• 《动手学深度学习 Pytorch版》 10.4 Bahdanau注意力

    10.4.1 模型

    Bahdanau 等人提出了一个没有严格单向对齐限制的可微注意力模型。在预测词元时，如果不是所有输入词元都相关，模型将仅对齐（或参与）输入序列中与当前预测相关的部分。这是通过将上下文变量视为注意力集中的输出来实现的。

    新的基于注意力的模型与 9.7 节中的模型相同，只不过 9.7 节中的上下文变量 c \boldsymbol{c} c 在任何解码时间步 t ′ \boldsymbol{t'} t′ 都会被 c t ′ \boldsymbol{c}_{t'} ct′​ 替换。假设输入序列中有 T \boldsymbol{T} T 个词元，解码时间步 t ′ \boldsymbol{t'} t′ 的上下文变量是注意力集中的输出：

    c t ′ = ∑ t = 1 T α ( s t ′ − 1 , h t ) h t \boldsymbol{c}_{t'}=\sum^T_{t=1}{\alpha{(\boldsymbol{s}_{t'-1},\boldsymbol{h}_t)\boldsymbol{h}_t}} ct′​=t=1∑T​α(st′−1​,ht​)ht​

    参数字典：

    • 遵循与 9.7 节中的相同符号表达

    • 时间步 t ′ − 1 \boldsymbol{t'-1} t′−1 时的解码器隐状态 s t ′ − 1 \boldsymbol{s}_{t'-1} st′−1​ 是查询

    • 编码器隐状态 h t \boldsymbol{h}_t ht​ 既是键，也是值

    • 注意力权重 α \alpha α 是使用上节所定义的加性注意力打分函数计算的

    在这里插入图片描述

    从图中可以看到，加入注意力机制后：

    • 将编码器对每次词的输出作为 key 和 value

    • 将解码器对上一个词的输出作为 querry

    • 将注意力的输出和下一个词的词嵌入合并作为解码器输入

    import torch
from torch import nn
from d2l import torch as d2l

    • 1
    • 2
    • 3

    10.4.2 定义注意力解码器

    AttentionDecoder 类定义了带有注意力机制解码器的基本接口

    #@save
class AttentionDecoder(d2l.Decoder):
    """带有注意力机制解码器的基本接口"""
    def __init__(self, **kwargs):
        super(AttentionDecoder, self).__init__(**kwargs)

    @property
    def attention_weights(self):
        raise NotImplementedError

    • 1
    • 2
    • 3
    • 4
    • 5
    • 6
    • 7
    • 8
    • 9

    在 Seq2SeqAttentionDecoder 类中实现带有 Bahdanau 注意力的循环神经网络解码器。初始化解码器的状态，需要下面的输入：

    • 编码器在所有时间步的最终层隐状态，将作为注意力的键和值；

    • 上一时间步的编码器全层隐状态，将作为初始化解码器的隐状态；

    • 编码器有效长度（排除在注意力池中填充词元）。

    class Seq2SeqAttentionDecoder(AttentionDecoder):
    def __init__(self, vocab_size, embed_size, num_hiddens, num_layers,
                 dropout=0, **kwargs):
        super(Seq2SeqAttentionDecoder, self).__init__(**kwargs)
        self.attention = d2l.AdditiveAttention(
            num_hiddens, num_hiddens, num_hiddens, dropout)
        self.embedding = nn.Embedding(vocab_size, embed_size)
        self.rnn = nn.GRU(
            embed_size + num_hiddens, num_hiddens, num_layers,
            dropout=dropout)
        self.dense = nn.Linear(num_hiddens, vocab_size)

    def init_state(self, enc_outputs, enc_valid_lens, *args):  # 新增 enc_valid_lens 表示有效长度
        # outputs的形状为(batch_size，num_steps，num_hiddens).
        # hidden_state的形状为(num_layers，batch_size，num_hiddens)
        outputs, hidden_state = enc_outputs
        return (outputs.permute(1, 0, 2), hidden_state, enc_valid_lens)

    def forward(self, X, state):
        # enc_outputs的形状为(batch_size,num_steps,num_hiddens).
        # hidden_state的形状为(num_layers,batch_size,num_hiddens)
        enc_outputs, hidden_state, enc_valid_lens = state
        # 输出X的形状为(num_steps,batch_size,embed_size)
        X = self.embedding(X).permute(1, 0, 2)
        outputs, self._attention_weights = [], []
        for x in X:
            # query的形状为(batch_size,1,num_hiddens)
            query = torch.unsqueeze(hidden_state[-1], dim=1)  # 解码器最终隐藏层的上一个输出添加querry个数的维度后作为querry
            # context的形状为(batch_size,1,num_hiddens)
            context = self.attention(
                query, enc_outputs, enc_outputs, enc_valid_lens)  # 编码器的输出作为key和value
            # 在特征维度上连结
            x = torch.cat((context, torch.unsqueeze(x, dim=1)), dim=-1)  # 并起来当解码器输入
            # 将x变形为(1,batch_size,embed_size+num_hiddens)
            out, hidden_state = self.rnn(x.permute(1, 0, 2), hidden_state)
            outputs.append(out)
            self._attention_weights.append(self.attention.attention_weights)  # 存一下注意力权重
        # 全连接层变换后，outputs的形状为 (num_steps,batch_size,vocab_size)
        outputs = self.dense(torch.cat(outputs, dim=0))
        return outputs.permute(1, 0, 2), [enc_outputs, hidden_state,
                                          enc_valid_lens]

    @property
    def attention_weights(self):
        return self._attention_weights

    • 1
    • 2
    • 3
    • 4
    • 5
    • 6
    • 7
    • 8
    • 9
    • 10
    • 11
    • 12
    • 13
    • 14
    • 15
    • 16
    • 17
    • 18
    • 19
    • 20
    • 21
    • 22
    • 23
    • 24
    • 25
    • 26
    • 27
    • 28
    • 29
    • 30
    • 31
    • 32
    • 33
    • 34
    • 35
    • 36
    • 37
    • 38
    • 39
    • 40
    • 41
    • 42
    • 43
    • 44
    • 45
    encoder = d2l.Seq2SeqEncoder(vocab_size=10, embed_size=8, num_hiddens=16,
                             num_layers=2)
encoder.eval()
decoder = Seq2SeqAttentionDecoder(vocab_size=10, embed_size=8, num_hiddens=16,
                                  num_layers=2)
decoder.eval()
X = torch.zeros((4, 7), dtype=torch.long)  # (batch_size,num_steps)
state = decoder.init_state(encoder(X), None)
output, state = decoder(X, state)
output.shape, len(state), state[0].shape, len(state[1]), state[1][0].shape

    • 1
    • 2
    • 3
    • 4
    • 5
    • 6
    • 7
    • 8
    • 9
    • 10
    (torch.Size([4, 7, 10]), 3, torch.Size([4, 7, 16]), 2, torch.Size([4, 16]))

    • 1

    10.4.3 训练

    embed_size, num_hiddens, num_layers, dropout = 32, 32, 2, 0.1
batch_size, num_steps = 64, 10
lr, num_epochs, device = 0.005, 250, d2l.try_gpu()

train_iter, src_vocab, tgt_vocab = d2l.load_data_nmt(batch_size, num_steps)
encoder = d2l.Seq2SeqEncoder(
    len(src_vocab), embed_size, num_hiddens, num_layers, dropout)
decoder = Seq2SeqAttentionDecoder(
    len(tgt_vocab), embed_size, num_hiddens, num_layers, dropout)
net = d2l.EncoderDecoder(encoder, decoder)
d2l.train_seq2seq(net, train_iter, lr, num_epochs, tgt_vocab, device)

    • 1
    • 2
    • 3
    • 4
    • 5
    • 6
    • 7
    • 8
    • 9
    • 10
    • 11
    loss 0.020, 7252.9 tokens/sec on cuda:0

    • 1

    在这里插入图片描述

    engs = ['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, dec_attention_weight_seq = d2l.predict_seq2seq(
        net, eng, src_vocab, tgt_vocab, num_steps, device, True)
    print(f'{eng} => {translation}, ',
          f'bleu {d2l.bleu(translation, fra, k=2):.3f}')

    • 1
    • 2
    • 3
    • 4
    • 5
    • 6
    • 7
    go . => va !,  bleu 1.000
i lost . => j'ai perdu .,  bleu 1.000
he's calm . => il est mouillé .,  bleu 0.658
i'm home . => je suis chez moi .,  bleu 1.000

    • 1
    • 2
    • 3
    • 4

    训练结束后，下面通过可视化注意力权重会发现，每个查询都会在键值对上分配不同的权重，这说明在每个解码步中，输入序列的不同部分被选择性地聚集在注意力池中。

    attention_weights = torch.cat([step[0][0][0] for step in dec_attention_weight_seq], 0).reshape((
    1, 1, -1, num_steps))

# 加上一个包含序列结束词元
d2l.show_heatmaps(
    attention_weights[:, :, :, :len(engs[-1].split()) + 1].cpu(),
    xlabel='Key positions', ylabel='Query positions')

    • 1
    • 2
    • 3
    • 4
    • 5
    • 6
    • 7

    ​
    在这里插入图片描述

    ​

    练习

    （1）在实验中用LSTM替换GRU。

    class Seq2SeqEncoder_LSTM(d2l.Encoder):
    def __init__(self, vocab_size, embed_size, num_hiddens, num_layers,
                 dropout=0, **kwargs):
        super(Seq2SeqEncoder_LSTM, self).__init__(**kwargs)
        self.embedding = nn.Embedding(vocab_size, embed_size)
        self.lstm = nn.LSTM(embed_size, num_hiddens, num_layers,  # 更换为 LSTM
                          dropout=dropout)

    def forward(self, X, *args):
        X = self.embedding(X)
        X = X.permute(1, 0, 2)
        output, state = self.lstm(X)
        return output, state

class Seq2SeqAttentionDecoder_LSTM(AttentionDecoder):
    def __init__(self, vocab_size, embed_size, num_hiddens, num_layers,
                 dropout=0, **kwargs):
        super(Seq2SeqAttentionDecoder_LSTM, self).__init__(**kwargs)
        self.attention = d2l.AdditiveAttention(
            num_hiddens, num_hiddens, num_hiddens, dropout)
        self.embedding = nn.Embedding(vocab_size, embed_size)
        self.rnn = nn.LSTM(
            embed_size + num_hiddens, num_hiddens, num_layers,
            dropout=dropout)
        self.dense = nn.Linear(num_hiddens, vocab_size)

    def init_state(self, enc_outputs, enc_valid_lens, *args):
        outputs, hidden_state = enc_outputs
        return (outputs.permute(1, 0, 2), hidden_state, enc_valid_lens)

    def forward(self, X, state):
        enc_outputs, hidden_state, enc_valid_lens = state
        X = self.embedding(X).permute(1, 0, 2)
        outputs, self._attention_weights = [], []
        for x in X:
            query = torch.unsqueeze(hidden_state[-1][0], dim=1)  # 解码器最终隐藏层的上一个输出添加querry个数的维度后作为querry
            context = self.attention(
                query, enc_outputs, enc_outputs, enc_valid_lens)
            x = torch.cat((context, torch.unsqueeze(x, dim=1)), dim=-1)
            out, hidden_state = self.rnn(x.permute(1, 0, 2), hidden_state)
            outputs.append(out)
            self._attention_weights.append(self.attention.attention_weights)
        outputs = self.dense(torch.cat(outputs, dim=0))
        return outputs.permute(1, 0, 2), [enc_outputs, hidden_state,
                                          enc_valid_lens]

    @property
    def attention_weights(self):
        return self._attention_weights

    • 1
    • 2
    • 3
    • 4
    • 5
    • 6
    • 7
    • 8
    • 9
    • 10
    • 11
    • 12
    • 13
    • 14
    • 15
    • 16
    • 17
    • 18
    • 19
    • 20
    • 21
    • 22
    • 23
    • 24
    • 25
    • 26
    • 27
    • 28
    • 29
    • 30
    • 31
    • 32
    • 33
    • 34
    • 35
    • 36
    • 37
    • 38
    • 39
    • 40
    • 41
    • 42
    • 43
    • 44
    • 45
    • 46
    • 47
    • 48
    • 49
    embed_size_LSTM, num_hiddens_LSTM, num_layers_LSTM, dropout_LSTM = 32, 32, 2, 0.1
batch_size_LSTM, num_steps_LSTM = 64, 10
lr_LSTM, num_epochs_LSTM, device_LSTM = 0.005, 250, d2l.try_gpu()

train_iter_LSTM, src_vocab_LSTM, tgt_vocab_LSTM = d2l.load_data_nmt(batch_size_LSTM, num_steps_LSTM)
encoder_LSTM = Seq2SeqEncoder_LSTM(
    len(src_vocab_LSTM), embed_size_LSTM, num_hiddens_LSTM, num_layers_LSTM, dropout_LSTM)
decoder_LSTM = Seq2SeqAttentionDecoder_LSTM(
    len(tgt_vocab_LSTM), embed_size_LSTM, num_hiddens_LSTM, num_layers_LSTM, dropout_LSTM)
net_LSTM = d2l.EncoderDecoder(encoder_LSTM, decoder_LSTM)
d2l.train_seq2seq(net_LSTM, train_iter_LSTM, lr_LSTM, num_epochs_LSTM, tgt_vocab_LSTM, device_LSTM)

    • 1
    • 2
    • 3
    • 4
    • 5
    • 6
    • 7
    • 8
    • 9
    • 10
    • 11
    loss 0.021, 7280.8 tokens/sec on cuda:0

    • 1

    在这里插入图片描述

    engs = ['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, dec_attention_weight_seq_LSTM = d2l.predict_seq2seq(
        net_LSTM, eng, src_vocab_LSTM, tgt_vocab_LSTM, num_steps_LSTM, device_LSTM, True)
    print(f'{eng} => {translation}, ',
          f'bleu {d2l.bleu(translation, fra, k=2):.3f}')

    • 1
    • 2
    • 3
    • 4
    • 5
    • 6
    • 7
    go . => va !,  bleu 1.000
i lost . => j'ai perdu .,  bleu 1.000
he's calm . => puis-je   .,  bleu 0.000
i'm home . => je suis chez moi .,  bleu 1.000

    • 1
    • 2
    • 3
    • 4
    attention_weights_LSTM = torch.cat([step[0][0][0] for step in dec_attention_weight_seq_LSTM], 0).reshape((
    1, 1, -1, num_steps_LSTM))

# 加上一个包含序列结束词元
d2l.show_heatmaps(
    attention_weights_LSTM[:, :, :, :len(engs[-1].split()) + 1].cpu(),
    xlabel='Key positions', ylabel='Query positions')

    • 1
    • 2
    • 3
    • 4
    • 5
    • 6
    • 7

    ​
    在这里插入图片描述

    ​

    （2）修改实验以将加性注意力打分函数替换为缩放点积注意力，它如何影响训练效率？

    class Seq2SeqAttentionDecoder_Dot(AttentionDecoder):
    def __init__(self, vocab_size, embed_size, num_hiddens, num_layers,
                 dropout=0, **kwargs):
        super(Seq2SeqAttentionDecoder, self).__init__(**kwargs)
        self.attention = d2l.DotProductAttention(  # 替换为缩放点积注意力
            num_hiddens, num_hiddens, num_hiddens, dropout)
        self.embedding = nn.Embedding(vocab_size, embed_size)
        self.rnn = nn.GRU(
            embed_size + num_hiddens, num_hiddens, num_layers,
            dropout=dropout)
        self.dense = nn.Linear(num_hiddens, vocab_size)

    def init_state(self, enc_outputs, enc_valid_lens, *args):
        outputs, hidden_state = enc_outputs
        return (outputs.permute(1, 0, 2), hidden_state, enc_valid_lens)

    def forward(self, X, state):
        enc_outputs, hidden_state, enc_valid_lens = state
        X = self.embedding(X).permute(1, 0, 2)
        outputs, self._attention_weights = [], []
        for x in X:
            query = torch.unsqueeze(hidden_state[-1], dim=1)
            context = self.attention(
                query, enc_outputs, enc_outputs, enc_valid_lens)
            x = torch.cat((context, torch.unsqueeze(x, dim=1)), dim=-1)
            out, hidden_state = self.rnn(x.permute(1, 0, 2), hidden_state)
            outputs.append(out)
            self._attention_weights.append(self.attention.attention_weights)
        outputs = self.dense(torch.cat(outputs, dim=0))
        return outputs.permute(1, 0, 2), [enc_outputs, hidden_state,
                                          enc_valid_lens]

    @property
    def attention_weights(self):
        return self._attention_weights

    • 1
    • 2
    • 3
    • 4
    • 5
    • 6
    • 7
    • 8
    • 9
    • 10
    • 11
    • 12
    • 13
    • 14
    • 15
    • 16
    • 17
    • 18
    • 19
    • 20
    • 21
    • 22
    • 23
    • 24
    • 25
    • 26
    • 27
    • 28
    • 29
    • 30
    • 31
    • 32
    • 33
    • 34
    • 35
    embed_size_Dot, num_hiddens_Dot, num_layers_Dot, dropout_Dot = 32, 32, 2, 0.1
batch_size_Dot, num_steps_Dot = 64, 10
lr_Dot, num_epochs_Dot, device_Dot = 0.005, 250, d2l.try_gpu()

train_iter_Dot, src_vocab_Dot, tgt_vocab_Dot = d2l.load_data_nmt(batch_size_Dot, num_steps_Dot)
encoder_Dot = Seq2SeqEncoder_LSTM(
    len(src_vocab_Dot), embed_size_LSTM, num_hiddens_Dot, num_layers_Dot, dropout_Dot)
decoder_Dot = Seq2SeqAttentionDecoder_LSTM(
    len(tgt_vocab_Dot), embed_size_Dot, num_hiddens_Dot, num_layers_Dot, dropout_Dot)
net_Dot = d2l.EncoderDecoder(encoder_Dot, decoder_Dot)
d2l.train_seq2seq(net_Dot, train_iter_Dot, lr_Dot, num_epochs_Dot, tgt_vocab_Dot, device_Dot)

    • 1
    • 2
    • 3
    • 4
    • 5
    • 6
    • 7
    • 8
    • 9
    • 10
    • 11
    loss 0.021, 7038.8 tokens/sec on cuda:0

    • 1

    在这里插入图片描述

    engs = ['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, dec_attention_weight_seq_Dot = d2l.predict_seq2seq(
        net_Dot, eng, src_vocab_Dot, tgt_vocab_Dot, num_steps_Dot, device_Dot, True)
    print(f'{eng} => {translation}, ',
          f'bleu {d2l.bleu(translation, fra, k=2):.3f}')

    • 1
    • 2
    • 3
    • 4
    • 5
    • 6
    • 7
    go . => va !,  bleu 1.000
i lost . => j'ai perdu .,  bleu 1.000
he's calm . => il est riche .,  bleu 0.658
i'm home . => je suis chez moi .,  bleu 1.000

    • 1
    • 2
    • 3
    • 4
    attention_weights_Dot = torch.cat([step[0][0][0] for step in dec_attention_weight_seq_Dot], 0).reshape((
    1, 1, -1, num_steps_Dot))

# 加上一个包含序列结束词元
d2l.show_heatmaps(
    attention_weights_Dot[:, :, :, :len(engs[-1].split()) + 1].cpu(),
    xlabel='Key positions', ylabel='Query positions')

    • 1
    • 2
    • 3
    • 4
    • 5
    • 6
    • 7

    ​
    在这里插入图片描述

    ​

  • 相关阅读:
    C++ 虚函数表
    现在做跨境电商还需要全球代理IP吗？全球代理IP哪家靠谱？
    9、IOC 之基于注释的容器配置
    ArmSoM-W3之RK3588 Debian11详解
    【C++】list容器
    Microsoft Edge不能工作了，可能原因不少，那么如何修复呢
    【Web_环境搭建_Python3_pip】pip的升级、安装、更新、卸载，以及pipupgrade和pip-review的基础使用
    TCP、UDP、ARP协议（持续更新）
    神经网络和深度神经网络,深度神经网络类型包括
    Codeforces Round 892 (Div. 2) - E. Maximum Monogonosity 思维dp 详细解析
  • 原文地址：https://blog.csdn.net/qq_43941037/article/details/134043136