informer模型代码

以下是一个简单的informer模型的PyTorch代码实现：

import torch
import torch.nn as nn
import torch.nn.functional as F

class Informer(nn.Module):
    def __init__(self, enc_in, dec_in, out_seq_len, n_heads, e_layers, d_layers, d_ff, dropout, activation='relu'):
        super(Informer, self).__init__()
        
        self.encoder = Encoder(enc_in, n_heads, e_layers, d_ff, dropout, activation)
        self.decoder = Decoder(dec_in, out_seq_len, n_heads, d_layers, d_ff, dropout, activation)
        self.out = nn.Linear(dec_in, out_seq_len)
        
    def forward(self, x):
        enc_out, attn = self.encoder(x)
        dec_out = self.decoder(enc_out, attn)
        out = self.out(dec_out)
        return out

class Encoder(nn.Module):
    def __init__(self, input_dim, n_heads, n_layers, d_ff, dropout, activation):
        super(Encoder, self).__init__()

        self.layers = nn.ModuleList()
        for i in range(n_layers):
            self.layers.append(EncoderLayer(input_dim, n_heads, d_ff, dropout, activation))

    def forward(self, x):
        attn_weights = []
        for layer in self.layers:
            x, attn_weight = layer(x)
            attn_weights.append(attn_weight)
        return x, attn_weights

class EncoderLayer(nn.Module):
    def __init__(self, input_dim, n_heads, d_ff, dropout, activation):
        super(EncoderLayer, self).__init__()

        self.self_attn = MultiHeadAttention(n_heads, input_dim, input_dim, dropout)
        self.feed_forward = FeedForward(input_dim, d_ff, activation, dropout)
        self.norm1 = nn.LayerNorm(input_dim)
        self.norm2 = nn.LayerNorm(input_dim)
        self.dropout1 = nn.Dropout(dropout)
        self.dropout2 = nn.Dropout(dropout)

    def forward(self, x):
        # self-attention
        residual = x
        x, attn_weight = self.self_attn(x, x, x)
        x = self.norm1(residual + self.dropout1(x))

        # feed forward
        residual = x
        x = self.feed_forward(x)
        x = self.norm2(residual + self.dropout2(x))

        return x, attn_weight

class Decoder(nn.Module):
    def __init__(self, input_dim, out_seq_len, n_heads, n_layers, d_ff, dropout, activation):
        super(Decoder, self).__init__()

        self.layers = nn.ModuleList()
        for i in range(n_layers):
            self.layers.append(DecoderLayer(input_dim, n_heads, d_ff, dropout, activation))

        self.out_seq_len = out_seq_len
        self.linear = nn.Linear(input_dim, out_seq_len)

    def forward(self, enc_out, attn_weights):
        # mask future positions
        mask = torch.triu(torch.ones(self.out_seq_len, self.out_seq_len), diagonal=1)
        mask = mask.unsqueeze(0).bool().to(enc_out.device)

        # self-attention
        x = torch.zeros(enc_out.shape[0], self.out_seq_len, enc_out.shape[-1]).to(enc_out.device)
        for i in range(self.out_seq_len):
            residual = x[:, i, :]
            x[:, i, :], attn_weight = self.layers[i](x[:, :i+1, :], enc_out, mask, attn_weights)
            x[:, i, :] = residual + x[:, i, :]

        # linear
        out = self.linear(x)
        return out

class DecoderLayer(nn.Module):
    def __init__(self, input_dim, n_heads, d_ff, dropout, activation):
        super(DecoderLayer, self).__init__()

        self.self_attn = MultiHeadAttention(n_heads, input_dim, input_dim, dropout)
        self.enc_attn = MultiHeadAttention(n_heads, input_dim, input_dim, dropout)
        self.feed_forward = FeedForward(input_dim, d_ff, activation, dropout)
        self.norm1 = nn.LayerNorm(input_dim)
        self.norm2 = nn.LayerNorm(input_dim)
        self.norm3 = nn.LayerNorm(input_dim)
        self.dropout1 = nn.Dropout(dropout)
        self.dropout2 = nn.Dropout(dropout)
        self.dropout3 = nn.Dropout(dropout)

    def forward(self, x, enc_out, mask, attn_weights):
        # self-attention
        residual = x[:, -1, :]
        x[:, -1, :], attn_weight1 = self.self_attn(x[:, -1:, :], x[:, -1:, :], x[:, -1:, :], mask)
        x[:, -1, :] = residual + self.dropout1(x[:, -1, :])

        # encoder-decoder attention
        residual = x[:, -1, :]
        x[:, -1, :], attn_weight2 = self.enc_attn(x[:, -1:, :], enc_out, enc_out)
        x[:, -1, :] = residual + self.dropout2(x[:, -1, :])

        # feed forward
        residual = x[:, -1, :]
        x[:, -1, :] = self.feed_forward(x[:, -1, :])
        x[:, -1, :] = residual + self.dropout3(x[:, -1, :])

        attn_weights.append(torch.cat([attn_weight1, attn_weight2], dim=1))
        return x, attn_weights

class MultiHeadAttention(nn.Module):
    def __init__(self, n_heads, q_dim, k_dim, dropout):
        super(MultiHeadAttention, self).__init__()

        self.n_heads = n_heads
        self.q_dim = q_dim
        self.k_dim = k_dim

        self.query = nn.Linear(q_dim, q_dim * n_heads)
        self.key = nn.Linear(k_dim, k_dim * n_heads)
        self.value = nn.Linear(k_dim, k_dim * n_heads)
        self.out = nn.Linear(k_dim * n_heads, q_dim)
        self.dropout = nn.Dropout(dropout)

    def forward(self, query, key, value, mask=None):
        batch_size = query.shape[0]

        # linear
        query = self.query(query).view(batch_size, -1, self.n_heads, self.q_dim // self.n_heads).transpose(1, 2)
        key = self.key(key).view(batch_size, -1, self.n_heads, self.k_dim // self.n_heads).transpose(1, 2)
        value = self.value(value).view(batch_size, -1, self.n_heads, self.k_dim // self.n_heads).transpose(1, 2)

        # dot product attention
        attn_weight = torch.matmul(query, key.transpose(-2, -1)) / torch.sqrt(torch.tensor(self.k_dim // self.n_heads).float().to(query.device))
        if mask is not None:
            attn_weight = attn_weight.masked_fill(mask == False, -1e9)
        attn_weight = F.softmax(attn_weight, dim=-1)
        attn_weight = self.dropout(attn_weight)

        # linear
        output = torch.matmul(attn_weight, value).transpose(1, 2).contiguous().view(batch_size, -1, self.q_dim)
        output = self.out(output)

        return output, attn_weight

class FeedForward(nn.Module):
    def __init__(self, input_dim, hidden_dim, activation, dropout):
        super(FeedForward, self).__init__()

        self.linear1 = nn.Linear(input_dim, hidden_dim)
        self.linear2 = nn.Linear(hidden_dim, input_dim)
        self.activation = getattr(F, activation)
        self.dropout = nn.Dropout(dropout)

    def forward(self, x):
        x = self.linear1(x)
        x = self.activation(x)
        x = self.dropout(x)
        x = self.linear2(x)
        return x

这里实现了一个简单的Informer模型，包括Encoder、Decoder和MultiHeadAttention等模块。你可以根据具体的任务和数据来调整模型的参数和结构，以获得更好的性能。