基于Pytorch 实现LSTM ,不调用nn.LSTM()

实现LSTM可以分为以下几个步骤：

1. 初始化权重和偏置

def init_weights(input_size, hidden_size):
    weight_ih = torch.randn(4 * hidden_size, input_size)
    weight_hh = torch.randn(4 * hidden_size, hidden_size)
    bias_ih = torch.zeros(4 * hidden_size)
    bias_hh = torch.zeros(4 * hidden_size)
    return weight_ih, weight_hh, bias_ih, bias_hh

2. 实现LSTM的前向传播

def forward(x, weight_ih, weight_hh, bias_ih, bias_hh, hidden, cell):
    # 计算输入门、遗忘门、输出门以及候选细胞状态
    gates = torch.mm(x, weight_ih.t()) + torch.mm(hidden, weight_hh.t()) + bias_ih + bias_hh
    input_gate, forget_gate, output_gate, candidate_cell_state = gates.chunk(4, 1)

    # 计算细胞状态和隐藏状态
    input_gate = torch.sigmoid(input_gate)
    forget_gate = torch.sigmoid(forget_gate)
    output_gate = torch.sigmoid(output_gate)
    candidate_cell_state = torch.tanh(candidate_cell_state)
    cell = forget_gate * cell + input_gate * candidate_cell_state
    hidden = output_gate * torch.tanh(cell)
    
    return hidden, cell

3. 启动LSTM的训练过程

# 定义超参数
input_size = 10
hidden_size = 20
seq_len = 5
batch_size = 3

# 初始化权重和偏置
weight_ih, weight_hh, bias_ih, bias_hh = init_weights(input_size, hidden_size)

# 初始化隐藏状态和细胞状态
hidden = torch.zeros(batch_size, hidden_size)
cell = torch.zeros(batch_size, hidden_size)

# 生成随机输入序列
inputs = torch.randn(seq_len, batch_size, input_size)

# 前向传播
for i in range(seq_len):
    hidden, cell = forward(inputs[i], weight_ih, weight_hh, bias_ih, bias_hh, hidden, cell)

# 输出最后一个隐藏状态
print(hidden[-1])

这样，我们就实现了一个基于Pytorch的LSTM模型。