171 lines
5.3 KiB
Python
Executable File
171 lines
5.3 KiB
Python
Executable File
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
|