exp12 残差

config/EXP/PEMSD3.yaml

@@ -17,6 +17,7 @@ model:
   input_dim: 1
   output_dim: 1
   embed_dim: 10
+  in_len: 12
   rnn_units: 64
   num_layers: 1
   cheb_order: 2

config/EXP/PEMSD4.yaml

@@ -18,20 +18,7 @@ model:
   input_dim: 1
   output_dim: 1
   in_len: 12
-  dropout: 0.3
-  supports: null
-  gcn_bool: true
-  addaptadj: true
-  aptinit: null
-  in_dim: 2
-  out_dim: 12
-  residual_channels: 32
-  dilation_channels: 32
-  skip_channels: 256
-  end_channels: 512
-  kernel_size: 2
-  blocks: 4
-  layers: 2
+
 
 train:
   loss_func: mae

config/EXP/PEMSD7.yaml

@@ -0,0 +1,45 @@
+data:
+  num_nodes: 883
+  lag: 12
+  horizon: 12
+  val_ratio: 0.2
+  test_ratio: 0.2
+  tod: False
+  normalizer: std
+  column_wise: False
+  default_graph: True
+  add_time_in_day: True
+  add_day_in_week: True
+  steps_per_day: 288
+  days_per_week: 7
+
+model:
+  batch_size: 64
+  input_dim: 1
+  output_dim: 1
+  in_len: 12
+
+
+train:
+  loss_func: mae
+  seed: 10
+  batch_size: 64
+  epochs: 300
+  lr_init: 0.003
+  weight_decay: 0
+  lr_decay: False
+  lr_decay_rate: 0.3
+  lr_decay_step: "5,20,40,70"
+  early_stop: True
+  early_stop_patience: 15
+  grad_norm: False
+  max_grad_norm: 5
+  real_value: True
+
+test:
+  mae_thresh: null
+  mape_thresh: 0.0
+
+log:
+  log_step: 200
+  plot: False

config/EXP/PEMSD8.yaml

@@ -14,18 +14,11 @@ data:
   days_per_week: 7
 
 model:
+  batch_size: 64
   input_dim: 1
   output_dim: 1
-  embed_dim: 12
-  rnn_units: 64
-  num_layers: 1
-  cheb_order: 2
-  use_day: True
-  use_week: True
-  graph_size: 30
-  expert_nums: 8
-  top_k: 2
-  hidden_dim: 64
+  in_len: 12
+
 
 train:
   loss_func: mae

model/EXP/EXP10.py

@@ -0,0 +1,156 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+"""
+KAN网络
+"""
+
+class KANLinear(nn.Module):
+    """
+    A simple Kolmogorov–Arnold Network linear layer.
+    y_k = sum_{q=1}^Q alpha_{kq} * phi_q( sum_{i=1}^I beta_{qi} * x_i )
+    """
+    def __init__(self, in_features, out_features, hidden_funcs=10):
+        super().__init__()
+        self.in_features = in_features
+        self.out_features = out_features
+        self.num_hidden = hidden_funcs
+        # mixing weights from input to Q hidden functions
+        self.beta = nn.Parameter(torch.randn(hidden_funcs, in_features))
+        # one univariate phi function per hidden channel
+        self.phi = nn.ModuleList([
+            nn.Sequential(nn.Linear(1, 1), nn.ReLU())
+            for _ in range(hidden_funcs)
+        ])
+        # mixing weights from hidden functions to outputs
+        self.alpha = nn.Parameter(torch.randn(out_features, hidden_funcs))
+
+    def forward(self, x):
+        # x: (..., in_features)
+        # compute univariate projections for each hidden func: u_q = sum_i beta_{qi} * x_i
+        u = torch.einsum('...i,qi->...q', x, self.beta)  # (..., Q)
+        # apply phi elementwise
+        u_phi = torch.stack([
+            self.phi[q](u[..., q].unsqueeze(-1)).squeeze(-1)
+            for q in range(self.num_hidden)
+        ], dim=-1)  # (..., Q)
+        # mix to out_features
+        y = torch.einsum('...q,kq->...k', u_phi, self.alpha)  # (..., out_features)
+        return y
+
+
+class DynamicGraphConstructor(nn.Module):
+    def __init__(self, node_num, embed_dim):
+        super().__init__()
+        self.nodevec1 = nn.Parameter(torch.randn(node_num, embed_dim), requires_grad=True)
+        self.nodevec2 = nn.Parameter(torch.randn(node_num, embed_dim), requires_grad=True)
+
+    def forward(self):
+        # (N, D) @ (D, N) -> (N, N)
+        adj = torch.matmul(self.nodevec1, self.nodevec2.T)
+        adj = F.relu(adj)
+        adj = F.softmax(adj, dim=-1)
+        return adj
+
+
+class GraphConvBlock(nn.Module):
+    def __init__(self, input_dim, output_dim, kan_hidden=8):
+        super().__init__()
+        self.theta = KANLinear(input_dim, output_dim, hidden_funcs=kan_hidden)
+        self.residual = input_dim == output_dim
+        if not self.residual:
+            self.res_proj = KANLinear(input_dim, output_dim, hidden_funcs=kan_hidden)
+
+    def forward(self, x, adj):
+        # x: (B, N, C) / adj: (N, N)
+        res = x
+        x = torch.matmul(adj, x)
+        # apply KAN-based linear mapping
+        B, N, C = x.shape
+        x = x.view(B * N, C)
+        x = self.theta(x)
+        x = x.view(B, N, -1)
+        if self.residual:
+            x = x + res
+        else:
+            x = x + self.res_proj(res.view(B*N, C)).view(B, N, -1)
+        return F.relu(x)
+
+
+class MANBA_Block(nn.Module):
+    def __init__(self, input_dim, hidden_dim):
+        super().__init__()
+        self.attn = nn.MultiheadAttention(embed_dim=input_dim, num_heads=4, batch_first=True)
+        self.ffn = nn.Sequential(
+            nn.Linear(input_dim, hidden_dim),
+            nn.ReLU(),
+            nn.Linear(hidden_dim, input_dim)
+        )
+        self.norm1 = nn.LayerNorm(input_dim)
+        self.norm2 = nn.LayerNorm(input_dim)
+
+    def forward(self, x):
+        # x: (B, N, C) -> treat N as temporal for attention
+        res = x
+        # swap dims to (B, T, C) for attn if needed
+        x_attn, _ = self.attn(x, x, x)
+        x = self.norm1(res + x_attn)
+
+        res2 = x
+        x_ffn = self.ffn(x)
+        x = self.norm2(res2 + x_ffn)
+        return x
+
+
+class EXP(nn.Module):
+    def __init__(self, args):
+        super().__init__()
+        self.horizon = args['horizon']
+        self.output_dim = args['output_dim']
+        self.seq_len = args.get('in_len', 12)
+        self.hidden_dim = args.get('hidden_dim', 64)
+        self.num_nodes = args['num_nodes']
+        kan_hidden = args.get('kan_hidden', 8)
+
+        # 动态图构建
+        self.graph = DynamicGraphConstructor(self.num_nodes, embed_dim=16)
+
+        # 输入映射：KAN替代线性层
+        self.input_proj = KANLinear(self.seq_len, self.hidden_dim, hidden_funcs=kan_hidden)
+
+        # 图卷积
+        self.gc = GraphConvBlock(self.hidden_dim, self.hidden_dim, kan_hidden=kan_hidden)
+
+        # 时间建模：保持MANBA
+        self.manba = MANBA_Block(self.hidden_dim, self.hidden_dim * 2)
+
+        # 输出映射：KAN替代线性层
+        out_size = self.horizon * self.output_dim
+        self.out_proj = KANLinear(self.hidden_dim, out_size, hidden_funcs=kan_hidden)
+
+    def forward(self, x):
+        # x: (B, T, N, D_total)
+        x = x[..., 0]
+        B, T, N = x.shape
+        assert T == self.seq_len
+
+        # 输入投影 (B, T, N) -> (B, N, T) -> (B*N, T)
+        x = x.permute(0, 2, 1).reshape(B * N, T)
+        h = self.input_proj(x)  # (B*N, hidden_dim)
+        h = h.view(B, N, self.hidden_dim)
+
+        # 动态图
+        adj = self.graph()
+
+        # 空间：图卷积
+        h = self.gc(h, adj)
+
+        # 时间：MANBA
+        h = self.manba(h)
+
+        # 输出
+        h_flat = h.view(B * N, self.hidden_dim)
+        out = self.out_proj(h_flat)
+        out = out.view(B, N, self.horizon, self.output_dim).permute(0, 2, 1, 3)
+        return out

model/EXP/EXP11.py

@@ -0,0 +1,123 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+"""
+含残差版本 + 时间-空间-时间（三明治结构）
+"""
+
+class DynamicGraphConstructor(nn.Module):
+    def __init__(self, node_num, embed_dim):
+        super().__init__()
+        self.nodevec1 = nn.Parameter(torch.randn(node_num, embed_dim), requires_grad=True)
+        self.nodevec2 = nn.Parameter(torch.randn(node_num, embed_dim), requires_grad=True)
+
+    def forward(self):
+        # (N, D) @ (D, N) -> (N, N)
+        adj = torch.matmul(self.nodevec1, self.nodevec2.T)
+        adj = F.relu(adj)
+        adj = F.softmax(adj, dim=-1)
+        return adj
+
+
+class GraphConvBlock(nn.Module):
+    def __init__(self, input_dim, output_dim):
+        super().__init__()
+        self.theta = nn.Linear(input_dim, output_dim)
+        self.residual = input_dim == output_dim
+        if not self.residual:
+            self.res_proj = nn.Linear(input_dim, output_dim)
+
+    def forward(self, x, adj):
+        # x: (B, N, C) / adj: (N, N)
+        res = x
+        x = torch.matmul(adj, x)  # (B, N, C)
+        x = self.theta(x)
+
+        # 残差连接
+        if self.residual:
+            x = x + res
+        else:
+            x = x + self.res_proj(res)
+
+        return F.relu(x)
+
+
+class MANBA_Block(nn.Module):
+    def __init__(self, input_dim, hidden_dim):
+        super().__init__()
+        self.attn = nn.MultiheadAttention(embed_dim=input_dim, num_heads=4, batch_first=True)
+        self.ffn = nn.Sequential(
+            nn.Linear(input_dim, hidden_dim),
+            nn.ReLU(),
+            nn.Linear(hidden_dim, input_dim)
+        )
+        self.norm1 = nn.LayerNorm(input_dim)
+        self.norm2 = nn.LayerNorm(input_dim)
+
+    def forward(self, x):
+        # x: (B, T, C) 或 (B, N, C) 当 N 视为 时间维度
+        res = x
+        x_attn, _ = self.attn(x, x, x)
+        x = self.norm1(res + x_attn)
+
+        res2 = x
+        x_ffn = self.ffn(x)
+        x = self.norm2(res2 + x_ffn)
+
+        return x
+
+
+class EXP(nn.Module):
+    def __init__(self, args):
+        super().__init__()
+        self.horizon = args['horizon']
+        self.output_dim = args['output_dim']
+        self.seq_len = args.get('in_len', 12)
+        self.hidden_dim = args.get('hidden_dim', 64)
+        self.num_nodes = args['num_nodes']
+
+        # 动态图构建
+        self.graph = DynamicGraphConstructor(self.num_nodes, embed_dim=16)
+
+        # 输入映射层
+        self.input_proj = nn.Linear(self.seq_len, self.hidden_dim)
+
+        # 图卷积
+        self.gc = GraphConvBlock(self.hidden_dim, self.hidden_dim)
+
+        # MANBA block（时间建模）
+        self.manba = MANBA_Block(self.hidden_dim, self.hidden_dim * 2)
+
+        # 输出映射
+        self.out_proj = nn.Linear(self.hidden_dim, self.horizon * self.output_dim)
+
+    def forward(self, x):
+        # x: (B, T, N, D_total)
+        x = x[..., 0]  # 只用主通道 (B, T, N)
+        B, T, N = x.shape
+        assert T == self.seq_len
+
+        # 输入投影 (B, T, N) -> (B, N, T) -> (B*N, T) -> (B*N, hidden_dim)
+        x_flat = x.permute(0, 2, 1).reshape(B * N, T)
+        h = self.input_proj(x_flat)  # (B*N, hidden_dim)
+        h = h.view(B, N, self.hidden_dim)  # (B, N, hidden_dim)
+
+        # === 时间建模（首次） ===
+        # 将 N 视作 时间维度进行注意力
+        h_time1 = self.manba(h)  # (B, N, hidden_dim)
+
+        # 动态图构建
+        adj = self.graph()  # (N, N)
+
+        # === 空间建模 ===
+        h_space = self.gc(h_time1, adj)  # (B, N, hidden_dim)
+
+        # === 时间建模（再一次） ===
+        h_time2 = self.manba(h_space)  # (B, N, hidden_dim)
+
+        # 输出映射
+        out = self.out_proj(h_time2)  # (B, N, horizon * output_dim)
+        out = out.view(B, N, self.horizon, self.output_dim).permute(0, 2, 1, 3)
+        return out  # (B, horizon, N, output_dim)
+

model/EXP/EXP12.py

@@ -0,0 +1,128 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+"""
+含残差的双层三明治结构模型
+第一层：时间 -> 空间 -> 时间
+残差连接：层输出 + 层输入
+第二层：同样三明治结构 -> 最终输出
+有效
+"""
+
+class DynamicGraphConstructor(nn.Module):
+    def __init__(self, node_num, embed_dim):
+        super().__init__()
+        self.nodevec1 = nn.Parameter(torch.randn(node_num, embed_dim), requires_grad=True)
+        self.nodevec2 = nn.Parameter(torch.randn(node_num, embed_dim), requires_grad=True)
+
+    def forward(self):
+        # (N, D) @ (D, N) -> (N, N)
+        adj = torch.matmul(self.nodevec1, self.nodevec2.T)
+        adj = F.relu(adj)
+        adj = F.softmax(adj, dim=-1)
+        return adj
+
+
+class GraphConvBlock(nn.Module):
+    def __init__(self, input_dim, output_dim):
+        super().__init__()
+        self.theta = nn.Linear(input_dim, output_dim)
+        self.residual = (input_dim == output_dim)
+        if not self.residual:
+            self.res_proj = nn.Linear(input_dim, output_dim)
+
+    def forward(self, x, adj):
+        # x: (B, N, C); adj: (N, N)
+        res = x
+        x = torch.matmul(adj, x)         # 空间卷积
+        x = self.theta(x)
+        # 残差
+        x = x + (res if self.residual else self.res_proj(res))
+        return F.relu(x)
+
+
+class MANBA_Block(nn.Module):
+    def __init__(self, input_dim, hidden_dim):
+        super().__init__()
+        self.attn = nn.MultiheadAttention(embed_dim=input_dim, num_heads=4, batch_first=True)
+        self.ffn = nn.Sequential(
+            nn.Linear(input_dim, hidden_dim),
+            nn.ReLU(),
+            nn.Linear(hidden_dim, input_dim)
+        )
+        self.norm1 = nn.LayerNorm(input_dim)
+        self.norm2 = nn.LayerNorm(input_dim)
+
+    def forward(self, x):
+        # x: (B, N, C) 视 N 维为时间序列长度
+        res = x
+        x_attn, _ = self.attn(x, x, x)
+        x = self.norm1(res + x_attn)
+        res2 = x
+        x_ffn = self.ffn(x)
+        x = self.norm2(res2 + x_ffn)
+        return x
+
+
+class SandwichBlock(nn.Module):
+    """
+    时间-空间-时间 三明治结构
+    输入/输出： (B, N, hidden_dim)
+    """
+    def __init__(self, num_nodes, embed_dim, hidden_dim):
+        super().__init__()
+        self.manba1 = MANBA_Block(hidden_dim, hidden_dim * 2)
+        self.graph_constructor = DynamicGraphConstructor(num_nodes, embed_dim)
+        self.gc = GraphConvBlock(hidden_dim, hidden_dim)
+        self.manba2 = MANBA_Block(hidden_dim, hidden_dim * 2)
+
+    def forward(self, h):
+        # h: (B, N, hidden_dim)
+        h1 = self.manba1(h)
+        adj = self.graph_constructor()         # (N, N)
+        h2 = self.gc(h1, adj)
+        h3 = self.manba2(h2)
+        return h3
+
+
+class EXP(nn.Module):
+    def __init__(self, args):
+        super().__init__()
+        self.horizon = args['horizon']
+        self.output_dim = args['output_dim']
+        self.seq_len = args.get('in_len', 12)
+        self.hidden_dim = args.get('hidden_dim', 64)
+        self.num_nodes = args['num_nodes']
+        self.embed_dim = args.get('embed_dim', 16)
+
+        # 输入映射
+        self.input_proj = nn.Linear(self.seq_len, self.hidden_dim)
+        # 两层三明治块
+        self.sandwich1 = SandwichBlock(self.num_nodes, self.embed_dim, self.hidden_dim)
+        self.sandwich2 = SandwichBlock(self.num_nodes, self.embed_dim, self.hidden_dim)
+        # 输出映射
+        self.out_proj = nn.Linear(self.hidden_dim, self.horizon * self.output_dim)
+
+    def forward(self, x):
+        # x: (B, T, N, D_total)
+        x_main = x[..., 0]              # (B, T, N)
+        B, T, N = x_main.shape
+        assert T == self.seq_len
+
+        # 输入投影 (B, T, N) -> (B*N, T) -> (B, N, hidden_dim)
+        x_flat = x_main.permute(0, 2, 1).reshape(B * N, T)
+        h0 = self.input_proj(x_flat).view(B, N, self.hidden_dim)
+
+        # 第一层三明治 + 残差
+        h1 = self.sandwich1(h0)
+        h1 = h1 + h0
+
+        # 第二层三明治
+        h2 = self.sandwich2(h1)
+
+        # 输出映射
+        out = self.out_proj(h2)          # (B, N, H*D_out)
+        out = out.view(B, N, self.horizon, self.output_dim)
+        out = out.permute(0, 2, 1, 3)    # (B, horizon, N, output_dim)
+        return out

model/EXP/EXP13.py

@@ -0,0 +1,125 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+"""
+含残差的双层 空间->时间->空间 结构模型 无效
+"""
+
+class DynamicGraphConstructor(nn.Module):
+    def __init__(self, node_num, embed_dim):
+        super().__init__()
+        self.nodevec1 = nn.Parameter(torch.randn(node_num, embed_dim), requires_grad=True)
+        self.nodevec2 = nn.Parameter(torch.randn(node_num, embed_dim), requires_grad=True)
+
+    def forward(self):
+        # 构造动态邻接矩阵 (N, N)
+        adj = torch.matmul(self.nodevec1, self.nodevec2.T)
+        adj = F.relu(adj)
+        adj = F.softmax(adj, dim=-1)
+        return adj
+
+
+class GraphConvBlock(nn.Module):
+    def __init__(self, input_dim, output_dim):
+        super().__init__()
+        self.theta = nn.Linear(input_dim, output_dim)
+        self.residual = (input_dim == output_dim)
+        if not self.residual:
+            self.res_proj = nn.Linear(input_dim, output_dim)
+
+    def forward(self, x, adj):
+        # x: (B, N, C)
+        res = x
+        x = torch.matmul(adj, x)
+        x = self.theta(x)
+        x = x + (res if self.residual else self.res_proj(res))
+        return F.relu(x)
+
+
+class MANBA_Block(nn.Module):
+    def __init__(self, input_dim, hidden_dim):
+        super().__init__()
+        self.attn = nn.MultiheadAttention(embed_dim=input_dim, num_heads=4, batch_first=True)
+        self.ffn = nn.Sequential(
+            nn.Linear(input_dim, hidden_dim),
+            nn.ReLU(),
+            nn.Linear(hidden_dim, input_dim)
+        )
+        self.norm1 = nn.LayerNorm(input_dim)
+        self.norm2 = nn.LayerNorm(input_dim)
+
+    def forward(self, x):
+        # x: (B, N, C) 当 N 视为时间序列长度
+        res = x
+        x_attn, _ = self.attn(x, x, x)
+        x = self.norm1(res + x_attn)
+        res2 = x
+        x_ffn = self.ffn(x)
+        x = self.norm2(res2 + x_ffn)
+        return x
+
+
+class SandwichBlock(nn.Module):
+    """
+    空间 -> 时间 -> 空间 三明治结构
+    输入/输出： (B, N, hidden_dim)
+    """
+    def __init__(self, num_nodes, embed_dim, hidden_dim):
+        super().__init__()
+        self.graph_constructor = DynamicGraphConstructor(num_nodes, embed_dim)
+        self.gc1 = GraphConvBlock(hidden_dim, hidden_dim)
+        self.manba = MANBA_Block(hidden_dim, hidden_dim * 2)
+        self.gc2 = GraphConvBlock(hidden_dim, hidden_dim)
+
+    def forward(self, h):
+        # 第一步：空间卷积
+        adj = self.graph_constructor()
+        h1 = self.gc1(h, adj)
+        # 第二步：时间注意力
+        h2 = self.manba(h1)
+        # 第三步：空间卷积
+        h3 = self.gc2(h2, adj)
+        return h3
+
+
+class EXP(nn.Module):
+    def __init__(self, args):
+        super().__init__()
+        self.horizon = args['horizon']
+        self.output_dim = args['output_dim']
+        self.seq_len = args.get('in_len', 12)
+        self.hidden_dim = args.get('hidden_dim', 64)
+        self.num_nodes = args['num_nodes']
+        self.embed_dim = args.get('embed_dim', 16)
+
+        # 输入映射
+        self.input_proj = nn.Linear(self.seq_len, self.hidden_dim)
+        # 两层 空间-时间-空间 三明治块
+        self.sandwich1 = SandwichBlock(self.num_nodes, self.embed_dim, self.hidden_dim)
+        self.sandwich2 = SandwichBlock(self.num_nodes, self.embed_dim, self.hidden_dim)
+        # 输出映射
+        self.out_proj = nn.Linear(self.hidden_dim, self.horizon * self.output_dim)
+
+    def forward(self, x):
+        # x: (B, T, N, D)
+        x_main = x[..., 0]            # (B, T, N)
+        B, T, N = x_main.shape
+        assert T == self.seq_len
+
+        # 投影到隐藏维 (B,N,hidden)
+        x_flat = x_main.permute(0, 2, 1).reshape(B * N, T)
+        h0 = self.input_proj(x_flat).view(B, N, self.hidden_dim)
+
+        # 第一层三明治 + 残差
+        h1 = self.sandwich1(h0)
+        h1 = h1 + h0
+        # 第二层三明治
+        h2 = self.sandwich2(h1)
+
+        # 输出
+        out = self.out_proj(h2)       # (B, N, H*D_out)
+        out = out.view(B, N, self.horizon, self.output_dim)
+        out = out.permute(0, 2, 1, 3)  # (B, horizon, N, output_dim)
+        return out
+

model/EXP/EXP14.py

@@ -0,0 +1,147 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+"""
+含时间/空间额外特征的双层 时间->空间->时间 三明治结构模型
+使用 x[...,0] 主通道，x[...,1] time_in_day，x[...,2] day_in_week
+通过独立投影融合三路信息
+无改进
+"""
+
+class DynamicGraphConstructor(nn.Module):
+    def __init__(self, node_num, embed_dim):
+        super().__init__()
+        self.nodevec1 = nn.Parameter(torch.randn(node_num, embed_dim), requires_grad=True)
+        self.nodevec2 = nn.Parameter(torch.randn(node_num, embed_dim), requires_grad=True)
+
+    def forward(self):
+        # 构造动态邻接矩阵 (N, N)
+        adj = torch.matmul(self.nodevec1, self.nodevec2.T)
+        adj = F.relu(adj)
+        adj = F.softmax(adj, dim=-1)
+        return adj
+
+
+class GraphConvBlock(nn.Module):
+    def __init__(self, input_dim, output_dim):
+        super().__init__()
+        self.theta = nn.Linear(input_dim, output_dim)
+        self.residual = (input_dim == output_dim)
+        if not self.residual:
+            self.res_proj = nn.Linear(input_dim, output_dim)
+
+    def forward(self, x, adj):
+        # x: (B, N, C)
+        res = x
+        x = torch.matmul(adj, x)
+        x = self.theta(x)
+        x = x + (res if self.residual else self.res_proj(res))
+        return F.relu(x)
+
+
+class MANBA_Block(nn.Module):
+    def __init__(self, input_dim, hidden_dim):
+        super().__init__()
+        self.attn = nn.MultiheadAttention(embed_dim=input_dim, num_heads=4, batch_first=True)
+        self.ffn = nn.Sequential(
+            nn.Linear(input_dim, hidden_dim),
+            nn.ReLU(),
+            nn.Linear(hidden_dim, input_dim)
+        )
+        self.norm1 = nn.LayerNorm(input_dim)
+        self.norm2 = nn.LayerNorm(input_dim)
+
+    def forward(self, x):
+        # x: (B, N, C) 视 N 维为时间序列长度
+        res = x
+        x_attn, _ = self.attn(x, x, x)
+        x = self.norm1(res + x_attn)
+        res2 = x
+        x_ffn = self.ffn(x)
+        x = self.norm2(res2 + x_ffn)
+        return x
+
+
+class SandwichBlock(nn.Module):
+    """
+    时间 -> 空间 -> 时间 三明治结构
+    输入/输出： (B, N, hidden_dim)
+    """
+    def __init__(self, num_nodes, embed_dim, hidden_dim):
+        super().__init__()
+        self.manba1 = MANBA_Block(hidden_dim, hidden_dim * 2)
+        self.graph_constructor = DynamicGraphConstructor(num_nodes, embed_dim)
+        self.gc = GraphConvBlock(hidden_dim, hidden_dim)
+        self.manba2 = MANBA_Block(hidden_dim, hidden_dim * 2)
+
+    def forward(self, h):
+        # h: (B, N, hidden_dim)
+        # 第一步：时间注意力
+        h1 = self.manba1(h)
+        # 第二步：空间卷积
+        adj = self.graph_constructor()
+        h2 = self.gc(h1, adj)
+        # 第三步：时间注意力
+        h3 = self.manba2(h2)
+        return h3
+
+
+class EXP(nn.Module):
+    def __init__(self, args):
+        super().__init__()
+        self.horizon = args['horizon']
+        self.output_dim = args['output_dim']
+        self.seq_len = args.get('in_len', 12)
+        self.hidden_dim = args.get('hidden_dim', 64)
+        self.num_nodes = args['num_nodes']
+        self.embed_dim = args.get('embed_dim', 16)
+
+        # 对三路输入分别投影到隐藏维度
+        self.main_proj = nn.Linear(self.seq_len, self.hidden_dim)
+        self.time_proj = nn.Linear(self.seq_len, self.hidden_dim)
+        self.week_proj = nn.Linear(self.seq_len, self.hidden_dim)
+
+        # 双层 时间->空间->时间 三明治块
+        self.sandwich1 = SandwichBlock(self.num_nodes, self.embed_dim, self.hidden_dim)
+        self.sandwich2 = SandwichBlock(self.num_nodes, self.embed_dim, self.hidden_dim)
+
+        # 输出映射
+        self.out_proj = nn.Linear(self.hidden_dim, self.horizon * self.output_dim)
+
+    def forward(self, x):
+        # x: (B, T, N, D_total)
+        x_main = x[..., 0]   # (B, T, N)
+        x_time = x[..., 1]   # (B, T, N)
+        x_week = x[..., 2]   # (B, T, N)
+        B, T, N = x_main.shape
+        assert T == self.seq_len
+
+        # 将三路特征分别映射后叠加
+        x_main_flat = x_main.permute(0, 2, 1).reshape(B * N, T)
+        h_main = self.main_proj(x_main_flat).view(B, N, self.hidden_dim)
+        x_time_flat = x_time.permute(0, 2, 1).reshape(B * N, T)
+        h_time = self.time_proj(x_time_flat).view(B, N, self.hidden_dim)
+        x_week_flat = x_week.permute(0, 2, 1).reshape(B * N, T)
+        h_week = self.week_proj(x_week_flat).view(B, N, self.hidden_dim)
+
+        # 初始隐藏表示，融合三路信息
+        h0 = h_main + h_time + h_week
+
+        # 第一层三明治 + 残差
+        h1 = self.sandwich1(h0)
+        h1 = h1 + h0
+        # 第二层三明治
+        h2 = self.sandwich2(h1)
+
+        # 输出
+        out = self.out_proj(h2)
+        out = out.view(B, N, self.horizon, self.output_dim)
+        out = out.permute(0, 2, 1, 3)  # (B, horizon, N, D_out)
+        return out
+
+# 示例测试
+# args = {'horizon':12,'output_dim':1,'in_len':12,'hidden_dim':64,'num_nodes':307,'embed_dim':16}
+# model = EXP(args)
+# x = torch.randn(16, 12, 307, 3)
+# print(model(x).shape)  # (16,12,307,1)

model/EXP/EXP15.py

@@ -0,0 +1,139 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+"""
+含残差的双层三明治结构模型
+第一层：时间 -> 空间 -> 时间 -> Conv -> Residual差分 -> 输入第二层
+第二层：时间 -> 空间 -> 时间 -> Conv -> 最终输出
+无效但接近
+"""
+
+class DynamicGraphConstructor(nn.Module):
+    def __init__(self, node_num, embed_dim):
+        super().__init__()
+        self.nodevec1 = nn.Parameter(torch.randn(node_num, embed_dim), requires_grad=True)
+        self.nodevec2 = nn.Parameter(torch.randn(node_num, embed_dim), requires_grad=True)
+
+    def forward(self):
+        # (N, D) @ (D, N) -> (N, N)
+        adj = torch.matmul(self.nodevec1, self.nodevec2.T)
+        adj = F.relu(adj)
+        adj = F.softmax(adj, dim=-1)
+        return adj
+
+
+class GraphConvBlock(nn.Module):
+    def __init__(self, input_dim, output_dim):
+        super().__init__()
+        self.theta = nn.Linear(input_dim, output_dim)
+        self.residual = (input_dim == output_dim)
+        if not self.residual:
+            self.res_proj = nn.Linear(input_dim, output_dim)
+
+    def forward(self, x, adj):
+        # x: (B, N, C); adj: (N, N)
+        res = x
+        x = torch.matmul(adj, x)         # 空间卷积
+        x = self.theta(x)
+        # 残差
+        x = x + (res if self.residual else self.res_proj(res))
+        return F.relu(x)
+
+
+class MANBA_Block(nn.Module):
+    def __init__(self, input_dim, hidden_dim):
+        super().__init__()
+        self.attn = nn.MultiheadAttention(embed_dim=input_dim, num_heads=4, batch_first=True)
+        self.ffn = nn.Sequential(
+            nn.Linear(input_dim, hidden_dim),
+            nn.ReLU(),
+            nn.Linear(hidden_dim, input_dim)
+        )
+        self.norm1 = nn.LayerNorm(input_dim)
+        self.norm2 = nn.LayerNorm(input_dim)
+
+    def forward(self, x):
+        # x: (B, N, C) 视 N 维为时间序列长度
+        res = x
+        x_attn, _ = self.attn(x, x, x)
+        x = self.norm1(res + x_attn)
+        res2 = x
+        x_ffn = self.ffn(x)
+        x = self.norm2(res2 + x_ffn)
+        return x
+
+
+class SandwichBlock(nn.Module):
+    """
+    时间-空间-时间 三明治结构
+    输入/输出： (B, N, hidden_dim)
+    """
+    def __init__(self, num_nodes, embed_dim, hidden_dim):
+        super().__init__()
+        self.manba1 = MANBA_Block(hidden_dim, hidden_dim * 2)
+        self.graph_constructor = DynamicGraphConstructor(num_nodes, embed_dim)
+        self.gc = GraphConvBlock(hidden_dim, hidden_dim)
+        self.manba2 = MANBA_Block(hidden_dim, hidden_dim * 2)
+
+    def forward(self, h):
+        # h: (B, N, hidden_dim)
+        h1 = self.manba1(h)
+        adj = self.graph_constructor()         # (N, N)
+        h2 = self.gc(h1, adj)
+        h3 = self.manba2(h2)
+        return h3
+
+
+class EXP(nn.Module):
+    def __init__(self, args):
+        super().__init__()
+        self.horizon = args['horizon']
+        self.output_dim = args['output_dim']
+        self.seq_len = args.get('in_len', 12)
+        self.hidden_dim = args.get('hidden_dim', 64)
+        self.num_nodes = args['num_nodes']
+        self.embed_dim = args.get('embed_dim', 16)
+
+        # 输入映射: (batch*N, seq_len) -> hidden_dim
+        self.input_proj = nn.Linear(self.seq_len, self.hidden_dim)
+        # 两层三明治块
+        self.sandwich1 = SandwichBlock(self.num_nodes, self.embed_dim, self.hidden_dim)
+        self.sandwich2 = SandwichBlock(self.num_nodes, self.embed_dim, self.hidden_dim)
+        # 卷积层用于残差处理
+        self.res_conv1 = nn.Conv1d(in_channels=self.hidden_dim, out_channels=self.hidden_dim, kernel_size=1)
+        self.res_conv2 = nn.Conv1d(in_channels=self.hidden_dim, out_channels=self.hidden_dim, kernel_size=1)
+        # 输出映射
+        self.out_proj = nn.Linear(self.hidden_dim, self.horizon * self.output_dim)
+
+    def forward(self, x):
+        # x: (B, T, N, D_total)
+        x_main = x[..., 0]              # (B, T, N)
+        B, T, N = x_main.shape
+        assert T == self.seq_len
+
+        # 输入投影 (B, T, N) -> (B*N, T) -> (B, N, hidden_dim)
+        x_flat = x_main.permute(0, 2, 1).reshape(B * N, T)
+        h0 = self.input_proj(x_flat).view(B, N, self.hidden_dim)
+
+        # 第一层三明治
+        h1_sand = self.sandwich1(h0)   # (B, N, hidden_dim)
+        # 卷积残差 (节点维度视为长度)
+        h1_perm = h1_sand.permute(0, 2, 1)  # (B, C, N)
+        h1_conv = self.res_conv1(h1_perm)
+        h1 = h1_conv.permute(0, 2, 1)       # (B, N, hidden_dim)
+
+        # 计算差分残差作为第二层输入
+        h2_input = h1 - h0
+
+        # 第二层三明治
+        h2_sand = self.sandwich2(h2_input)
+        # 再次卷积处理
+        h2_perm = h2_sand.permute(0, 2, 1)
+        h2 = self.res_conv2(h2_perm).permute(0, 2, 1)
+
+        # 输出映射
+        out = self.out_proj(h2)          # (B, N, H*D_out)
+        out = out.view(B, N, self.horizon, self.output_dim)
+        out = out.permute(0, 2, 1, 3)    # (B, horizon, N, output_dim)
+        return out

model/EXP/EXP16.py

@@ -0,0 +1,119 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+"""
+含残差的双层三明治结构模型
+第一层：时间 -> 空间 -> 时间
+残差连接：层输出 + 层输入
+第二层：同样三明治结构 -> 最终输出
+无小残差
+"""
+
+class DynamicGraphConstructor(nn.Module):
+    def __init__(self, node_num, embed_dim):
+        super().__init__()
+        self.nodevec1 = nn.Parameter(torch.randn(node_num, embed_dim), requires_grad=True)
+        self.nodevec2 = nn.Parameter(torch.randn(node_num, embed_dim), requires_grad=True)
+
+    def forward(self):
+        # (N, D) @ (D, N) -> (N, N)
+        adj = torch.matmul(self.nodevec1, self.nodevec2.T)
+        adj = F.relu(adj)
+        adj = F.softmax(adj, dim=-1)
+        return adj
+
+
+class GraphConvBlock(nn.Module):
+    def __init__(self, input_dim, output_dim):
+        super().__init__()
+        self.theta = nn.Linear(input_dim, output_dim)
+
+    def forward(self, x, adj):
+        # x: (B, N, C) / adj: (N, N)
+        x = torch.matmul(adj, x)  # (B, N, C)
+        x = self.theta(x)
+        return F.relu(x)
+
+
+class MANBA_Block(nn.Module):
+    def __init__(self, input_dim, hidden_dim):
+        super().__init__()
+        self.attn = nn.MultiheadAttention(embed_dim=input_dim, num_heads=4, batch_first=True)
+        self.ffn = nn.Sequential(
+            nn.Linear(input_dim, hidden_dim),
+            nn.ReLU(),
+            nn.Linear(hidden_dim, input_dim)
+        )
+        self.norm1 = nn.LayerNorm(input_dim)
+        self.norm2 = nn.LayerNorm(input_dim)
+
+    def forward(self, x):
+        # x: (B, T, C)
+        x_attn, _ = self.attn(x, x, x)
+        x = self.norm1(x + x_attn)
+        x_ffn = self.ffn(x)
+        return self.norm2(x + x_ffn)
+
+
+class SandwichBlock(nn.Module):
+    """
+    时间-空间-时间 三明治结构
+    输入/输出： (B, N, hidden_dim)
+    """
+    def __init__(self, num_nodes, embed_dim, hidden_dim):
+        super().__init__()
+        self.manba1 = MANBA_Block(hidden_dim, hidden_dim * 2)
+        self.graph_constructor = DynamicGraphConstructor(num_nodes, embed_dim)
+        self.gc = GraphConvBlock(hidden_dim, hidden_dim)
+        self.manba2 = MANBA_Block(hidden_dim, hidden_dim * 2)
+
+    def forward(self, h):
+        # h: (B, N, hidden_dim)
+        h1 = self.manba1(h)
+        adj = self.graph_constructor()         # (N, N)
+        h2 = self.gc(h1, adj)
+        h3 = self.manba2(h2)
+        return h3
+
+
+class EXP(nn.Module):
+    def __init__(self, args):
+        super().__init__()
+        self.horizon = args['horizon']
+        self.output_dim = args['output_dim']
+        self.seq_len = args.get('in_len', 12)
+        self.hidden_dim = args.get('hidden_dim', 64)
+        self.num_nodes = args['num_nodes']
+        self.embed_dim = args.get('embed_dim', 16)
+
+        # 输入映射
+        self.input_proj = nn.Linear(self.seq_len, self.hidden_dim)
+        # 两层三明治块
+        self.sandwich1 = SandwichBlock(self.num_nodes, self.embed_dim, self.hidden_dim)
+        self.sandwich2 = SandwichBlock(self.num_nodes, self.embed_dim, self.hidden_dim)
+        # 输出映射
+        self.out_proj = nn.Linear(self.hidden_dim, self.horizon * self.output_dim)
+
+    def forward(self, x):
+        # x: (B, T, N, D_total)
+        x_main = x[..., 0]              # (B, T, N)
+        B, T, N = x_main.shape
+        assert T == self.seq_len
+
+        # 输入投影 (B, T, N) -> (B*N, T) -> (B, N, hidden_dim)
+        x_flat = x_main.permute(0, 2, 1).reshape(B * N, T)
+        h0 = self.input_proj(x_flat).view(B, N, self.hidden_dim)
+
+        # 第一层三明治 + 残差
+        h1 = self.sandwich1(h0)
+        h1 = h1 + h0
+
+        # 第二层三明治
+        h2 = self.sandwich2(h1)
+
+        # 输出映射
+        out = self.out_proj(h2)          # (B, N, H*D_out)
+        out = out.view(B, N, self.horizon, self.output_dim)
+        out = out.permute(0, 2, 1, 3)    # (B, horizon, N, output_dim)
+        return out

model/EXP/EXP7.py

@@ -2,6 +2,10 @@ import torch
 import torch.nn as nn
 import torch.nn.functional as F
 
+"""
+不含残差版本
+"""
+
 
 class DynamicGraphConstructor(nn.Module):
     def __init__(self, node_num, embed_dim):
@@ -75,7 +79,7 @@ class EXP(nn.Module):
 
     def forward(self, x):
         # x: (B, T, N, D_total)
-        x = x.sum(dim=-1)  # (B, T, N)
+        x = x[..., 0]  # 只用主通道 (B, T, N)
         B, T, N = x.shape
         assert T == self.seq_len
 

model/EXP/EXP8.py

@@ -2,6 +2,9 @@ import torch
 import torch.nn as nn
 import torch.nn.functional as F
 
+"""
+含残差版本
+"""
 
 class DynamicGraphConstructor(nn.Module):
     def __init__(self, node_num, embed_dim):

model/EXP/EXP9.py

@@ -2,6 +2,9 @@ import torch
 import torch.nn as nn
 import torch.nn.functional as F
 
+"""
+加入混合专家
+"""
 
 class DynamicGraphConstructor(nn.Module):
     def __init__(self, node_num, embed_dim):

model/model_selector.py

@@ -13,7 +13,7 @@ from model.STFGNN.STFGNN import STFGNN
 from model.STSGCN.STSGCN import STSGCN
 from model.STGODE.STGODE import ODEGCN
 from model.PDG2SEQ.PDG2Seq import PDG2Seq
-from model.EXP.EXP9 import EXP as EXP
+from model.EXP.EXP16 import EXP as EXP
 
 def model_selector(model):
     match model['type']:

run.py

@@ -30,10 +30,12 @@ def main():
     else:
         args['device'] = 'cpu'
         args['model']['device'] = args['device']
-
+    init_seed(args['train']['seed'])
     # Initialize model
     model = init_model(args['model'], device=args['device'])
 
+
+
     if args['mode'] == "benchmark":
         # 支持计算消耗分析，设置 mode为 benchmark
         import torch.profiler as profiler