RNN() 内部含有 W, b 参数之类的

1
batch_size, num_steps = 32, 35
2
train_iter, vocab = d2l.load_data_time_machine(batch_size, num_steps)
3

4
num_hiddens = 256
5
# RNN() 内部含有 W, b 参数之类的
6
rnn_layer = nn.RNN(len(vocab), num_hiddens)
7

8
# state 形状 [隐藏层数，批量大小，隐藏单元数]
9
state = torch.zeros((1, batch_size, num_hiddens))
10
# X.shape = [时间步，批量，词典大小(独热)]
11
X = torch.rand(size=(num_steps, batch_size, len(vocab)))
12
# Y.shape = [时间步，批量，隐藏单元数], state 形状不变
13
Y, state = self.rnn(X, state)

1
class RNNModel(nn.Module):
2
    """循环神经网络模型"""
3
    def __init__(self, rnn_layer, vocab_size, **kwargs):
4
        super(RNNModel, self).__init__(**kwargs)
5
        self.rnn = rnn_layer
6
        self.vocab_size = vocab_size
7
        self.num_hiddens = self.rnn.hidden_size
8
        # 如果RNN是双向的（之后将介绍），num_directions应该是2，否则应该是1
9
        if not self.rnn.bidirectional:
10
            self.num_directions = 1
11
            self.linear = nn.Linear(self.num_hiddens, self.vocab_size)
12
        else:
13
            self.num_directions = 2
14
            self.linear = nn.Linear(self.num_hiddens * 2, self.vocab_size)
15

23 collapsed lines
16
    def forward(self, inputs, state):
17
        X = F.one_hot(inputs.T.long(), self.vocab_size)
18
        X = X.to(torch.float32)
19
        Y, state = self.rnn(X, state)
20
        # 全连接层首先将Y的形状改为(时间步数*批量大小,隐藏单元数)
21
        # 它的输出形状是(时间步数*批量大小,词表大小)。
22
        output = self.linear(Y.reshape((-1, Y.shape[-1])))
23
        return output, state
24

25
    def begin_state(self, device, batch_size=1):
26
        if not isinstance(self.rnn, nn.LSTM):
27
            # nn.GRU以张量作为隐状态
28
            return  torch.zeros((self.num_directions * self.rnn.num_layers,
29
                                 batch_size, self.num_hiddens),
30
                                device=device)
31
        else:
32
            # nn.LSTM以元组作为隐状态
33
            return (torch.zeros((
34
                self.num_directions * self.rnn.num_layers,
35
                batch_size, self.num_hiddens), device=device),
36
                    torch.zeros((
37
                        self.num_directions * self.rnn.num_layers,
38
                        batch_size, self.num_hiddens), device=device))