zoukankan      html  css  js  c++  java
  • pytorch lstm crf 代码理解

    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/

    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!

    Pytorch Bi-LSTM + CRF 代码详解
    阅读数 9317

    久闻LSTM+CRF的效果强大,最近在看Pytorch官网文档的时候,看到了这段代码,前前后后查了很多资料,终于把代码弄懂了。我希望在后来人看这段代码的时候,直接就看我的博客就能完全弄懂这段代码。看这...
    博文
    来自: Call Me Hi Johnny~~


    ancient_wizard_wjs: 你好,请问一下你知道如何将完整的代码在GPU上跑通吗,我的老是报错
    Traceback (most recent call last):
    File "lstm_crf.py", line 281, in <module>
    loss = model.neg_log_likelihood(sentence_in, targets)
    File "lstm_crf.py", line 205, in neg_log_likelihood
    feats = self._get_lstm_features(sentence)
    File "lstm_crf.py", line 120, in _get_lstm_features
    lstm_out, self.hidden = self.lstm(embeds, self.hidden) # 11*1*4
    File "/home/sjwang/py/python3/lib/python3.6/site-packages/torch/nn/modules/module.py", line 489, in __call__
    result = self.forward(*input, **kwargs)
    File "/home/sjwang/py/python3/lib/python3.6/site-packages/torch/nn/modules/rnn.py", line 179, in forward
    self.dropout, self.training, self.bidirectional, self.batch_first)
    RuntimeError: Input and hidden tensors are not at the same device, found input tensor at cuda:0 and hidden tensor at cpu
    (3个月前#3楼)查看回复(1)

    for_android_d: 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。博主看明白这一行代码了吗,特别是vec-max_score_broadcast不知道是怎么回事,可以提供一个推导过程吗
    ---------------------
    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!

  • 相关阅读:
    IO模型
    协程
    线程
    进程总结
    HashMap和Hashtable有什么区别
    HashMap 1.7 与 1.8 的 区别,说明 1.8 做了哪些优化,如何优化的
    GC线程是否为守护线程?
    float f=3.4;是否正确?
    final、finally和finalized的区别?
    Eureka注册中心是什么?
  • 原文地址:https://www.cnblogs.com/jfdwd/p/11162752.html
Copyright © 2011-2022 走看看