### Bi-LSTM Conditional Random Field
### pytorch tutorials https://pytorch.org/tutorials/beginner/nlp/advanced_tutorial.html
### 模型主要结构：
pytorch bilstm crf的教程，注意 这里不支持批处理
Python version: 3.7.4 (default, Aug 13 2019, 20:35:49) [GCC 7.3.0]
Torch version: 1.4.0
# Author: Robert Guthrie
import torch
import torch.autograd as autograd
import torch.nn as nn
import torch.optim as optim
torch.manual_seed(1)
def argmax(vec):
# return the argmax as a python int
# 返回vec的dim为1维度上的最大值索引
_, idx = torch.max(vec, 1)
return idx.item()
def prepare_sequence(seq, to_ix):
# 将句子转化为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
# 前向算法是不断累积之前的结果，这样就会有个缺点
# 指数和累积到一定程度后，会超过计算机浮点值的最大值，变成inf，这样取log后也是inf
# 为了避免这种情况，用一个合适的值clip去提指数和的公因子，这样就不会使某项变得过大而无法计算
# SUM = log(exp(s1)+exp(s2)+...+exp(s100))
# = log{exp(clip)*[exp(s1-clip)+exp(s2-clip)+...+exp(s100-clip)]}
# = clip + log[exp(s1-clip)+exp(s2-clip)+...+exp(s100-clip)]
# where clip=max
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)))
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 # word embedding dim
self.hidden_dim = hidden_dim # Bi-LSTM 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.
# 将BiLSTM提取的特征向量映射到特征空间，即经过全连接得到发射分数
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.
# 转移矩阵的参数初始化，transitions[i,j]代表的是从第j个tag转移到第i个tag的转移分数
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
# 初始化所有其他tag转移到START_TAG的分数非常小，即不可能由其他tag转移到START_TAG
# 初始化STOP_TAG转移到所有其他tag的分数非常小，即不可能由STOP_TAG转移到其他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):
# 初始化LSTM的参数
return (torch.randn(2, 1, self.hidden_dim // 2),
torch.randn(2, 1, self.hidden_dim // 2))
def _get_lstm_features(self, sentence):
# 通过Bi-LSTM提取特征
self.hidden = self.init_hidden()
# 因此，embeds 的最终数据形式是一个三维张量，形状为 (seq_len, 1, embed_dim)，其中：
# seq_len 是句子的长度（即单词的数量）。
# 1 表示批次大小，表明当前处理的是单个句子。
# embed_dim 是每个单词的嵌入向量维度。
# 这种形状非常适合直接传递给 PyTorch 的 LSTM 层进行处理，因为 LSTM 层期望输入有三个维度，分别对应序列长度
# 、批次大小和特征数（或输入大小）。如果你希望模型能够处理多个句子（即更大的批次），你应该相应地调整代码，
# 使得 sentence 可以同时包含多条序列，并且批次大小不固定为1。
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
# 计算给定tag序列的分数，即一条路径的分数
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):
# 递推计算路径分数：转移分数 + 发射分数
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 _forward_alg(self, feats):
# Do the forward algorithm to compute the partition function
# 通过前向算法递推计算
init_alphas = torch.full((1, self.tagset_size), -10000.)
# START_TAG has all of the score.
# 初始化step 0即START位置的发射分数，START_TAG取0其他位置取-10000
init_alphas[0][self.tag_to_ix[START_TAG]] = 0.
# Wrap in a variable so that we will get automatic backprop
# 将初始化START位置为0的发射分数赋值给previous
previous = init_alphas
# Iterate through the sentence
# 迭代整个句子
for obs in feats:
# The forward tensors at this timestep
# 当前时间步的前向tensor
alphas_t = []
for next_tag in range(self.tagset_size):
# broadcast the emission score: it is the same regardless of the previous tag
# 取出当前tag的发射分数，与之前时间步的tag无关
emit_score = obs[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
# 取出当前tag由之前tag转移过来的转移分数
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
# 当前路径的分数：之前时间步分数 + 转移分数 + 发射分数
next_tag_var = previous + trans_score + emit_score
# The forward variable for this tag is log-sum-exp of all the scores.
# 对当前分数取log-sum-exp
alphas_t.append(log_sum_exp(next_tag_var).view(1))
# 更新previous 递推计算下一个时间步
previous = torch.cat(alphas_t).view(1, -1)
# 考虑最终转移到STOP_TAG
terminal_var = previous + self.transitions[self.tag_to_ix[STOP_TAG]]
# 计算最终的分数
scores = log_sum_exp(terminal_var)
return scores
def _viterbi_decode(self, feats):
backpointers = []
# Initialize the viterbi variables in log space
# 初始化viterbi的previous变量
init_vvars = torch.full((1, self.tagset_size), -10000.)
init_vvars[0][self.tag_to_ix[START_TAG]] = 0
previous = init_vvars
for obs in feats:
# holds the backpointers for this step
# 保存当前时间步的回溯指针
bptrs_t = []
# holds the viterbi variables for this step
# 保存当前时间步的viterbi变量
viterbivars_t = []
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)
# 维特比算法记录最优路径时只考虑上一步的分数以及上一步tag转移到当前tag的转移分数
# 并不取决与当前tag的发射分数
next_tag_var = previous + 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
# 更新previous，加上当前tag的发射分数obs
previous = (torch.cat(viterbivars_t) + obs).view(1, -1)
# 回溯指针记录当前时间步各个tag来源前一步的tag
backpointers.append(bptrs_t)
# Transition to STOP_TAG
# 考虑转移到STOP_TAG的转移分数
terminal_var = previous + 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.
# 通过回溯指针解码出最优路径
best_path = [best_tag_id]
# best_tag_id作为线头，反向遍历backpointers找到最优路径
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)
# 去除START_TAG
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):
# CRF损失函数由两部分组成，真实路径的分数和所有路径的总分数。
# 真实路径的分数应该是所有路径中分数最高的。
# log真实路径的分数/log所有可能路径的分数，越大越好，构造crf loss函数取反，loss越小越好
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
# 通过BiLSTM提取发射分数
lstm_feats = self._get_lstm_features(sentence)
# Find the best path, given the features.
# 根据发射分数以及转移分数，通过viterbi解码找到一条最优路径
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
# 构造一些训练数据
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
# 训练前检查模型预测结果
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
# 第一步，pytorch梯度累积，需要清零梯度
model.zero_grad()
# Step 2. Get our inputs ready for the network, that is,
# turn them into Tensors of word indices.
# 第二步，将输入转化为tensors
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.
# 进行前向计算，取出crf loss
loss = model.neg_log_likelihood(sentence_in, targets)
# Step 4. Compute the loss, gradients, and update the parameters by
# calling optimizer.step()
# 第四步，计算loss，梯度，通过optimier更新参数
loss.backward()
optimizer.step()
# Check predictions after training
# 训练结束查看模型预测结果，对比观察模型是否学到
with torch.no_grad():
precheck_sent = prepare_sequence(training_data[0][0], word_to_ix)
print(model(precheck_sent))
# We got it!
改成批处理关键代码 previous_score = score[t - 1].view(batch_size, -1, 1)
def viterbi_decode(self, h: FloatTensor, mask: BoolTensor) -> List[List[int]]:
"""
decode labels using viterbi algorithm
:param h: hidden matrix (batch_size, seq_len, num_labels)
:param mask: mask tensor of each sequence
in mini batch (batch_size, batch_size)
:return: labels of each sequence in mini batch
"""
batch_size, seq_len, _ = h.size()
# prepare the sequence lengths in each sequence
seq_lens = mask.sum(dim=1)
# In mini batch, prepare the score
# from the start sequence to the first label
score = [self.start_trans.data + h[:, 0]]
path = []
for t in range(1, seq_len):
# extract the score of previous sequence
# (batch_size, num_labels, 1)
previous_score = score[t - 1].view(batch_size, -1, 1)
# extract the score of hidden matrix of sequence
# (batch_size, 1, num_labels)
h_t = h[:, t].view(batch_size, 1, -1)
# extract the score in transition
# from label of t-1 sequence to label of sequence of t
# self.trans_matrix has the score of the transition
# from sequence A to sequence B
# (batch_size, num_labels, num_labels)
score_t = previous_score + self.trans_matrix + h_t
# keep the maximum value
# and point where maximum value of each sequence
# (batch_size, num_labels)
best_score, best_path = score_t.max(1)
score.append(best_score)
path.append(best_path)
torchcrf 使用 支持批处理，torchcrf的简单使用-CSDN博客文章浏览阅读9.7k次，点赞5次，收藏33次。本文介绍了如何在PyTorch中安装和使用TorchCRF库，重点讲解了CRF模型参数设置、自定义掩码及损失函数的计算。作者探讨了如何将CRF的NLL损失与交叉熵结合，并通过自适应权重优化训练过程。虽然在单任务中效果不显著，但对于多任务学习提供了有价值的方法。
https://blog.csdn.net/csdndogo/article/details/125541213
torchcrf的简单使用-CSDN博客
为了防止文章丢失 ，吧内容转发在这里
https://blog.csdn.net/csdndogo/article/details/125541213
. 安装torchcrf，模型使用
安装：pip install TorchCRF
CRF的使用：在官网里有简单的使用说明
注意输入的格式。在其他地方下载的torchcrf有多个版本，有些版本有batch_first参数，有些没有，要看清楚有没有这个参数，默认batch_size是第一维度。
这个代码是我用来熟悉使用crf模型和损失函数用的，模拟多分类任务输入为随机数据和随机标签，所以最后的结果预测不能很好的跟标签对应。
import torch
import torch.nn as nn
import numpy as np
import random
from TorchCRF import CRF
from torch.optim import Adam
seed = 100
def seed_everything(seed=seed):
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
torch.backends.cudnn.benchmark = False
torch.backends.cudnn.deterministic = True
num_tags = 5
model = CRF(num_tags, batch_first=True) # 这里根据情况而定
seq_len = 3
batch_size = 50
seed_everything()
trainset = torch.randn(batch_size, seq_len, num_tags) # features
traintags = (torch.rand([batch_size, seq_len])*4).floor().long() # (batch_size, seq_len)
testset = torch.randn(5, seq_len, num_tags) # features
testtags = (torch.rand([5, seq_len])*4).floor().long() # (batch_size, seq_len)
# 训练阶段
for e in range(50):
optimizer = Adam(model.parameters(), lr=0.05)
model.train()
optimizer.zero_grad()
loss = -model(trainset, traintags)
print('epoch{}: loss score is {}'.format(e, loss))
loss.backward()
torch.nn.utils.clip_grad_norm_(model.parameters(),5)
optimizer.step()
#测试阶段
model.eval()
loss = model(testset, testtags)
model.decode(testset)
1.1模型参数，自定义掩码mask注意事项
def forward(self, emissions, labels: LongTensor, mask: BoolTensor)
1
分别为发射矩阵（各标签的预测值），标签，掩码（注意这里的mask类型为BoolTensor）
注意：此处自定义mask掩码时，使用LongTensor类型的[1,1,1,1,0,0]会报错，需要转换成ByteTensor，下面是一个简单的获取mask的函数，输入为标签数据：
def get_crfmask(self, labels):
crfmask = []
for batch in labels:
res = [0 if d == -1 else 1 for d in batch]
crfmask.append(res)
return torch.ByteTensor(crfmask)
运行运行
2. CRF的损失函数是什么？
损失函数由真实转移路径值和所有可能情况路径转移值两部分组成，损失函数的公式为
分子为真实转移路径值，分母为所有路径总分数，上图公式在crf原始代码中为：
def forward(
self, h: FloatTensor, labels: LongTensor, mask: BoolTensor) -> FloatTensor:
log_numerator = self._compute_numerator_log_likelihood(h, labels, mask)
log_denominator = self._compute_denominator_log_likelihood(h, mask)
return log_numerator - log_denominator
CRF损失函数值为负对数似然函数（NLL），所以如果原来的模型损失函数使用的是交叉熵损失函数，两个损失函数相加时要对CRF返回的损失取负。
loss = -model(trainset, traintags)
1
3. 如何联合CRF的损失函数和自己的网络模型的交叉熵损失函数进行训练？
我想在自己的模型上添加CRF，就需要联合原本的交叉熵损失函数和CRF的损失函数，因为CRF输出的时NLL，所以在模型在我仅对该损失函数取负之后和原先函数相加。
loss2 = -crf_layer(log_prob, label, mask=crfmask)
loss1 = loss_function(log_prob.permute(0, 2, 1), label)
loss = loss1 + loss2
loss.backward()
缺陷： 效果不佳，可以尝试对loss2添加权重。此处贴一段包含两个损失函数的自适应权重训练的函数。
3.1.自适应损失函数权重
由于CRF返回的损失与原来的损失数值不在一个量级，所以产生了自适应权重调整两个权重的大小来达到优化的目的。自适应权重原本属于多任务学习部分，未深入了解，代码源自某篇复现论文的博客。
class AutomaticWeightedLoss(nn.Module):
def __init__(self, num=2):
super(AutomaticWeightedLoss, self).__init__()
params = torch.ones(num, requires_grad=True)
self.params = torch.nn.Parameter(params)
def forward(self, *x):
loss_sum = 0
for i, loss in enumerate(x):
loss_sum += 0.5 / (self.params[i] ** 2) * loss + torch.log(1 + self.params[i] ** 2)
return loss_sum
