机器翻译之Bahdanau注意力机制在Seq2Seq中的应用
目录
1.创建 添加了Bahdanau的decoder
2. 训练
3.定义评估函数BLEU
4.预测
5.知识点个人理解
1.创建 添加了Bahdanau的decoder
import torch
from torch import nn
import dltools#定义注意力解码器基类
class AttentionDecoder(dltools.Decoder): #继承dltools.Decoder写注意力编码器的基类def __init__(self, **kwargs):super().__init__(**kwargs)@property #装饰器， 定义的函数方法可以像类的属性一样被调用def attention_weights(self):#raise用于引发(或抛出)异常raise NotImplementedError #通常用于抽象基类中，作为占位符，提醒子类必须实现这个方法。 #创建 添加了Bahdanau的decoder
#继承AttentionDecoder这个基类创建Seq2SeqAttentionDecoder子类， 子类必须实现父类中NotImplementedError占位的方法
class Seq2SeqAttentionDecoder(AttentionDecoder): #初始化属性和方法def __init__(self, vocab_size, embed_size, num_hiddens, num_layers, dropout=0, **kwargs):"""vocab_size:此表大小, 相当于输入数据的特征数features, 也是输出数据的特征数embed_size：嵌入层的大小：将输入数据处理成小批量的数据num_hiddens：隐藏层神经元的数量num_layers：循环网络的层数dropout=0：不释放模型的参数（比如：神经元）"""super().__init__(**kwargs)#初始化注意力机制的评分函数方法self.attention = dltools.AdditiveAttention(key_size=num_hiddens,query_size=num_hiddens, num_hiddens=num_hiddens,dropout=dropout)#初始化嵌入层：将输入的数据处理成小批量的tensor数据 （文本--->数值的映射转化）self.embedding = nn.Embedding(num_embeddings=vocab_size, embedding_dim=embed_size)#初始化循环网络self.rnn = nn.GRU(embed_size+num_hiddens, num_hiddens, num_layers, dropout=dropout)#初始化线性层 （输出层）self.dense = nn.Linear(num_hiddens, vocab_size)#初始化隐藏层的状态state (计算state，需要编码器的输出结果、序列的有效长度)def init_state(self, enc_outputs, enc_valid_lens, *args):#enc_outputs是一个元组（输出结果，隐藏状态）#outputs的shape=（batch_size, num_steps, num_hiddens）#hidden_state的shape=（num_layers, batch_size, num_hiddens）outputs, hidden_state = enc_outputs#返回一个元组（，），可以用一个变量接收#outputs.permute(1, 0, 2)转换数据的维度是因为rnn循环神经网络的输入要求是先num_steps，再batch_size,return (outputs.permute(1, 0, 2), hidden_state, enc_valid_lens)#定义前向传播 (输入数据X，state)def forward(self, X, state):#变量赋值：接收编码器encoder的输出结果、隐藏状态、序列有效长度#enc_outputs的shape=（batch_size, num_steps, num_hiddens）#hidden_state的shape=（num_layers, batch_size, num_hiddens）enc_outputs, hidden_state, enc_valid_lens = state#X的shape=（batch_size, num_steps, vocab_size）X = self.embedding(X) #将X输入embedding嵌入层后， X的shape=（batch_size, num_steps, embed_size）#调换X的0维度和1维度数据X = X.permute(1, 0, 2) #X的shape=（num_steps, batch_size, embed_size）outputs, self._attention_weights = [], [] #创建空列表，用于存储数据for x in X: #遍历每一批数据#获取query#hidden_state[-1]表示最后一层循环网络的隐藏层状态 （有两层循环网络）#hidden_state[-1]的shape=（batch_size, num_hiddens） #dim=1表示在原索引1的维度增加一个维度query = torch.unsqueeze(hidden_state[-1], dim=1)
# print('query的shape:', query.shape) #query的shape=(batch_size, 1, num_hiddens)#通过注意力机制获取上下文序列context = self.attention(query, enc_outputs, enc_outputs, enc_valid_lens)
# print('context的shape:', context.shape) #context的shape=(batch_size, 1, num_hiddens)#用最后一个维度 拼接context, x 数据x = torch.cat((context, torch.unsqueeze(x, dim=1)), dim=-1)
# print('x的shape:', x.shape) #x的shape=(batch_size, 1, num_hiddens+embed_size)#将x和hidden_state输入循环神经网络中，获取输出结果和新的hidden_stateout, hidden_state = self.rnn(x.permute(1, 0, 2), hidden_state)
# print('out的shape:', out.shape) #out的shape=(1, batch_size, num_hiddens)
# print('hidden_state的shape:', hidden_state.shape) #两层循环层：hidden_state的shape=(2, batch_size, num_hiddens)#将输出结果添加到列表中outputs.append(out)self._attention_weights.append(self.attention_weights)outputs = self.dense(torch.cat(outputs, dim=0))
# print('outputs的shape:', outputs.shape) #outputs的shape=(num_steps, batch_size, vocab_size)return outputs.permute(1, 0, 2), [enc_outputs, hidden_state, enc_valid_lens]@propertydef attention_weights(self):return self._attention_weights#测试代码
#创建编码器对象
encoder = dltools.Seq2SeqEncoder(vocab_size=10, embed_size=8, num_hiddens=16, num_layers=2)
#需要预测， 要加encoder.eval()
encoder.eval()
#创建解码器对象
decoder = Seq2SeqAttentionDecoder(vocab_size=10, embed_size=8, num_hiddens=16, num_layers=2)
decoder.eval()#假设数据
batch_size, num_steps = 4, 7
X = torch.zeros((4, 7), dtype = torch.long)
#初始化状态state
state = decoder.init_state(encoder(X), None)
outputs, state = decoder(X, state)
#state包含三个东西（enc_outputs, hidden_state, enc_valid_lens）
#state[0]是 enc_outputs
#state[1]是 hidden_state， 两层循环层，就会有两个hidden_state， state[1][0]是第一层的hidden_state
outputs.shape, len(state), state[0].shape, len(state[1]), state[1][0].shapequery的shape: torch.Size([4, 1, 16]) context的shape: torch.Size([4, 1, 16]) x的shape: torch.Size([4, 1, 24]) out的shape: torch.Size([1, 4, 16]) hidden_state的shape: torch.Size([2, 4, 16]) query的shape: torch.Size([4, 1, 16]) context的shape: torch.Size([4, 1, 16]) x的shape: torch.Size([4, 1, 24]) out的shape: torch.Size([1, 4, 16]) hidden_state的shape: torch.Size([2, 4, 16]) query的shape: torch.Size([4, 1, 16]) context的shape: torch.Size([4, 1, 16]) x的shape: torch.Size([4, 1, 24]) out的shape: torch.Size([1, 4, 16]) hidden_state的shape: torch.Size([2, 4, 16]) query的shape: torch.Size([4, 1, 16]) context的shape: torch.Size([4, 1, 16]) x的shape: torch.Size([4, 1, 24]) out的shape: torch.Size([1, 4, 16]) hidden_state的shape: torch.Size([2, 4, 16]) query的shape: torch.Size([4, 1, 16]) context的shape: torch.Size([4, 1, 16]) x的shape: torch.Size([4, 1, 24]) out的shape: torch.Size([1, 4, 16]) hidden_state的shape: torch.Size([2, 4, 16]) query的shape: torch.Size([4, 1, 16]) context的shape: torch.Size([4, 1, 16]) x的shape: torch.Size([4, 1, 24]) out的shape: torch.Size([1, 4, 16]) hidden_state的shape: torch.Size([2, 4, 16]) query的shape: torch.Size([4, 1, 16]) context的shape: torch.Size([4, 1, 16]) x的shape: torch.Size([4, 1, 24]) out的shape: torch.Size([1, 4, 16]) hidden_state的shape: torch.Size([2, 4, 16]) outputs的shape: torch.Size([7, 4, 10])Out[11]:
(torch.Size([4, 7, 10]), 3, torch.Size([4, 7, 16]), 2, torch.Size([4, 16]))
2. 训练
#声明变量
embed_size, num_hiddens, num_layers, dropout = 32, 32, 2, 0.1
batch_size, num_steps = 64, 10
lr, num_epochs, device = 0.005, 200, dltools.try_gpu()#加载数据
train_iter, src_vocab, tgt_vocab = dltools.load_data_nmt(batch_size, num_steps)#创建编辑器对象
encoder = dltools.Seq2SeqEncoder(len(src_vocab), embed_size, num_hiddens, num_layers, dropout)
#创建编辑器对象
decoder = Seq2SeqAttentionDecoder(len(tgt_vocab), embed_size, num_hiddens, num_layers, dropout)#创建网络模型
net = dltools.EncoderDecoder(encoder, decoder)#模型训练
dltools.train_seq2seq(net, train_iter, lr, num_epochs, tgt_vocab, device) 
3.定义评估函数BLEU
def bleu(pred_seq, label_seq, k):print('pred_seq:', pred_seq)print('label_seq:', label_seq)#将pred_seq, label_seq分别进行空格分隔pred_tokens, label_tokens = pred_seq.split(' '), label_seq.split(' ')#获取pred_seq, label_seq的长度len_pred, len_label = len(pred_seq), len(label_seq)score = math.exp(min(0, 1 - (len_label / len_pred)))for n in range(1, k+1): #n的取值范围， range（）左闭右开num_matches, label_subs = 0, collections.defaultdict(int)for i in range(len_label - n + 1):label_subs[' '.join(label_tokens[i: i+n])] += 1for i in range(len_pred - n + 1):if label_subs[' '.join(pred_tokens[i: i+n])] > 0:num_matches += 1label_subs[' '.join(pred_tokens[i: i+n])] -=1score *= math.pow(num_matches / (len_pred -n + 1), math.pow(0.5, n))return score4.预测
import math
import collectionsengs = ['go .', 'i lost .', 'he\'s calm .', 'i\'m home .']
fras = ['va !', 'j\'ai perdu .', 'il est calme .', 'je suis chez moi .']
for eng, fra in zip(engs, fras):translation = dltools.predict_seq2seq(net, eng, src_vocab, tgt_vocab, num_steps, device)print(f'{eng} => {translation}, bleu {dltools.bleu(translation[0], fra, k=2):.3f}') go . => ('va !', []), bleu 1.000
i lost . => ("j'ai perdu .", []), bleu 1.000
he's calm . => ('il est bon .', []), bleu 0.658
i'm home . => ('je suis chez moi .', []), bleu 1.000
5.知识点个人理解
