Files
timelib/models/MSGNet.py
flowerstonezl f7291b1221 first commit
2026-01-29 17:08:54 +08:00

247 lines
9.3 KiB
Python

import numpy as np
# import pywt
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.fft
from layers.Embed import DataEmbedding
from layers.MSGBlock import GraphBlock, simpleVIT, Attention_Block, Predict
def FFT_for_Period(x, k=2):
# [B, T, C]
xf = torch.fft.rfft(x, dim=1)
frequency_list = abs(xf).mean(0).mean(-1)
frequency_list[0] = 0
_, top_list = torch.topk(frequency_list, k)
top_list = top_list.detach().cpu().numpy()
period = x.shape[1] // top_list
return period, abs(xf).mean(-1)[:, top_list]
class ScaleGraphBlock(nn.Module):
def __init__(self, configs):
super(ScaleGraphBlock, self).__init__()
self.seq_len = configs.seq_len
self.pred_len = configs.pred_len
self.k = configs.top_k
self.att0 = Attention_Block(configs.d_model, configs.d_ff,
n_heads=configs.n_heads, dropout=configs.dropout, activation="gelu")
self.norm = nn.LayerNorm(configs.d_model)
self.gelu = nn.GELU()
self.gconv = nn.ModuleList()
for i in range(self.k):
self.gconv.append(
GraphBlock(configs.c_out , configs.d_model , configs.conv_channel, configs.skip_channel,
configs.gcn_depth , configs.dropout, configs.propalpha ,configs.seq_len,
configs.node_dim))
def forward(self, x):
B, T, N = x.size()
scale_list, scale_weight = FFT_for_Period(x, self.k)
res = []
for i in range(self.k):
scale = scale_list[i]
#Gconv
x = self.gconv[i](x)
# paddng
if (self.seq_len) % scale != 0:
length = (((self.seq_len) // scale) + 1) * scale
padding = torch.zeros([x.shape[0], (length - (self.seq_len)), x.shape[2]]).to(x.device)
out = torch.cat([x, padding], dim=1)
else:
length = self.seq_len
out = x
out = out.reshape(B, length // scale, scale, N)
#for Mul-attetion
out = out.reshape(-1 , scale , N)
out = self.norm(self.att0(out))
out = self.gelu(out)
out = out.reshape(B, -1 , scale , N).reshape(B ,-1 ,N)
# #for simpleVIT
# out = self.att(out.permute(0, 3, 1, 2).contiguous()) #return
# out = out.permute(0, 2, 3, 1).reshape(B, -1 ,N)
out = out[:, :self.seq_len, :]
res.append(out)
res = torch.stack(res, dim=-1)
# adaptive aggregation
scale_weight = F.softmax(scale_weight, dim=1)
scale_weight = scale_weight.unsqueeze(1).unsqueeze(1).repeat(1, T, N, 1)
res = torch.sum(res * scale_weight, -1)
# residual connection
res = res + x
return res
class Model(nn.Module):
def __init__(self, configs):
super(Model, self).__init__()
self.configs = configs
self.task_name = configs.task_name
self.seq_len = configs.seq_len
self.label_len = configs.label_len
self.pred_len = configs.pred_len
self.device = "cuda" if torch.cuda.is_available() else "cpu"
# for graph
# self.num_nodes = configs.c_out
# self.subgraph_size = configs.subgraph_size
# self.node_dim = configs.node_dim
# to return adj (node , node)
# self.graph = constructor_graph()
self.model = nn.ModuleList([ScaleGraphBlock(configs) for _ in range(configs.e_layers)])
self.enc_embedding = DataEmbedding(configs.enc_in, configs.d_model,
configs.embed, configs.freq, configs.dropout)
self.layer = configs.e_layers
self.layer_norm = nn.LayerNorm(configs.d_model)
self.predict_linear = nn.Linear(
self.seq_len, self.pred_len + self.seq_len)
self.projection = nn.Linear(configs.d_model, configs.c_out, bias=True)
self.seq2pred = Predict(configs.individual, configs.c_out,
configs.seq_len, configs.pred_len, configs.dropout)
if self.task_name == 'imputation' or self.task_name == 'anomaly_detection':
self.projection = nn.Linear(configs.d_model, configs.c_out, bias=True)
if self.task_name == 'classification':
self.act = F.gelu
self.dropout = nn.Dropout(configs.dropout)
self.projection = nn.Linear(
configs.d_model * configs.seq_len, configs.num_class)
def forecast(self, x_enc, x_mark_enc, x_dec, x_mark_dec, mask=None):
# Normalization from Non-stationary Transformer
means = x_enc.mean(1, keepdim=True).detach()
x_enc = x_enc - means
stdev = torch.sqrt(
torch.var(x_enc, dim=1, keepdim=True, unbiased=False) + 1e-5)
x_enc /= stdev
# embedding
enc_out = self.enc_embedding(x_enc, x_mark_enc) # [B,T,C]
# adp = self.graph(torch.arange(self.num_nodes).to(self.device))
for i in range(self.layer):
enc_out = self.layer_norm(self.model[i](enc_out))
# porject back
dec_out = self.projection(enc_out)
dec_out = self.seq2pred(dec_out.transpose(1, 2)).transpose(1, 2)
# De-Normalization from Non-stationary Transformer
dec_out = dec_out * \
(stdev[:, 0, :].unsqueeze(1).repeat(
1, self.pred_len, 1))
dec_out = dec_out + \
(means[:, 0, :].unsqueeze(1).repeat(
1, self.pred_len, 1))
return dec_out[:, -self.pred_len:, :]
def imputation(self, x_enc, x_mark_enc, x_dec, x_mark_dec, mask=None):
# Normalization from Non-stationary Transformer
means = x_enc.mean(1, keepdim=True).detach()
x_enc = x_enc - means
stdev = torch.sqrt(
torch.var(x_enc, dim=1, keepdim=True, unbiased=False) + 1e-5)
x_enc /= stdev
_, L, N = x_enc.shape
# embedding
enc_out = self.enc_embedding(x_enc, x_mark_enc) # [B,T,C]
# adp = self.graph(torch.arange(self.num_nodes).to(self.device))
for i in range(self.layer):
enc_out = self.layer_norm(self.model[i](enc_out))
# porject back
dec_out = self.projection(enc_out)
# dec_out = self.seq2pred(dec_out.transpose(1, 2)).transpose(1, 2)
# print(dec_out.shape)
# De-Normalization from Non-stationary Transformer
dec_out = dec_out * \
(stdev[:, 0, :].unsqueeze(1).repeat(
1, L, 1))
dec_out = dec_out + \
(means[:, 0, :].unsqueeze(1).repeat(
1, L, 1))
return dec_out
def anomaly_detection(self, x_enc):
# Normalization from Non-stationary Transformer
means = x_enc.mean(1, keepdim=True).detach()
x_enc = x_enc - means
stdev = torch.sqrt(
torch.var(x_enc, dim=1, keepdim=True, unbiased=False) + 1e-5)
x_enc /= stdev
_, L, N = x_enc.shape
# embedding
enc_out = self.enc_embedding(x_enc, None) # [B,T,C]
# adp = self.graph(torch.arange(self.num_nodes).to(self.device))
for i in range(self.layer):
enc_out = self.layer_norm(self.model[i](enc_out))
# porject back
dec_out = self.projection(enc_out)
# dec_out = self.seq2pred(dec_out.transpose(1, 2)).transpose(1, 2)
# print(dec_out.shape)
# De-Normalization from Non-stationary Transformer
dec_out = dec_out * \
(stdev[:, 0, :].unsqueeze(1).repeat(
1, L, 1))
dec_out = dec_out + \
(means[:, 0, :].unsqueeze(1).repeat(
1, L, 1))
return dec_out
def classification(self, x_enc, x_mark_enc):
# Normalization from Non-stationary Transformer
means = x_enc.mean(1, keepdim=True).detach()
x_enc = x_enc - means
stdev = torch.sqrt(
torch.var(x_enc, dim=1, keepdim=True, unbiased=False) + 1e-5)
x_enc /= stdev
# embedding
enc_out = self.enc_embedding(x_enc, None) # [B,T,C]
# adp = self.graph(torch.arange(self.num_nodes).to(self.device))
for i in range(self.layer):
enc_out = self.layer_norm(self.model[i](enc_out))
output = self.act(enc_out)
output = self.dropout(output)
# zero-out padding embeddings
output = output * x_mark_enc.unsqueeze(-1)
# (batch_size, seq_length * d_model)
output = output.reshape(output.shape[0], -1)
output = self.projection(output) # (batch_size, num_classes)
return output
def forward(self, x_enc, x_mark_enc, x_dec, x_mark_dec, mask=None):
if self.task_name == 'long_term_forecast' or self.task_name == 'short_term_forecast':
dec_out = self.forecast(x_enc, x_mark_enc, x_dec, x_mark_dec)
return dec_out[:, -self.pred_len:, :] # [B, L, D]
if self.task_name == 'imputation':
dec_out = self.imputation(
x_enc, x_mark_enc, x_dec, x_mark_dec, mask)
return dec_out # [B, L, D]
if self.task_name == 'anomaly_detection':
dec_out = self.anomaly_detection(x_enc)
return dec_out # [B, L, D]
if self.task_name == 'classification':
dec_out = self.classification(x_enc, x_mark_enc)
return dec_out # [B, N]
return None