训练bilstm + crf模型时，发射矩阵不变

如何解决训练bilstm + crf模型时，发射矩阵不变

我想构建BiLSTM + CRF模型来解决NER问题。建立基于pytorch的模型。 但是，每个Epoche的发射矩阵都没有改变。

我修改了pytorch的官方模型代码。在bilstm层中，我压缩了每个批次并将其输入到lstm层中，但是输出发射矩阵保持不变。我不知道为什么BiLSTM层不起作用。 请帮助我，谢谢。

这是我的代码。

import utils


#1、先用embedding测试一下
'''random.seed(1)
np.random.seed(1)
torch.manual_seed(1)'''


class BILSTM_CRF(nn.Module):
    def __init__(self,batch_size,vocab_size,tag_to_idx,embed_dim,hidden_dim,word_to_idx):
        super(BILSTM_CRF,self).__init__()
        self.batch_size = batch_size
        self.embed_dim = embed_dim
        self.hidden_dim = hidden_dim
        self.tag_to_idx = tag_to_idx
        self.vocab_size = vocab_size
        self.target_size = len(tag_to_idx) # 要带着start和end
        self.word_to_idx = word_to_idx

        self.embedding = nn.Embedding(vocab_size,embed_dim)

        self.lstm = nn.LSTM(embed_dim,hidden_dim // 2,bidirectional=True,batch_first=True)

        # 发射函数
        self.emission = nn.Linear(hidden_dim,self.target_size)

        # crf 部分，参数的初始化
        # 注意：t[i][j]是指由tag j转移到tag i
        self.transitions = nn.Parameter(torch.randn(self.target_size,self.target_size,requires_grad=True))

        # 其中start不会由任何tag转移
        # end不会转移任何到tag，所以更新这两个的转移值变得很小
        self.transitions.data[tag_to_idx["<start>"],:] = -10000
        self.transitions.data[:,tag_to_idx["<end>"]] = -10000

    # 通过bilstm层获得发射函数
    def _get_lstm_features(self,sents):
        h0 = torch.randn(2,self.batch_size,self.hidden_dim // 2)
        c0 = torch.randn(2,self.hidden_dim // 2)

        lengths = utils.get_lengths(sents) # 导入的序列长度

        # 下面是修改了官方代码，将整个batch压缩后放进lstm层计算
        input_tensor = utils.pad_sents(sents,self.word_to_idx) # 填充与编码sents
        # print("input",input_tensor)

        emb = self.embedding(input_tensor)
        # print(emb.shape)
        # 压缩
        input_packed = torch.nn.utils.rnn.pack_padded_sequence(emb,lengths,batch_first=True)
        output_tensor,_ = self.lstm(input_packed)
        # 解压缩
        output_tensor,_ = torch.nn.utils.rnn.pad_packed_sequence(output_tensor,batch_first=True)
        emission = self.emission(output_tensor) # size[batch,length,tag_size]
        return emission


    def neg_log_likelihood(self,sents,tags): # loss函数
        # sents[batch,length] tags[batch,length]
        feats = self._get_lstm_features(sents) # size[batch,tag_size]

        # ！！！！！！For each epoche,the output here is the same
        print("feats",feats)
 
        forward_score = self._forward_alg(feats) # size[batch,1]
        # print("forward_score",forward_score.shape)
        gold_score = self._score_sentence(feats,tags) # size[batch,1]
        # print("gold_score",gold_score.shape)

        losses = forward_score - gold_score # size[batch,1]
        loss = torch.sum(losses)
        return loss,feats

utils是一个包含一些工具功能的软件包 显示在下面

# 一些小函数
import torch

# 统计序列长度
def get_lengths(sents):
    lengths = []
    for sent in sents:
        lengths.append(len(sent))
    return lengths

# 填充序列等长
# 放进来的sents是排好序的
# pad是填充内容
# default是不在字典中的默认内容，<unk>
def pad_sents(sents,word_map):
    UNK = word_map.get('<unk>')
    PAD = word_map.get('<pad>')
    max_len = len(sents[0])
    # batch_tensor = torch.ones(len(sents),max_len,dtype=torch.long,requires_grad=True)*pad
    batch_tensor = torch.ones(len(sents),requires_grad=True).long()*PAD

    for i,l in enumerate(sents):
        for j,e in enumerate(l):
            batch_tensor[i][j] = word_map.get(e,UNK)

    return batch_tensor

# 句子按照长短排序用
def sort_pairs(sents_list,tag_list):
    pairs = list(zip(sents_list,tag_list))
    # print(pairs)
    indices = sorted(range(len(pairs)),key = lambda k:len(pairs[k][0]),reverse = True)
    pairs = [pairs[i] for i in indices]

    sents_list,tag_list = list(zip(*pairs))

    return sents_list,tag_list