ASTRA-V3

.vscode/launch.json

@@ -234,6 +234,14 @@
             "console": "integratedTerminal",
             "args": "--config ./config/AEPSA/v2_SolarEnergy.yaml"
         },
+        {
+            "name": "AEPSA_v3: METR-LA",
+            "type": "debugpy",
+            "request": "launch",
+            "program": "run.py",
+            "console": "integratedTerminal",
+            "args": "--config ./config/AEPSA/v3_METR-LA.yaml"
+        },
         {
             "name": "EXPB: NYCBike-InFlow",
             "type": "debugpy",

config/AEPSA/v3_METR-LA.yaml

@@ -0,0 +1,57 @@
+basic:
+  dataset: METR-LA
+  device: cuda:0
+  mode: train
+  model: AEPSA_v3
+  seed: 2023
+
+data:
+  batch_size: 16
+  column_wise: false
+  days_per_week: 7
+  horizon: 24
+  input_dim: 1
+  lag: 24
+  normalizer: std
+  num_nodes: 207
+  steps_per_day: 288
+  test_ratio: 0.2
+  val_ratio: 0.2
+
+model:
+  chebyshev_order: 3
+  d_ff: 128
+  d_model: 64
+  dropout: 0.2
+  graph_hidden_dim: 32
+  gpt_layers: 9
+  gpt_path: ./GPT-2
+  input_dim: 1
+  n_heads: 1
+  num_nodes: 207
+  patch_len: 6
+  pred_len: 24
+  seq_len: 24
+  stride: 7
+  word_num: 1000
+
+train:
+  batch_size: 16
+  debug: false
+  early_stop: true
+  early_stop_patience: 15
+  epochs: 100
+  grad_norm: false
+  log_step: 1000
+  loss_func: mae
+  lr_decay: true
+  lr_decay_rate: 0.3
+  lr_decay_step: 5,20,40,70
+  lr_init: 0.003
+  mae_thresh: None
+  mape_thresh: 0.001
+  max_grad_norm: 5
+  output_dim: 1
+  plot: false
+  real_value: true
+  weight_decay: 0

config/AEPSA/v3_PEMS-BAY.yaml

@@ -0,0 +1,54 @@
+basic:
+  dataset: PEMS-BAY
+  device: cuda:0
+  mode: train
+  model: AEPSA_v3
+  seed: 2023
+
+data:
+  batch_size: 16
+  column_wise: false
+  days_per_week: 7
+  horizon: 24
+  input_dim: 1
+  lag: 24
+  normalizer: std
+  num_nodes: 325
+  steps_per_day: 288
+  test_ratio: 0.2
+  val_ratio: 0.2
+
+model:
+  d_ff: 128
+  d_model: 64
+  dropout: 0.2
+  gpt_layers: 9
+  gpt_path: ./GPT-2
+  input_dim: 1
+  n_heads: 1
+  num_nodes: 325
+  patch_len: 6
+  pred_len: 24
+  seq_len: 24
+  stride: 7
+  word_num: 1000
+
+train:
+  batch_size: 16
+  debug: false
+  early_stop: true
+  early_stop_patience: 15
+  epochs: 100
+  grad_norm: false
+  log_step: 100
+  loss_func: mae
+  lr_decay: true
+  lr_decay_rate: 0.3
+  lr_decay_step: 5,20,40,70
+  lr_init: 0.003
+  mae_thresh: None
+  mape_thresh: 0.001
+  max_grad_norm: 5
+  output_dim: 1
+  plot: false
+  weight_decay: 0

model/AEPSA/aepsav3.py

@@ -0,0 +1,209 @@
+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
+
+# 基于动态图增强的时空序列预测模型实现
+
+class DynamicGraphEnhancer(nn.Module):
+    """动态图增强编码器"""
+    def __init__(self, num_nodes, in_dim, embed_dim=10):
+        super().__init__()
+        self.num_nodes = num_nodes  # 节点个数
+        self.embed_dim = embed_dim  # 节点嵌入维度
+        
+        self.node_embeddings = nn.Parameter(torch.randn(num_nodes, embed_dim), requires_grad=True)  # 节点嵌入参数
+        
+        self.feature_transform = nn.Sequential(  # 特征转换网络
+            nn.Linear(in_dim, 16),
+            nn.Sigmoid(),
+            nn.Linear(16, 2),
+            nn.Sigmoid(),
+            nn.Linear(2, embed_dim)
+        )
+        
+        self.register_buffer("eye", torch.eye(num_nodes))  # 注册单位矩阵
+    
+    def get_laplacian(self, graph, I, normalize=True):
+        D_inv = torch.diag_embed(torch.sum(graph, -1) ** (-0.5))  # 度矩阵的逆平方根
+        D_inv[torch.isinf(D_inv)] = 0.0  # 处理零除问题
+        if normalize:
+            return torch.matmul(torch.matmul(D_inv, graph), D_inv)  # 归一化拉普拉斯矩阵
+        else:
+            return torch.matmul(torch.matmul(D_inv, graph + I), D_inv)  # 带自环的归一化拉普拉斯矩阵
+    
+    def forward(self, X):
+        """生成动态拉普拉斯矩阵"""
+        batch_size = X.size(0)  # 批次大小
+        laplacians = []  # 存储各批次的拉普拉斯矩阵
+        I = self.eye.to(X.device)  # 移动单位矩阵到目标设备
+        
+        for b in range(batch_size):
+            filt = self.feature_transform(X[b])  # 特征转换
+            nodevec = torch.tanh(self.node_embeddings * filt)  # 计算节点嵌入
+            adj = F.relu(torch.matmul(nodevec, nodevec.transpose(0, 1)))  # 计算邻接矩阵
+            laplacian = self.get_laplacian(adj, I)  # 计算拉普拉斯矩阵
+            laplacians.append(laplacian)
+        return torch.stack(laplacians, dim=0)  # 堆叠并返回
+
+class GraphEnhancedEncoder(nn.Module):
+    """图增强编码器"""
+    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):
+        super().__init__()
+        self.K = K  # Chebyshev多项式阶数
+        self.in_dim = in_dim  # 输入特征维度
+        self.hidden_dim = hidden_dim  # 隐藏层维度
+        self.device = device  # 运行设备
+        self.temporal_dim = temporal_dim  # 时间序列长度
+        self.num_features = num_features  # 特征通道数量
+        
+        self.input_projection = nn.Sequential(  # 输入投影层
+            nn.Conv2d(num_features, 16, kernel_size=(1, 3), padding=(0, 1)),
+            nn.ReLU(),
+            nn.Conv2d(16, in_dim, kernel_size=(1, temporal_dim)),
+            nn.ReLU()
+        )
+        
+        self.graph_enhancer = DynamicGraphEnhancer(num_nodes, in_dim, embed_dim)  # 动态图增强器
+        self.alpha = nn.Parameter(torch.randn(K + 1, 1))  # 谱系数
+        self.W = nn.ParameterList([nn.Parameter(torch.randn(in_dim, hidden_dim)) for _ in range(K + 1)])  # 传播权重
+        self.to(device)  # 移动到指定设备
+
+    def chebyshev_polynomials(self, L_tilde, X):
+        """计算Chebyshev多项式展开"""
+        T_k_list = [X]  # T_0(X) = X
+        if self.K >= 1:
+            T_k_list.append(torch.matmul(L_tilde, X))  # T_1(X) = L_tilde * X
+        for k in range(2, self.K + 1):
+            T_k_list.append(2 * torch.matmul(L_tilde, T_k_list[-1]) - T_k_list[-2])  # 递推计算
+        return T_k_list  # 返回多项式列表
+
+    def forward(self, X):
+        """输入特征[B,N,C,T]，返回增强特征[B,N,hidden_dim*(K+1)]"""
+        batch_size = X.size(0)  # 批次大小
+        num_nodes = X.size(1)  # 节点数量
+
+        x = X.permute(0, 2, 1, 3)  # [B,C,N,T]
+        x_proj = self.input_projection(x).squeeze(-1)  # [B,in_dim,N]
+        x_proj = x_proj.permute(0, 2, 1)  # [B,N,in_dim]
+        
+        enhanced_features = []  # 存储增强特征
+        laplacians = self.graph_enhancer(x_proj)  # 生成动态拉普拉斯矩阵
+        
+        for b in range(batch_size):
+            L = laplacians[b]  # 当前批次的拉普拉斯矩阵
+            
+            # 特征值缩放
+            try:
+                lambda_max = torch.linalg.eigvalsh(L).max().real  # 最大特征值
+                lambda_max = 1.0 if lambda_max < 1e-6 else lambda_max  # 防止除零
+                L_tilde = (2.0 / lambda_max) * L - torch.eye(L.size(0), device=L.device)  # 归一化拉普拉斯
+            except:
+                L_tilde = torch.eye(num_nodes, device=X.device)  # 异常处理
+            
+            # 计算展开并应用权重
+            T_k_list = self.chebyshev_polynomials(L_tilde, x_proj[b])  # 计算Chebyshev多项式
+            H_list = [torch.matmul(T_k_list[k], self.W[k]) for k in range(self.K + 1)]  # 应用权重
+            X_enhanced = torch.cat(H_list, dim=-1)  # 拼接特征
+            enhanced_features.append(X_enhanced)
+        
+        return torch.stack(enhanced_features, dim=0)  # 堆叠返回[B,N,hidden_dim*(K+1)]，每个节点在每个k阶下的切比雪夫特征
+
+class AEPSA(nn.Module):
+    """自适应特征投影时空自注意力模型"""
+    def __init__(self, configs):
+        super(AEPSA, self).__init__()
+        self.device = configs['device']  # 运行设备
+        self.pred_len = configs['pred_len']  # 预测序列长度
+        self.seq_len = configs['seq_len']  # 输入序列长度
+        self.patch_len = configs['patch_len']  # 补丁长度
+        self.input_dim = configs['input_dim']  # 输入特征维度
+        self.stride = configs['stride']  # 步长
+        self.dropout = configs['dropout']  # Dropout概率
+        self.gpt_layers = configs['gpt_layers']  # 使用的GPT2层数
+        self.d_ff = configs['d_ff']  # 前馈网络隐藏层维度
+        self.gpt_path = configs['gpt_path']  # 预训练GPT2模型路径
+        self.num_nodes = configs.get('num_nodes', 325)  # 节点数量
+
+        self.word_choice = GumbelSoftmax(configs['word_num'])  # 词汇选择层
+
+        self.d_model = configs['d_model']  # 模型维度
+        self.n_heads = configs['n_heads']  # 注意力头数量
+        self.d_keys = None  # 键维度
+        self.d_llm = 768  # GPT2隐藏层维度
+
+        self.patch_nums = int((self.seq_len - self.patch_len) / self.stride + 2)  # 补丁数量
+        self.head_nf = self.d_ff * self.patch_nums  # 头特征维度
+        
+        # 初始化GPT2模型
+        self.gpts = GPT2Model.from_pretrained(self.gpt_path, output_attentions=True, output_hidden_states=True)  # GPT2模型
+        self.gpts.h = self.gpts.h[:self.gpt_layers]  # 截取指定层数
+        self.gpts.apply(self.reset_parameters)  # 重置参数
+
+        self.word_embeddings = self.gpts.get_input_embeddings().weight.to(self.device)  # 词嵌入权重
+        self.vocab_size = self.word_embeddings.shape[0]  # 词汇表大小
+        self.mapping_layer = nn.Linear(self.vocab_size, 1)  # 映射层
+        self.reprogramming_layer = ReprogrammingLayer(self.d_model + configs.get('graph_hidden_dim', 32) * (configs.get('chebyshev_order', 3) + 1), self.n_heads, self.d_keys, self.d_llm)  # 重编程层
+        
+        # 初始化图增强编码器
+        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
+        )
+        
+        self.graph_projection = nn.Linear(  # 图特征投影层，每一k阶的切比雪夫权重映射到隐藏维度
+            configs.get('graph_hidden_dim', 32) * (configs.get('chebyshev_order', 3) + 1),  # 输入维度
+            self.d_model  # 输出维度
+        )
+    
+        self.out_mlp = nn.Sequential(
+            nn.Linear(self.d_llm, 128),
+            nn.ReLU(),
+            nn.Linear(128, self.pred_len)
+        )
+
+        # 设置参数可训练性 wps=word position embeddings
+        for name, param in self.gpts.named_parameters():
+                param.requires_grad = 'wpe' in name
+
+    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):
+        # 数据处理
+        x = x[..., :1]  # [B,T,N,1]
+        x_enc = rearrange(x, 'b t n c -> b n c t')  # [B,N,1,T]
+        
+        # 图编码
+        H_t = self.graph_encoder(x_enc)  # [B,N,1,T] -> [B, N, hidden_dim*(K+1)]
+        X_t_1 = self.graph_projection(H_t)  # [B,N,d_model]
+        enc_out = torch.cat([H_t, X_t_1], dim = -1) # [B, N, d_model + hidden_dim*(K+1)]
+        
+        # 词嵌入处理
+        self.mapping_layer(self.word_embeddings.permute(1, 0)).permute(1, 0)
+        masks = self.word_choice(self.mapping_layer.weight.data.permute(1,0))  # [d_llm,1]
+        source_embeddings = self.word_embeddings[masks==1]  # [selected_words,d_llm]
+   
+        # 重编程与预测
+        enc_out = self.reprogramming_layer(enc_out, source_embeddings, source_embeddings)
+        enc_out = self.gpts(inputs_embeds=enc_out).last_hidden_state  # [B,N,d_llm]
+        dec_out = self.out_mlp(enc_out)  # [B,N,pred_len]
+        
+        # 维度调整
+        outputs = dec_out.unsqueeze(dim=-1)  # [B,N,pred_len,1]
+        outputs = outputs.permute(0, 2, 1, 3)  # [B,pred_len,N,1]
+
+        return outputs

model/model_selector.py

@@ -25,6 +25,7 @@ from model.STAWnet.STAWnet import STAWnet
 from model.REPST.repst import repst as REPST
 from model.AEPSA.aepsa import AEPSA as AEPSA
 from model.AEPSA.aepsav2 import AEPSA as AEPSAv2
+from model.AEPSA.aepsav3 import AEPSA as AEPSAv3
 
 
 
@@ -86,3 +87,5 @@ def model_selector(config):
             return AEPSA(model_config)
         case "AEPSA_v2":
             return AEPSAv2(model_config)
+        case "AEPSA_v3":
+            return AEPSAv3(model_config)