentence): # dont confuse this with _forward_alg above.
# Get the emission scores from the BiLSTM
lstm_feats = self._get_lstm_features(sentence)
# Find the best path, given the features.
score, tag_seq = self._viterbi_decode(lstm_feats)
return score, tag_seq
START_TAG = "<START>"
STOP_TAG = "<STOP>"
EMBEDDING_DIM = 5
HIDDEN_DIM = 4
# Make up some training data
training_data = [("the wall street journal reported today that apple corporation made money".split(), "B I I I O O O B I O O".split()),
("georgia tech is a university in georgia".split(), "B I O O O O B".split())]
word_to_ix = {}
for sentence, tags in training_data:
for word in sentence:
if word not in word_to_ix:
word_to_ix[word] = len(word_to_ix)
tag_to_ix = {"B": 0, "I": 1, "O": 2, START_TAG: 3, STOP_TAG: 4}
model = BiLSTM_CRF(len(word_to_ix), tag_to_ix, EMBEDDING_DIM, HIDDEN_DIM)
optimizer = optim.SGD(model.parameters(), lr=0.01, weight_decay=1e-4)
# Check predictions before training
# precheck_sent = prepare_sequence(training_data[0][0], word_to_ix)
# precheck_tags = torch.LongTensor([tag_to_ix[t] for t in training_data[0][1]])
# print(model(precheck_sent))
# Make sure prepare_sequence from earlier in the LSTM section is loaded
for epoch in range(1): # again, normally you would NOT do 300 epochs, it is toy data
for sentence, tags in training_data:
# Step 1. Remember that Pytorch accumulates gradients.
# We need to clear them out before each instance
model.zero_grad()
# Step 2. Get our inputs ready for the network, that is,
# turn them into Variables of word indices.
sentence_in = prepare_sequence(sentence, word_to_ix)
targets = torch.LongTensor([tag_to_ix[t] for t in tags])
# Step 3. Run our forward pass.
neg_log_likelihood = model.neg_log_likelihood(sentence_in, targets)#tensor([ 15.4958]) 最大的可能的值与 根据随机转移矩阵 计算的真实值 的差
# Step 4. Compute the loss, gradients, and update the parameters by
# calling optimizer.step()
neg_log_likelihood.backward()#卧槽，这就能更新啦？？？进行了反向传播，算了梯度值。debug中可以看到，transition的_grad 有了值 torch.Size([5, 5])
optimizer.step()
# Check predictions after training
precheck_sent = prepare_sequence(training_data[0][0], word_to_ix)
print(model(precheck_sent)[0]) #得分
print('^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^')
print(model(precheck_sent)[1]) #tag sequence
心得的地方：
反向传播不需要一定使用forward()，而且不需要定义loss=nn.MSError()等，直接score1 - score2 ，就可以反向传播了。
无论两个矩阵你咋操作，只要满足，不管你是只取一行，还是几行，加减乘数。只要能够满足这个式子，就能够反向传播，前提是 self.transitions = nn.Parameter(torch.randn(self.tagset_size, self.tagset_size)) 将你想要更新的矩阵，放入到module的参数中，这样才能够更新。
即便你是这样用的：
for feat in feats: #feat的维度是５ 依次把每一行取出来~
alphas_t = [] # The forward variables at this timestep
for next_tag in range(self.tagset_size):#next tag 就是简单 i，从0到len
# broadcast the emission score: it is the same regardless of
# the previous tag
emit_score = feat[next_tag].view(1, -1).expand(1, self.tagset_size) #维度是1*5 噢噢！原来，LSTM后的那个矩阵，就被当做是emit score了
# the ith entry of trans_score is the score of transitioning to
# next_tag from i
trans_score = self.transitions[next_tag].view(1, -1) #维度是１＊５
# The ith entry of next_tag_var is the value for the
# edge (i -> next_tag) before we do log-sum-exp
#第一次迭代时理解：
# trans_score所有其他标签到Ｂ标签的概率
# 由lstm运行进入隐层再到输出层得到标签Ｂ的概率，emit_score维度是１＊５，5个值是相同的
next_tag_var = forward_var + trans_score + emit_score
# The forward variable for this tag is log-sum-exp of all the
# scores.
alphas_t.append(log_sum_exp(next_tag_var).unsqueeze(0))
#此时的alphas t 是一个长度为5，例如<class 'list'>: [tensor(0.8259), tensor(2.1739), tensor(1.3526), tensor(-9999.7168), tensor(-0.7102)]
forward_var = torch.cat(alphas_t).view(1, -1)#到第（t-1）step时５个标签的各自分数
terminal_var = forward_var + self.transitions[self.tag_to_ix[STOP_TAG]] #最后只将最后一个单词的forward var与转移 stop tag的概率相加 tensor([[ 21.1036, 18.8673, 20.7906, -9982.2734, -9980.3135]])
alpha = log_sum_exp(terminal_var) #alpha是一个0维的tensor
或者是这样用的
score = autograd.Variable(torch.Tensor([0]))
tags = torch.cat([torch.LongTensor([self.tag_to_ix[START_TAG]]), tags]) #将START_TAG的标签３拼接到tag序列最前面，这样tag就是12个了
for i, feat in enumerate(feats):
#self.transitions[tags[i + 1], tags[i]] 实际得到的是从标签i到标签i+1的转移概率
#feat[tags[i+1]], feat是step i 的输出结果，有５个值，对应B, I, E, START_TAG, END_TAG, 取对应标签的值
#transition【j,i】 就是从i ->j 的转移概率值
score = score + self.transitions[tags[i + 1], tags[i]] + feat[tags[i + 1]]
score = score + self.transitions[self.tag_to_ix[STOP_TAG], tags[-1]]
return score
看样子，每个循环里只是去了转移矩阵的一行，或者就是一个值，进行操作，但是！转移矩阵就是能够更新！！！至于为什么能够更新！！我不知道：（
----------------------------------------------------------------------------------------------------------------------------------------------------------------------
2018.12.12
心得，发射矩阵是 lstm算出来的，是要通过网络学习的。转移矩阵是单独定义的，要学习的。初始矩阵，是[-1000,-1000,-1000,0,-1000]固定的，因为当加了开始符号后，则第一个位置是开始符号的概率是100%。
显式的加入了start标记，隐式的使用了end标记（总是最后多一步转移到end）的分数
----------------------------------------------------------------------------------------------------------------------------------------------------------------------
以上都是个人理解，恳请勘误！
---------------------
作者：Jason__Liang
来源：CSDN
原文：https://blog.csdn.net/Jason__Liang/article/details/81772632
版权声明：本文为博主原创文章，转载请附上博文链接！