这仅仅是pytorch 给的一个 BiLSTM CRF简单示例 这里分析下源码方便对crf有个清晰的认识
开始分析代码
def argmax(vec):
# return the argmax as a python int
# idx 是 最大值所在的索引
# 找出向量的最大索引
_, idx = torch.max(vec, 1)
return idx.item()
def prepare_sequence(seq, to_ix):
# 输入的sentence转成 id
idxs = [to_ix[w] for w in seq]
return torch.tensor(idxs, dtype=torch.long)
# Compute log sum exp in a numerically stable way for the forward algorithm
def log_sum_exp(vec):
# vec shape (1, 5)
max_score = vec[0, argmax(vec)]
# 将取出的最大值扩展到和vec 同shape (1, 5)
max_score_broadcast = max_score.view(1, -1).expand(1, vec.size()[1])
# 这里减去最大值 在加上最大值 是为了防止log数据溢出
# max + log Σexp(xi - max) = max + log (Σexp(xi) / emax) = max + logΣexp(xi) - max
return max_score + \
torch.log(torch.sum(torch.exp(vec - max_score_broadcast)))接下来看 BiLSTM_CRF类
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)
# 定义词向量
self.word_embeds = nn.Embedding(vocab_size, embedding_dim)
# 定义一个双向bi-lstm
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.
# 定义转移矩阵 transitions[i, j] 表示 从状态j转移到状态i
# 和正常表示想法
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
# 加上强制限制 1 没有状态能转移到 start tag 2 end stag 不能转移到任何状态
self.transitions.data[tag_to_ix[START_TAG], :] = -10000
self.transitions.data[:, tag_to_ix[STOP_TAG]] = -10000
# 初始化lstm 隐藏状态
self.hidden = self.init_hidden()
def init_hidden(self):
# 初始化lstm 隐藏状态参数
return (torch.randn(2, 1, self.hidden_dim // 2),
torch.randn(2, 1, self.hidden_dim // 2))
def _forward_alg(self, feats):
# Do the forward algorithm to compute the partition function
# 表示t0时刻到所有状态的分数score
init_alphas = torch.full((1, self.tagset_size), -10000.)
# START_TAG has all of the score.
# 表示t0时刻只能到start tag 状态
init_alphas[0][self.tag_to_ix[START_TAG]] = 0.
# Wrap in a variable so that we will get automatic backprop
# forward_var 存的是 t时刻 到每个状态的分数总和
# init_alphas
forward_var = init_alphas
# Iterate through the sentence
# 遍历 feats 也就是遍历 每个词 也就是 每个时刻
for feat in feats:
# 定义临时数组 存放当前时刻 到所有状态的分数总和
# 比如 alphas_t = [-1.3660, 1.6381, 0.2685， -9999.5635， -0.6746]
# alphas_t[i] 表示当前时刻 到 第i个状态的路分数总和
alphas_t = [] # The forward tensors at this timestep
for next_tag in range(self.tagset_size):
# 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)
# the ith entry of trans_score is the score of transitioning to
# next_tag from i
# 得到转移分数
trans_score = self.transitions[next_tag].view(1, -1)
# The ith entry of next_tag_var is the value for the
# edge (i -> next_tag) before we do log-sum-exp
# 对 trans_score + emit_score 分数进行累积
next_tag_var = forward_var + trans_score + emit_score
# The forward variable for this tag is log-sum-exp of all the
# scores.
# 对当前时刻转移到当前状态分数进行求和
alphas_t.append(log_sum_exp(next_tag_var).view(1))
forward_var = torch.cat(alphas_t).view(1, -1)
# 得到最后分数 添加 stop tag 分数
terminal_var = forward_var + self.transitions[self.tag_to_ix[STOP_TAG]]
# 对最后时刻所有状态分数求和
alpha = log_sum_exp(terminal_var)
return alpha
def _get_lstm_features(self, sentence):
# 得到 双向lstm的分数值
self.hidden = self.init_hidden()
embeds = self.word_embeds(sentence).view(len(sentence), 1, -1)
# (11, 1, 4)
lstm_out, self.hidden = self.lstm(embeds, self.hidden)
lstm_out = lstm_out.view(len(sentence), self.hidden_dim)
lstm_feats = self.hidden2tag(lstm_out)
return lstm_feats
def _score_sentence(self, feats, tags):
# Gives the score of a provided tag sequence
score = torch.zeros(1)
tags = torch.cat([torch.tensor([self.tag_to_ix[START_TAG]], dtype=torch.long), tags])
# 这里很好理解 得到real path 分数
for i, feat in enumerate(feats):
# 从这里也可以看出 self.transitions[i, j]表示从j 状态转移到 i状态
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
def _viterbi_decode(self, feats):
# 这里是用来进行 crf解码的 decoding
# backpointers 存储整个句子每个时刻到每个状态的最佳状态路径
backpointers = []
# Initialize the viterbi variables in log space
# 初始化 start tag 时刻到各个状态的最大分数值
init_vvars = torch.full((1, self.tagset_size), -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 存储了 t 时刻 到 各个状态的最大分数
forward_var = init_vvars
# 遍历sentence 也就是 所有时刻
for feat in feats:
bptrs_t = [] # holds the backpointers for this step
viterbivars_t = [] # holds the viterbi variables for this step
# 遍历所有状态
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)
# 得到当前时刻 到 该状态的转移分数值
next_tag_var = forward_var + self.transitions[next_tag]
# 选出到当前状态的最大的状态值
best_tag_id = argmax(next_tag_var)
# 记录下最大状态值
bptrs_t.append(best_tag_id)
# 记录最大状态的转移分数值
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
# 得到 emit score + transition score 并进行更新
forward_var = (torch.cat(viterbivars_t) + feat).view(1, -1)
# 记录当前时刻 到每个状态的最佳状态
# 比如 bptrs_t = [2, 2, 2, 2, 1] 中 bptrs_t[i] 表示 bptrs_t[i]状态 转移到 i 状态是分数最大的
# 最终 backpointers 存了整个句子每个时刻到每个状态的最佳状态路径
backpointers.append(bptrs_t)
# Transition to STOP_TAG
# 最后加上 转移到 stop_tag 分数
terminal_var = forward_var + self.transitions[self.tag_to_ix[STOP_TAG]]
# 从 后面到前面逆推 最佳路径
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]
# 从后到前得到最佳路径
for bptrs_t in reversed(backpointers):
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()
return path_score, best_path
def neg_log_likelihood(self, sentence, tags):
# 得到 lstm 分数 也就是 发射分数
feats = self._get_lstm_features(sentence)
# 计算所有路径总得分
forward_score = self._forward_alg(feats)
# 计算真实路径score
gold_score = self._score_sentence(feats, tags)
# 实现损失函数的定义
# 目标是极大化 p(y|x) = exp score(x, y) / Σexp score(x, yi)
# 因为我们计算的是对数概率 也就是 最大化 log p(y|x) = score(x, y) - Σexp score(x, yi)
# score(x, y) == gold_score 就是计算出来的真实路径
# forward_score 就是 Σexp score(x, yi)
# 所以最大化 gold_score - forward_score 等于 最小化 forward_score - gold_score
return forward_score - gold_score
def forward(self, sentence): # dont confuse this with _forward_alg above.
# 这个forward是用来解码的 就是计算出最大得分的路径
# Get the emission scores from the BiLSTM
# (11, 5)
lstm_feats = self._get_lstm_features(sentence)
# 通过 veterbi方法实现解码 得到分数最大的路径
# Find the best path, given the features.
score, tag_seq = self._viterbi_decode(lstm_feats)
return score, tag_seq本文章为转载内容，我们尊重原作者对文章享有的著作权。如有内容错误或侵权问题，欢迎原作者联系我们进行内容更正或删除文章。
