注意
跳至结尾 下载完整的示例代码。
高级:动态决策和 Bi-LSTM CRF#
创建于:2017 年 4 月 8 日 | 最后更新:2021 年 12 月 20 日 | 最后验证:2024 年 11 月 5 日
动态与静态深度学习工具包#
Pytorch 是一个动态神经网络工具包。另一个动态工具包的例子是 Dynet(我提到它是因为使用 Pytorch 和 Dynet 是相似的。如果你看到 Dynet 的例子,它很可能有助于你在 Pytorch 中实现它)。相反的是静态工具包,包括 Theano、Keras、TensorFlow 等。核心区别如下:
在静态工具包中,您定义一次计算图,对其进行编译,然后将实例流式传输给它。
在动态工具包中,您为每个实例定义一个计算图。它从未被编译,并且是即时执行的。
如果没有很多经验,很难体会到这种区别。一个例子是假设我们要构建一个深度成分分析器。假设我们的模型涉及大致以下步骤:
我们自底向上构建树
标记根节点(句子的词语)
然后,使用神经网络和词语的嵌入来查找形成成分的组合。每当您形成一个新成分时,请使用某种技术来获取该成分的嵌入。在这种情况下,我们的网络架构将完全取决于输入句子。在句子“The green cat scratched the wall”中,模型中的某个时刻,我们将希望组合跨度 \((i,j,r) = (1, 3, \text{NP})\)(即,NP 成分跨越词语 1 到词语 3,在本例中为“The green cat”)。
但是,另一个句子可能是“Somewhere, the big fat cat scratched the wall”。在这个句子中,我们将希望在某个时候形成成分 \((2, 4, NP)\)。我们将要形成的成分将取决于实例。如果像静态工具包那样只编译一次计算图,那么编程这种逻辑将非常困难甚至不可能。但是在动态工具包中,不仅仅有一个预定义的计算图。每个实例都可以有一个新的计算图,因此这个问题就消失了。
动态工具包还具有易于调试的优点,并且代码更接近宿主语言(我的意思是 Pytorch 和 Dynet 比 Keras 或 Theano 更像实际的 Python 代码)。
Bi-LSTM 条件随机场讨论#
在本节中,我们将看到一个用于命名实体识别的 Bi-LSTM 条件随机场的完整、复杂的示例。上面的 LSTM 标记器通常足以进行词性标注,但像 CRF 这样的序列模型对于 NER 的强大性能至关重要。假定对 CRF 有所了解。虽然这个名字听起来很吓人,但整个模型就是一个 CRF,其中 LSTM 提供了特征。然而,这是一个高级模型,比本教程中任何早期模型都要复杂得多。如果您想跳过它,也没关系。要判断您是否准备好,请看您是否可以:
写出第 i 步、标签为 k 的 Viterbi 变量的递推关系。
修改上述递推关系以计算前向变量。
再次修改上述递推关系以在对数空间中计算前向变量(提示:log-sum-exp)
如果您能完成这三件事,您应该就能理解下面的代码。回想一下,CRF 计算条件概率。令 \(y\) 为标签序列,\(x\) 为单词输入序列。然后我们计算:
其中分数由定义一些对数势 \(\log \psi_i(x,y)\) 来确定,使得:
为了使归一化因子可处理,势能只能查看局部特征。
在 Bi-LSTM CRF 中,我们定义了两种势能:发射(emission)和转移(transition)。索引为 \(i\) 的词语的发射势能来自 Bi-LSTM 在时间步 \(i\) 的隐藏状态。转移分数存储在一个 \(|T|x|T|\) 矩阵 \(\textbf{P}\) 中,其中 \(T\) 是标签集。在我的实现中,\(\textbf{P}_{j,k}\) 是从标签 \(k\) 转移到标签 \(j\) 的分数。所以:
其中在第二个表达式中,我们将标签视为分配了唯一的非负索引。
如果上述讨论过于简略,您可以查看 Michael Collins 关于 CRF 的 这篇 文档。
实现说明#
下面的示例在对数空间中实现了前向算法来计算归一化因子,并使用 Viterbi 算法进行解码。反向传播会自动为我们计算梯度。我们不必手动处理任何事情。
该实现未经优化。如果您理解正在发生的事情,您可能会很快发现前向算法中迭代下一个标签的操作可以一次性完成。我希望代码更具可读性。如果您想进行相关更改,您或许可以将此标记器用于实际任务。
# Author: Robert Guthrie
import torch
import torch.autograd as autograd
import torch.nn as nn
import torch.optim as optim
torch.manual_seed(1)
<torch._C.Generator object at 0x7f49c71889b0>
用于提高代码可读性的辅助函数。
def argmax(vec):
# return the argmax as a python int
_, idx = torch.max(vec, 1)
return idx.item()
def prepare_sequence(seq, to_ix):
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):
max_score = vec[0, argmax(vec)]
max_score_broadcast = max_score.view(1, -1).expand(1, vec.size()[1])
return max_score + \
torch.log(torch.sum(torch.exp(vec - max_score_broadcast)))
创建模型
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.
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 (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
init_alphas = torch.full((1, self.tagset_size), -10000.)
# START_TAG has all of the score.
init_alphas[0][self.tag_to_ix[START_TAG]] = 0.
# Wrap in a variable so that we will get automatic backprop
forward_var = init_alphas
# Iterate through the sentence
for feat in feats:
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
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
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).view(1))
forward_var = torch.cat(alphas_t).view(1, -1)
terminal_var = forward_var + self.transitions[self.tag_to_ix[STOP_TAG]]
alpha = log_sum_exp(terminal_var)
return alpha
def _get_lstm_features(self, sentence):
self.hidden = self.init_hidden()
embeds = self.word_embeds(sentence).view(len(sentence), 1, -1)
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])
for i, feat in enumerate(feats):
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):
backpointers = []
# Initialize the viterbi variables in log space
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 = 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]
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)
backpointers.append(bptrs_t)
# Transition to STOP_TAG
terminal_var = forward_var + self.transitions[self.tag_to_ix[STOP_TAG]]
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):
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):
feats = self._get_lstm_features(sentence)
forward_score = self._forward_alg(feats)
gold_score = self._score_sentence(feats, tags)
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
运行训练
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
with torch.no_grad():
precheck_sent = prepare_sequence(training_data[0][0], word_to_ix)
precheck_tags = torch.tensor([tag_to_ix[t] for t in training_data[0][1]], dtype=torch.long)
print(model(precheck_sent))
# Make sure prepare_sequence from earlier in the LSTM section is loaded
for epoch in range(
300): # 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 Tensors of word indices.
sentence_in = prepare_sequence(sentence, word_to_ix)
targets = torch.tensor([tag_to_ix[t] for t in tags], dtype=torch.long)
# Step 3. Run our forward pass.
loss = model.neg_log_likelihood(sentence_in, targets)
# Step 4. Compute the loss, gradients, and update the parameters by
# calling optimizer.step()
loss.backward()
optimizer.step()
# Check predictions after training
with torch.no_grad():
precheck_sent = prepare_sequence(training_data[0][0], word_to_ix)
print(model(precheck_sent))
# We got it!
(tensor(2.6907), [1, 2, 2, 2, 2, 2, 2, 2, 2, 2, 1])
(tensor(20.4906), [0, 1, 1, 1, 2, 2, 2, 0, 1, 2, 2])
练习:用于区分性标注的新损失函数#
在进行解码时,我们实际上没有必要创建计算图,因为我们不从 Viterbi 路径分数进行反向传播。既然我们已经有了,请尝试在损失函数是 Viterbi 路径分数与黄金标准路径分数之差的情况下训练标记器。应该很清楚,这个函数是非负的,并且在预测的标签序列是正确的标签序列时为 0。这本质上是结构化感知器。
此修改应该很简短,因为 Viterbi 和 score_sentence 已经实现。这是计算图的形状取决于训练实例的一个例子。尽管我没有尝试在静态工具包中实现这一点,但我可以想象这是可能的,但要困难得多。
拾取一些真实数据并进行比较!
脚本总运行时间: (0 分钟 7.061 秒)