import torch import torch.nn as nn from transformers.models.gpt2.modeling_gpt2 import GPT2Model from einops import rearrange from model.AEPSA.normalizer import GumbelSoftmax from model.AEPSA.reprogramming import ReprogrammingLayer import torch.nn.functional as F # 该文件实现了基于动态图增强的时空序列预测模型 # 主要包含三个类：DynamicGraphEnhancer（动态图增强器）、GraphEnhancedEncoder（图增强编码器）和AEPSA（主模型） # 每个操作都标注了输入输出shape以帮助理解数据流向 class DynamicGraphEnhancer(nn.Module): """ 动态图增强器，基于节点嵌入自动生成图结构 参考DDGCRN的设计，使用节点嵌入和特征信息动态计算邻接矩阵 """ def __init__(self, num_nodes, in_dim, embed_dim=10): # num_nodes: 节点数量 # in_dim: 输入特征维度 # embed_dim: 节点嵌入维度 super().__init__() self.num_nodes = num_nodes self.embed_dim = embed_dim # 节点嵌入参数 [num_nodes, embed_dim] self.node_embeddings = nn.Parameter( torch.randn(num_nodes, embed_dim), requires_grad=True ) # 特征转换层，用于生成动态调整的嵌入 # 输入: [N, in_dim] -> 输出: [N, embed_dim] self.feature_transform = nn.Sequential( nn.Linear(in_dim, 16), # [N, in_dim] -> [N, 16] nn.Sigmoid(), nn.Linear(16, 2), # [N, 16] -> [N, 2] nn.Sigmoid(), nn.Linear(2, embed_dim) # [N, 2] -> [N, embed_dim] ) # 注册单位矩阵作为固定的支持矩阵 [num_nodes, num_nodes] self.register_buffer("eye", torch.eye(num_nodes)) def get_laplacian(self, graph, I, normalize=True): """ 计算归一化拉普拉斯矩阵 参数: graph: 邻接矩阵 [N, N] I: 单位矩阵 [N, N] normalize: 是否使用标准化拉普拉斯矩阵 返回: laplacian: 归一化拉普拉斯矩阵 [N, N] """ # 计算度矩阵的逆平方根 [N, N] D_inv = torch.diag_embed(torch.sum(graph, -1) ** (-0.5)) # [N, N] D_inv[torch.isinf(D_inv)] = 0.0 # 处理零除问题 if normalize: # 归一化拉普拉斯矩阵: D^(-1/2) * graph * D^(-1/2) [N, N] return torch.matmul(torch.matmul(D_inv, graph), D_inv) # [N, N] else: # 拉普拉斯矩阵: D^(-1/2) * (graph + I) * D^(-1/2) [N, N] return torch.matmul(torch.matmul(D_inv, graph + I), D_inv) # [N, N] def forward(self, X): """ 参数: X: 输入特征 [B, N, D]，其中B为批次大小，N为节点数，D为特征维度 返回: laplacians: 动态生成的归一化拉普拉斯矩阵 [B, N, N] """ batch_size = X.size(0) laplacians = [] # 获取单位矩阵 [N, N] I = self.eye.to(X.device) for b in range(batch_size): # 使用特征转换层生成动态嵌入调整因子 [N, embed_dim] filt = self.feature_transform(X[b]) # [N, embed_dim] # 计算节点嵌入向量 [N, embed_dim] nodevec = torch.tanh(self.node_embeddings * filt) # [N, embed_dim] # 通过节点嵌入的点积计算邻接矩阵 [N, N] adj = F.relu(torch.matmul(nodevec, nodevec.transpose(0, 1))) # [N, N] # 计算归一化拉普拉斯矩阵 [N, N] laplacian = self.get_laplacian(adj, I) # [N, N] laplacians.append(laplacian) # 堆叠所有批次的拉普拉斯矩阵 [B, N, N] return torch.stack(laplacians, dim=0) # [B, N, N] class GraphEnhancedEncoder(nn.Module): """ 基于Chebyshev多项式和动态拉普拉斯矩阵的图增强编码器 用于将动态图结构信息整合到特征编码中 优化版本：支持直接处理原始时间序列输入 """ def __init__(self, K=3, in_dim=64, hidden_dim=32, num_nodes=325, embed_dim=10, device='cpu', temporal_dim=12, num_features=1): # K: Chebyshev多项式阶数 # in_dim: 输入特征维度 # hidden_dim: 隐藏层维度 # num_nodes: 节点数量 # embed_dim: 嵌入维度 # temporal_dim: 时间序列长度 # num_features: 特征通道数量 super().__init__() self.K = K # Chebyshev多项式阶数 self.in_dim = in_dim self.hidden_dim = hidden_dim self.device = device self.temporal_dim = temporal_dim self.num_features = num_features # 输入预处理层，适配正确的通道维度 # 输入: [B, C, N, T] -> 输出: [B, in_dim, N, 1] self.input_projection = nn.Sequential( # 2D卷积：[B, C, N, T] -> [B, 16, N, T] nn.Conv2d(num_features, 16, kernel_size=(1, 3), padding=(0, 1)), # [B, C, N, T] -> [B, 16, N, T] nn.ReLU(), # 2D卷积：[B, 16, N, T] -> [B, in_dim, N, 1]，时间维度上的全局卷积 nn.Conv2d(16, in_dim, kernel_size=(1, temporal_dim)), # [B, 16, N, T] -> [B, in_dim, N, 1] nn.ReLU() ) # 动态图增强器，用于生成动态拉普拉斯矩阵 # 输入: [B, N, in_dim] -> 输出: [B, N, N] self.graph_enhancer = DynamicGraphEnhancer(num_nodes, in_dim, embed_dim) # 谱系数 α_k (可学习参数) [K+1, 1] self.alpha = nn.Parameter(torch.randn(K + 1, 1)) # 传播权重 W_k (可学习参数) # 每个权重将Chebyshev多项式展开的结果从in_dim映射到hidden_dim # 输入: [N, in_dim] -> 输出: [N, hidden_dim] self.W = nn.ParameterList([ nn.Parameter(torch.randn(in_dim, hidden_dim)) for _ in range(K + 1) ]) self.to(device) def chebyshev_polynomials(self, L_tilde, X): """ 递归计算Chebyshev多项式展开 [T_0(L_tilde)X, ..., T_K(L_tilde)X] 参数: L_tilde: 归一化拉普拉斯矩阵 [N, N] X: 输入特征 [N, in_dim] 返回: T_k_list: Chebyshev多项式展开列表 [K+1, N, in_dim] """ # T_0(X) = X [N, in_dim] T_k_list = [X] if self.K >= 1: # T_1(X) = L_tilde * X [N, in_dim] T_k_list.append(torch.matmul(L_tilde, X)) for k in range(2, self.K + 1): # T_k(X) = 2*L_tilde*T_{k-1}(X) - T_{k-2}(X) [N, in_dim] T_k_list.append(2 * torch.matmul(L_tilde, T_k_list[-1]) - T_k_list[-2]) # 返回Chebyshev多项式展开列表 [K+1, N, in_dim] return T_k_list def forward(self, X): """ 参数: X: 输入特征 [B, N, C, T] 或 [B, N, T]（单特征情况） B: 批次大小, N: 节点数, C: 特征通道数, T: 时间序列长度 返回: 增强后的特征 [B, N, hidden_dim*(K+1)] """ batch_size = X.size(0) num_nodes = X.size(1) # 处理不同维度的输入 if len(X.shape) == 4: # [B, N, C, T] # 输入: [B, N, C, T] -> 输出: [B, C, N, T] # 将输入转换为[B, C, N, T]格式，适合2D卷积 x = X.permute(0, 2, 1, 3) # [B, C, N, T] else: # [B, N, T] # 输入: [B, N, T] -> 输出: [B, 1, N, T] # 添加通道维度 x = X.unsqueeze(1) # [B, 1, N, T] # 使用卷积投影层处理时间维度 # 输入: [B, C, N, T] -> 输出: [B, in_dim, N, 1] x_proj = self.input_projection(x) # 输入: [B, in_dim, N, 1] -> 输出: [B, in_dim, N] x_proj = x_proj.squeeze(-1) # [B, in_dim, N] # 输入: [B, in_dim, N] -> 输出: [B, N, in_dim] x_proj = x_proj.permute(0, 2, 1) # [B, N, in_dim] enhanced_features = [] # 动态生成拉普拉斯矩阵 # 输入: [B, N, in_dim] -> 输出: [B, N, N] laplacians = self.graph_enhancer(x_proj) # [B, N, N] for b in range(batch_size): # 获取当前批次的拉普拉斯矩阵 [N, N] L = laplacians[b] # [N, N] # 特征值缩放 try: # 计算最大特征值 [1] lambda_max = torch.linalg.eigvalsh(L).max().real # [1] # 避免除零问题 if lambda_max < 1e-6: lambda_max = 1.0 # 缩放拉普拉斯矩阵到[-1, 1]区间 [N, N] L_tilde = (2.0 / lambda_max) * L - torch.eye(L.size(0), device=L.device) # [N, N] except: # 如果计算特征值失败，使用单位矩阵 [N, N] L_tilde = torch.eye(num_nodes, device=X.device) # [N, N] # 计算Chebyshev多项式展开 # 输入: L_tilde [N, N], x_proj [N, in_dim] -> 输出: [K+1, N, in_dim] T_k_list = self.chebyshev_polynomials(L_tilde, x_proj[b]) # [K+1, N, in_dim] H_list = [] # 应用传播权重 for k in range(self.K + 1): # 矩阵乘法: [N, in_dim] × [in_dim, hidden_dim] -> [N, hidden_dim] H_k = torch.matmul(T_k_list[k], self.W[k]) # [N, hidden_dim] H_list.append(H_k) # 拼接所有尺度的特征 # 输入: [K+1, N, hidden_dim] -> 输出: [N, hidden_dim*(K+1)] X_enhanced = torch.cat(H_list, dim=-1) # [N, hidden_dim*(K+1)] enhanced_features.append(X_enhanced) # 堆叠所有批次的增强特征 # 输入: [B, N, hidden_dim*(K+1)] -> 输出: [B, N, hidden_dim*(K+1)] return torch.stack(enhanced_features, dim=0) # [B, N, hidden_dim*(K+1)] class AEPSA(nn.Module): """ 自适应特征投影时空自注意力模型（AEPSA） 整合动态图增强和预训练语言模型进行时空序列预测 """ def __init__(self, configs): # configs: 包含模型所有配置的字典 # 主要配置参数说明： # device: 运行设备 # pred_len: 预测序列长度 # seq_len: 输入序列长度 # patch_len: 补丁长度（已移除对应组件） # input_dim: 输入特征维度 # stride: 步长（已移除对应组件） # dropout: Dropout概率 # gpt_layers: 使用的GPT2层数 # d_ff: 前馈网络隐藏层维度 # gpt_path: 预训练GPT2模型路径 # num_nodes: 节点数量 # word_num: GumbelSoftmax词汇数量 # d_model: 模型维度 # n_heads: 注意力头数量 # chebyshev_order: Chebyshev多项式阶数 # graph_hidden_dim: 图编码器隐藏层维度 # graph_embed_dim: 图编码器嵌入维度 super(AEPSA, self).__init__() self.device = configs['device'] self.pred_len = configs['pred_len'] self.seq_len = configs['seq_len'] self.patch_len = configs['patch_len'] self.input_dim = configs['input_dim'] self.stride = configs['stride'] self.dropout = configs['dropout'] self.gpt_layers = configs['gpt_layers'] self.d_ff = configs['d_ff'] self.gpt_path = configs['gpt_path'] self.num_nodes = configs.get('num_nodes', 325) # 添加节点数量配置 # GumbelSoftmax层，用于词汇选择 # 输入: [vocab_size] -> 输出: [vocab_size]（one-hot近似分布） self.word_choice = GumbelSoftmax(configs['word_num']) self.d_model = configs['d_model'] self.n_heads = configs['n_heads'] self.d_keys = None self.d_llm = 768 # GPT2隐藏层维度 self.patch_nums = int((self.seq_len - self.patch_len) / self.stride + 2) self.head_nf = self.d_ff * self.patch_nums # 移除不再使用的patch_embedding层 # GPT2初始化 # 加载预训练GPT2模型，输出注意力权重和隐藏状态 self.gpts = GPT2Model.from_pretrained(self.gpt_path, output_attentions=True, output_hidden_states=True) self.gpts.h = self.gpts.h[:self.gpt_layers] # 截取指定层数 self.gpts.apply(self.reset_parameters) # 获取GPT2词嵌入权重 # 形状: [vocab_size, d_llm] self.word_embeddings = self.gpts.get_input_embeddings().weight.to(self.device) self.vocab_size = self.word_embeddings.shape[0] # 映射层，将词汇表维度映射到1维 # 输入: [vocab_size] -> 输出: [1] self.mapping_layer = nn.Linear(self.vocab_size, 1) # 重编程层，用于特征映射和注意力计算 # 输入: [B, N, d_model], [d_llm], [d_llm] -> 输出: [B, N, d_model] self.reprogramming_layer = ReprogrammingLayer(self.d_model, self.n_heads, self.d_keys, self.d_llm) # 动态图增强编码器 # 输入: [B, N, C, T] -> 输出: [B, N, hidden_dim*(K+1)] self.graph_encoder = GraphEnhancedEncoder( K=configs.get('chebyshev_order', 3), # Chebyshev多项式阶数 in_dim=self.d_model, # 输入特征维度 hidden_dim=configs.get('graph_hidden_dim', 32), # 隐藏层维度 num_nodes=self.num_nodes, # 节点数量 embed_dim=configs.get('graph_embed_dim', 10), # 节点嵌入维度 device=self.device, # 运行设备 temporal_dim=self.seq_len, # 时间序列长度 num_features=self.input_dim # 特征通道数 ) # 图特征投影层，将图增强特征维度转换为d_model # 输入: [B, N, hidden_dim*(K+1)] -> 输出: [B, N, d_model] self.graph_projection = nn.Linear( configs.get('graph_hidden_dim', 32) * (configs.get('chebyshev_order', 3) + 1), self.d_model ) self.out_mlp = nn.Sequential( nn.Linear(self.d_llm, 128), nn.ReLU(), nn.Linear(128, self.pred_len) ) for i, (name, param) in enumerate(self.gpts.named_parameters()): if 'wpe' in name: param.requires_grad = True else: param.requires_grad = False def reset_parameters(self, module): if hasattr(module, 'weight') and module.weight is not None: torch.nn.init.normal_(module.weight, mean=0.0, std=0.02) if hasattr(module, 'bias') and module.bias is not None: torch.nn.init.zeros_(module.bias) def forward(self, x): """ 前向传播函数 输入: x: 输入数据 [B, T, N, C]，其中B为批次大小，T为时间步长，N为节点数，C为特征通道数 返回: outputs: 预测结果 [B, pred_len, N, 1] """ # 只保留第一个特征通道 # 输入: [B, T, N, C] -> 输出: [B, T, N, 1] x = x[..., :1] # [B, T, N, 1] # 调整输入维度以适配图编码器 # 输入: [B, T, N, 1] -> 输出: [B, N, 1, T] x_enc = rearrange(x, 'b t n c -> b n c t') # [B, N, 1, T] # 应用图增强编码器获取增强特征 # 输入: [B, N, 1, T] -> 输出: [B, N, hidden_dim*(K+1)] graph_enhanced = self.graph_encoder(x_enc) # [B, N, hidden_dim*(K+1)] # 投影图增强特征到模型维度 # 输入: [B, N, hidden_dim*(K+1)] -> 输出: [B, N, d_model] enc_out = self.graph_projection(graph_enhanced) # [B, N, d_model] # 处理词嵌入权重，为注意力机制准备 # 输入: [vocab_size, d_llm] -> 输出: [d_llm, vocab_size] -> [d_llm, vocab_size] self.mapping_layer(self.word_embeddings.permute(1, 0)).permute(1, 0) # [vocab_size, d_llm] # 使用GumbelSoftmax选择词汇 # 输入: [d_llm, 1] -> 输出: [d_llm, 1] masks = self.word_choice(self.mapping_layer.weight.data.permute(1,0)) # [d_llm, 1] # 获取选中的源嵌入 # 输入: [vocab_size, d_llm] 与 masks -> 输出: [selected_words, d_llm] source_embeddings = self.word_embeddings[masks==1] # [selected_words, d_llm] # 应用重编程层处理特征和源嵌入 # 输入: [B, N, d_model], [selected_words, d_llm], [selected_words, d_llm] -> 输出: [B, N, d_model] enc_out = self.reprogramming_layer(enc_out, source_embeddings, source_embeddings) # [B, N, d_model] # 通过GPT2模型处理增强特征 # 输入: [B, N, d_model] -> 输出: [B, N, d_llm] enc_out = self.gpts(inputs_embeds=enc_out).last_hidden_state # [B, N, d_llm] # 使用MLP预测未来时间步 # 输入: [B, N, d_llm] -> 输出: [B, N, pred_len] dec_out = self.out_mlp(enc_out) # [B, N, pred_len] # 添加通道维度 # 输入: [B, N, pred_len] -> 输出: [B, N, pred_len, 1] outputs = dec_out.unsqueeze(dim=-1) # [B, N, pred_len, 1] # 调整维度顺序为 [B, pred_len, N, 1] # 输入: [B, N, pred_len, 1] -> 输出: [B, pred_len, N, 1] outputs = outputs.permute(0, 2, 1, 3) # [B, pred_len, N, 1] return outputs # [B, pred_len, N, 1]