129 lines
4.2 KiB
Python
Executable File
129 lines
4.2 KiB
Python
Executable File
import torch
|
|
import torch.nn as nn
|
|
import torch.nn.functional as F
|
|
|
|
|
|
"""
|
|
使用多层感知机替换输入输出的proj层
|
|
"""
|
|
|
|
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):
|
|
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):
|
|
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):
|
|
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):
|
|
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):
|
|
h1 = self.manba1(h)
|
|
adj = self.graph_constructor()
|
|
h2 = self.gc(h1, adj)
|
|
h3 = self.manba2(h2)
|
|
return h3
|
|
|
|
|
|
class MLP(nn.Module):
|
|
def __init__(self, in_dim, hidden_dims, out_dim, activation=nn.ReLU):
|
|
super().__init__()
|
|
dims = [in_dim] + hidden_dims + [out_dim]
|
|
layers = []
|
|
for i in range(len(dims)-2):
|
|
layers += [nn.Linear(dims[i], dims[i+1]), activation()]
|
|
layers += [nn.Linear(dims[-2], dims[-1])]
|
|
self.net = nn.Sequential(*layers)
|
|
|
|
def forward(self, x):
|
|
return self.net(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']
|
|
self.embed_dim = args.get('embed_dim', 16)
|
|
|
|
# 替换为MLP: input_proj(seq_len -> hidden_dim -> hidden_dim)
|
|
self.input_proj = MLP(self.seq_len, [self.hidden_dim], self.hidden_dim)
|
|
|
|
self.sandwich1 = SandwichBlock(self.num_nodes, self.embed_dim, self.hidden_dim)
|
|
self.sandwich2 = SandwichBlock(self.num_nodes, self.embed_dim, self.hidden_dim)
|
|
|
|
# 替换为MLP: out_proj(hidden_dim -> 2*hidden_dim -> horizon*output_dim)
|
|
self.out_proj = MLP(self.hidden_dim, [2 * 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
|
|
|
|
# (B, T, N) -> (B, N, T) -> (B*N, T) -> MLP -> (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)
|
|
|
|
h1 = self.sandwich1(h0)
|
|
h1 = h1 + h0
|
|
h2 = self.sandwich2(h1)
|
|
|
|
# MLP输出 -> (B, N, H*D_out)
|
|
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, output_dim)
|
|
return out
|