From e7ab59610bf2a6c9d59a12e3e57796105ad6899d Mon Sep 17 00:00:00 2001 From: RealSuSeven <122971162+RealSuSeven@users.noreply.github.com> Date: Mon, 17 Mar 2025 14:53:18 +0800 Subject: [PATCH 01/11] Add files via upload --- .../Weather_script/PPDformer.sh | 134 ++++++++++++++++++ 1 file changed, 134 insertions(+) create mode 100644 scripts/long_term_forecast/Weather_script/PPDformer.sh diff --git a/scripts/long_term_forecast/Weather_script/PPDformer.sh b/scripts/long_term_forecast/Weather_script/PPDformer.sh new file mode 100644 index 000000000..bdf3727c9 --- /dev/null +++ b/scripts/long_term_forecast/Weather_script/PPDformer.sh @@ -0,0 +1,134 @@ +export CUDA_VISIBLE_DEVICES=1 + + +model_name=PPDformer +patchH=2 +patchW=4 +strideH=1 +strideW=2 +python -u run.py \ + --task_name long_term_forecast \ + --is_training 1 \ + --root_path ./dataset/weather/ \ + --data_path weather.csv \ + --model_id weather_96_96 \ + --model $model_name \ + --data custom \ + --features M \ + --seq_len 96 \ + --label_len 48 \ + --pred_len 96 \ + --e_layers 3 \ + --d_layers 1 \ + --factor 3 \ + --enc_in 4 \ + --dec_in 21 \ + --c_out 21 \ + --des 'Exp' \ + --d_model 512\ + --d_ff 512\ + --itr 1\ + --patchH $patchH \ + --top_k 5 \ + --attention 0 \ + --normal 1 \ + --strideH $strideH\ + --strideW $strideW\ + --dropout 0.1 \ + --patchW $patchW + +python -u run.py \ + --task_name long_term_forecast \ + --is_training 1 \ + --root_path ./dataset/weather/ \ + --data_path weather.csv \ + --model_id weather_96_192 \ + --model $model_name \ + --data custom \ + --features M \ + --seq_len 96 \ + --label_len 48 \ + --pred_len 192 \ + --e_layers 3 \ + --d_layers 1 \ + --factor 3 \ + --enc_in 4 \ + --dec_in 21 \ + --c_out 21 \ + --des 'Exp' \ + --d_model 512\ + --d_ff 512\ + --itr 1\ + --patchH $patchH \ + --top_k 5 \ + --attention 0 \ + --normal 1 \ + --dropout 0.1 \ + --strideH $strideH\ + --strideW $strideW\ + --patchW $patchW + + +python -u run.py \ + --task_name long_term_forecast \ + --is_training 1 \ + --root_path ./dataset/weather/ \ + --data_path weather.csv \ + --model_id weather_96_336 \ + --model $model_name \ + --data custom \ + --features M \ + --seq_len 96 \ + --label_len 48 \ + --pred_len 336 \ + --e_layers 3 \ + --d_layers 1 \ + --factor 3 \ + --enc_in 4 \ + --dec_in 21 \ + --c_out 21 \ + --des 'Exp' \ + --d_model 512\ + --d_ff 512\ + --itr 1\ + --patchH $patchH \ + --top_k 5 \ + --attention 0 \ + --normal 0 \ + --dropout 0.1 \ + --strideH $strideH\ + --strideW $strideW\ + --patchW $patchW + + +python -u run.py \ + --task_name long_term_forecast \ + --is_training 1 \ + --root_path ./dataset/weather/ \ + --data_path weather.csv \ + --model_id weather_96_720 \ + --model $model_name \ + --data custom \ + --features M \ + --seq_len 96 \ + --label_len 48 \ + --pred_len 720 \ + --e_layers 3 \ + --d_layers 1 \ + --factor 3 \ + --enc_in 4 \ + --dec_in 21 \ + --c_out 21 \ + --des 'Exp' \ + --d_model 512\ + --d_ff 512\ + --itr 1\ + --patchH $patchH \ + --top_k 5 \ + --normal 0 \ + --attention 0 \ + --dropout 0.1 \ + --strideH $strideH\ + --strideW $strideW\ + --patchW $patchW + From b167b4504089511db1f419b62cd0f209b934cf5b Mon Sep 17 00:00:00 2001 From: RealSuSeven <122971162+RealSuSeven@users.noreply.github.com> Date: Mon, 17 Mar 2025 14:56:14 +0800 Subject: [PATCH 02/11] Add files via upload --- .../ETT_script/PPDformer_ETTm2.sh | 124 ++++++++++++++++++ 1 file changed, 124 insertions(+) create mode 100644 scripts/long_term_forecast/ETT_script/PPDformer_ETTm2.sh diff --git a/scripts/long_term_forecast/ETT_script/PPDformer_ETTm2.sh b/scripts/long_term_forecast/ETT_script/PPDformer_ETTm2.sh new file mode 100644 index 000000000..af581d0f2 --- /dev/null +++ b/scripts/long_term_forecast/ETT_script/PPDformer_ETTm2.sh @@ -0,0 +1,124 @@ +export CUDA_VISIBLE_DEVICES=2 + + +model_name=PPDformer +patchH=2 +# patchW=2 +strideH=1 +# strideW=4 +seq_len=96 + + +python -u run.py \ + --is_training 1 \ + --root_path ./dataset/ETT-small/ \ + --data_path ETTm2.csv \ + --model_id ETTm2_96_96 \ + --model $model_name \ + --data ETTm2 \ + --features M \ + --seq_len 96 \ + --pred_len 96 \ + --e_layers 2 \ + --enc_in 7 \ + --dec_in 7 \ + --c_out 7 \ + --des 'Exp' \ + --d_model 512 \ + --d_ff 512 \ + --patchH $patchH \ + --top_k 5 \ + --normal 0 \ + --dropout 0.1 \ + --attention 1 \ + --patchW 2\ + --strideH $strideH\ + --strideW 2\ + --itr 1 + +python -u run.py \ + --is_training 1 \ + --root_path ./dataset/ETT-small/ \ + --data_path ETTm2.csv \ + --model_id ETTm2_96_336 \ + --model $model_name \ + --data ETTm2 \ + --features M \ + --seq_len 96 \ + --pred_len 336 \ + --e_layers 3 \ + --factor 3 \ + --enc_in 7 \ + --dec_in 7 \ + --c_out 7 \ + --des 'Exp' \ + --d_model 512 \ + --d_ff 512 \ + --patchH $patchH \ + --top_k 5 \ + --attention 1 \ + --patchW 4\ + --normal 1 \ + --dropout 0.0 \ + --strideH $strideH\ + --strideW 4\ + --n_heads 8 \ + --itr 1 + +python -u run.py \ + --is_training 1 \ + --root_path ./dataset/ETT-small/ \ + --data_path ETTm2.csv \ + --model_id ETTm2_96_336 \ + --model $model_name \ + --data ETTm2 \ + --features M \ + --seq_len 96 \ + --pred_len 336 \ + --e_layers 1 \ + --factor 1 \ + --enc_in 7 \ + --dec_in 7 \ + --c_out 7 \ + --des 'Exp' \ + --d_model 512 \ + --d_ff 512 \ + --patchH $patchH \ + --top_k 5 \ + --attention 1 \ + --patchW 16\ + --normal 1 \ + --dropout 0.0 \ + --strideH $strideH\ + --strideW 8\ + --n_heads 4 \ + --itr 1 + +python -u run.py \ + --is_training 1 \ + --root_path ./dataset/ETT-small/ \ + --data_path ETTm2.csv \ + --model_id ETTm2_96_336 \ + --model $model_name \ + --data ETTm2 \ + --features M \ + --seq_len 96 \ + --pred_len 336 \ + --e_layers 2 \ + --factor 1 \ + --enc_in 7 \ + --dec_in 7 \ + --c_out 7 \ + --des 'Exp' \ + --d_model 512 \ + --d_ff 512 \ + --patchH $patchH \ + --top_k 5 \ + --attention 1 \ + --patchW 4\ + --normal 1 \ + --dropout 0.0 \ + --strideH $strideH\ + --strideW 4\ + --n_heads 16 \ + --itr 1 \ No newline at end of file From 881a4c623bc4ba0f8c3ff65e9e8b03dd8d4c4056 Mon Sep 17 00:00:00 2001 From: RealSuSeven <122971162+RealSuSeven@users.noreply.github.com> Date: Mon, 17 Mar 2025 14:57:01 +0800 Subject: [PATCH 03/11] Add files via upload --- .../ECL_script/PPDformer.sh | 144 ++++++++++++++++++ 1 file changed, 144 insertions(+) create mode 100644 scripts/long_term_forecast/ECL_script/PPDformer.sh diff --git a/scripts/long_term_forecast/ECL_script/PPDformer.sh b/scripts/long_term_forecast/ECL_script/PPDformer.sh new file mode 100644 index 000000000..ff963b15f --- /dev/null +++ b/scripts/long_term_forecast/ECL_script/PPDformer.sh @@ -0,0 +1,144 @@ +export CUDA_VISIBLE_DEVICES=0 + + +model_name=PPDformer +patchH=2 +patchW=4 +strideH=1 +strideW=1 +python -u run.py \ + --task_name long_term_forecast \ + --is_training 1 \ + --root_path ./dataset/electricity/ \ + --data_path electricity.csv \ + --model_id electricity_96_96 \ + --model $model_name \ + --data custom \ + --features M \ + --seq_len 96 \ + --label_len 48 \ + --pred_len 96 \ + --e_layers 3 \ + --d_layers 1 \ + --factor 3 \ + --enc_in 321 \ + --dec_in 321 \ + --c_out 321 \ + --des 'Exp' \ + --d_model 512\ + --d_ff 512\ + --batch_size 16 \ + --learning_rate 0.0005 \ + --itr 1\ + --patchH $patchH \ + --top_k 5 \ + --attention 0 \ + --strideH $strideH\ + --strideW $strideW\ + --normal 1 \ + --dropout 0.1 \ + --patchW $patchW + +python -u run.py \ + --task_name long_term_forecast \ + --is_training 1 \ + --root_path ./dataset/electricity/ \ + --data_path electricity.csv \ + --model_id electricity_96_192 \ + --model $model_name \ + --data custom \ + --features M \ + --seq_len 96 \ + --label_len 48 \ + --pred_len 192 \ + --e_layers 3 \ + --d_layers 1 \ + --factor 3 \ + --enc_in 321 \ + --dec_in 321 \ + --c_out 321 \ + --des 'Exp' \ + --d_model 512\ + --d_ff 512\ + --batch_size 16 \ + --learning_rate 0.0005 \ + --itr 1\ + --normal 1 \ + --dropout 0.1 \ + --attention 0 \ + --strideH $strideH\ + --strideW $strideW\ + --patchH $patchH \ + --top_k 5 \ + --patchW $patchW + + +python -u run.py \ + --task_name long_term_forecast \ + --is_training 1 \ + --root_path ./dataset/electricity/ \ + --data_path electricity.csv \ + --model_id electricity_96_336 \ + --model $model_name \ + --data custom \ + --features M \ + --seq_len 96 \ + --label_len 48 \ + --pred_len 336 \ + --e_layers 3 \ + --d_layers 1 \ + --factor 3 \ + --enc_in 321 \ + --dec_in 321 \ + --c_out 321 \ + --batch_size 16 \ + --learning_rate 0.0005 \ + --des 'Exp' \ + --d_model 512\ + --d_ff 512\ + --itr 1\ + --attention 0 \ + --normal 1 \ + --dropout 0.1 \ + --strideH $strideH\ + --strideW $strideW\ + --patchH $patchH \ + --top_k 5 \ + --patchW $patchW + + +python -u run.py \ + --task_name long_term_forecast \ + --is_training 1 \ + --root_path ./dataset/electricity/ \ + --data_path electricity.csv \ + --model_id electricity_96_720 \ + --model $model_name \ + --data custom \ + --features M \ + --normal 0 \ + --dropout 0.1 \ + --seq_len 96 \ + --label_len 48 \ + --pred_len 720 \ + --e_layers 3 \ + --d_layers 1 \ + --factor 3 \ + --batch_size 16 \ + --learning_rate 0.0005 \ + --enc_in 321 \ + --dec_in 321 \ + --c_out 321 \ + --des 'Exp' \ + --d_model 512\ + --d_ff 512\ + --itr 1\ + --normal 1 \ + --dropout 0.1 \ + --attention 0 \ + --strideH $strideH\ + --strideW $strideW\ + --patchH $patchH \ + --top_k 5 \ + --patchW $patchW + From 9bceb3d5fad620287074826c8e43d2bcb4ac9220 Mon Sep 17 00:00:00 2001 From: RealSuSeven <122971162+RealSuSeven@users.noreply.github.com> Date: Mon, 17 Mar 2025 14:58:15 +0800 Subject: [PATCH 04/11] Add files via upload --- models/PPDformer.py | 440 ++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 440 insertions(+) create mode 100644 models/PPDformer.py diff --git a/models/PPDformer.py b/models/PPDformer.py new file mode 100644 index 000000000..5826e800c --- /dev/null +++ b/models/PPDformer.py @@ -0,0 +1,440 @@ +import torch +import torch.nn as nn +import torch.nn.functional as F +from layers.Transformer_EncDec import Encoder, EncoderLayer +from layers.SelfAttention_Family import FullAttention, AttentionLayer +from layers.Embed import DataEmbedding_inverted +import numpy as np + +from scipy.signal import find_peaks +from scipy.signal import periodogram +import pywt +from scipy.signal import hilbert +import matplotlib.pyplot as plt +from collections import defaultdict +import torch +import torch.nn as nn +import numpy as np +import pandas as pd +from sklearn.preprocessing import StandardScaler +from scipy.signal import find_peaks +import pandas as pd +import torch +import torch.nn as nn +from sklearn.preprocessing import StandardScaler +import matplotlib.pyplot as plt + + +class Preprocessor(nn.Module): + def __init__(self, window_size=5): + super().__init__() + self.window_size = window_size + + def moving_average(self, signal): + pad_signal = torch.nn.functional.pad(signal, (0, 0, self.window_size - 1, 0), mode='reflect') + cumsum_signal = torch.cumsum(pad_signal, dim=1) + ma_signal = (cumsum_signal[:, self.window_size:] - cumsum_signal[:, :-self.window_size]) / self.window_size + return ma_signal + + # 设置频率阈值函数 + def apply_fft_denoising(self, signal, threshold_factor=0.1): + # 对每个通道进行傅里叶变换 + fft_result = torch.fft.fft(signal, dim=1) + + # 计算阈值 + threshold = threshold_factor * torch.max(torch.abs(fft_result), dim=1, keepdim=True)[0] + + # 过滤噪声 + fft_result[torch.abs(fft_result) < threshold] = 0 + + # 对每个通道进行逆傅里叶变换 + denoised_signal = torch.fft.ifft(fft_result, dim=1) + + return denoised_signal.real + + def forward(self, original_signal): + # ma_signal = self.moving_average(original_signal) + # ma_signal = torch.nn.functional.pad(ma_signal, (0, 0, self.window_size - 1, 0), mode='reflect')[:, + # :original_signal.size(1)] + ma_signal = self.apply_fft_denoising(original_signal) + reference_noise = original_signal - ma_signal + return reference_noise + + +class AdaptiveNoiseCanceller(nn.Module): + def __init__(self, filter_order, mu, batch, c_in): + super().__init__() + # self.device = device + self.filter_order = filter_order + self.mu = mu + #self.mu = nn.Parameter(torch.tensor(0.01)) + # self.weights = nn.Parameter(torch.zeros((32, filter_order, 21))) + # self.weights = torch.zeros((32, 10, 21), device=device) + # self.weights = torch.zeros((32, 10, 21), device=self.device) + self.weights = torch.randn((batch, filter_order, c_in)) * 0.1 + + def forward(self, noisy_signal, reference_signal): + B, L, C = noisy_signal.size() + denoised_signal = torch.zeros_like(noisy_signal) + # weights = self.weights.clone() # 复制权重以进行显式更新 + + for n in range(self.filter_order, L): + if n < self.filter_order: + continue + x = reference_signal[:, n - self.filter_order:n, :].flip(dims=[1]).to(noisy_signal.device) + weights = self.weights.to(x.device) + y = torch.sum(weights * x, dim=1) + + e = noisy_signal[:, n, :] - y + + probabilities = 2 * self.mu * e + + probabilities = probabilities.unsqueeze(1).expand(-1, 10, -1).to(x.device) + + weights.data += probabilities * x + denoised_signal[:, n, :] = e + + # 用原始数据中的点代替前几个数据点 + denoised_signal[:, :self.filter_order, :] = noisy_signal[:, :self.filter_order, :] + + + return denoised_signal + + +class NoiseCancellation(nn.Module): + def __init__(self, filter_order=5, mu=0.01, window_size=5, batch_size = 32, c_in = 21, + device=torch.device('cuda' if torch.cuda.is_available() else 'cpu')): + super().__init__() + self.device = device + self.filter_order = filter_order + # mu_tensor = torch.full((B, C), mu) + self.window_size = window_size + self.preprocessor = Preprocessor(window_size) + self.anc = AdaptiveNoiseCanceller(filter_order, mu, batch_size, c_in) + + def forward(self, x): + reference_noise = self.preprocessor(x) + + denoised_data = self.anc(x, reference_noise) + + return denoised_data + +##########full############# +def optimized_linear_interpolation_padding(x, seq_len, period,mode='linear'): + + B,C,T = x.shape + # 计算需要的总长度，使其为周期的整数倍 + if seq_len % period != 0: + length = ((seq_len // period) + 1) * period + + else: + length = seq_len + return x,length # 如果已经是周期的整数倍，直接返回 + # 获取最后一个周期的数据 + last_period_start = seq_len - (seq_len % period) + + remaining_length = seq_len - last_period_start + + last_period_data = x + + target = length + + if mode =='linear': + # 目标长度为一个完整的周期长度 + interpolated_data = F.interpolate(last_period_data, size=target, mode='linear', align_corners=False) + elif mode =='nearest': + interpolated_data = F.interpolate(last_period_data, size=target, mode='nearest') + else: + last_period_data_reshaped = last_period_data.unsqueeze(1) # 添加一个维度作为channel维 + if mode =='bicubic': + # 使用bicubic插值将图像的大小调整到48x48 + interpolated_data = F.interpolate(last_period_data_reshaped, size=(C, target), mode='bicubic', align_corners=True) + elif mode =='bilinear': + # 使用bicubic插值将图像的大小调整到48x48 + interpolated_data = F.interpolate(last_period_data_reshaped, size=(C, target), mode='bilinear', align_corners=True) + elif mode =='nearest2': + # 使用bicubic插值将图像的大小调整到48x48 + interpolated_data = F.interpolate(last_period_data_reshaped, size=(C, target), mode='nearest') + interpolated_data = interpolated_data.squeeze(1) + + # 将插值后的数据重塑回原始维度 [batch_size, C, new_seq_len] + + return interpolated_data,length + + +class ComplexWeightedSumModel(nn.Module): + def __init__(self, channels, d_model, normal, dropout): + super(ComplexWeightedSumModel, self).__init__() + + # 权重参数 + self.weights_inner = nn.Parameter(torch.randn(1, channels, 1)) + self.weights_whole = nn.Parameter(torch.randn(1, channels, 1)) + + # LayerNorm + self.norm_inner = nn.LayerNorm([channels, d_model]) + self.norm_whole = nn.LayerNorm([channels, d_model]) + + # MLP层和Dropout + self.linear1 = nn.Linear(d_model, d_model*2) + self.linear2 = nn.Linear(d_model*2, d_model*2) + self.linear3 = nn.Linear(d_model*2, d_model) + + self.dropout1 = nn.Dropout(dropout) + self.dropout2 = nn.Dropout(dropout) + self.normal = normal + self.drop = dropout + + def forward(self, inner, whole): + # 应用LayerNorm + if self.normal: + + inner = self.norm_inner(inner) + whole = self.norm_whole(whole) + + # 应用权重 + + weighted_inner = self.weights_inner * inner + weighted_whole = self.weights_whole * whole + + # 相加 + sum_result = weighted_inner + weighted_whole + + # MLP处理 + mlp_output = F.relu(self.linear1(sum_result)) + mlp_output = self.dropout1(mlp_output) + mlp_output = F.relu(self.linear2(mlp_output)) + mlp_output = self.dropout2(mlp_output) + mlp_output = F.gelu(self.linear3(mlp_output)) + + return mlp_output + + +class ComplexTensorProcessor(nn.Module): + def __init__(self, d_model,configs): + super().__init__() + # 使用不同大小的卷积核进行特征提取 + self.encoder1 = Encoder( + [ + EncoderLayer( + AttentionLayer( + FullAttention(False, configs.factor, attention_dropout=configs.dropout, + output_attention=configs.output_attention), configs.d_model, configs.n_heads), + configs.d_model, + configs.d_ff, + dropout=configs.dropout, + activation=configs.activation + ) for l in range(configs.e_layers) + ], + norm_layer=torch.nn.LayerNorm(configs.d_model) + ) + self.conv1 = nn.Conv1d(d_model, d_model *2, kernel_size=3, padding=0) + self.point_conv = nn.Conv1d(d_model*2, d_model*2, kernel_size=1) + self.point = nn.Conv1d(d_model*2, d_model, kernel_size=1) + self.mlp = nn.Sequential( + nn.Linear(d_model, d_model), + # nn.ReLU(), + # nn.Linear(d_model*2, d_model*2), + nn.ReLU(), + nn.Linear(d_model, d_model) + + ) + self.layer_norm1 = nn.LayerNorm(d_model) + self.attention = configs.attention + + def forward(self, x): + + B,C,N,P = x.shape + x = x.reshape(B*C,N,P) + x = x.transpose(1, 2) # 交换N和T的维度 + x = self.conv1(x) + x = self.point_conv(x) + x = self.point(x) + # 注意力层 + if self.attention : + + x = x.transpose(1, 2) # 为自注意力准备维度 + x = self.layer_norm1(x) # 添加层归一化 + x, attns = self.encoder1(x, attn_mask=None) + x = x.transpose(1, 2) # 恢复维度 + x = F.adaptive_avg_pool1d(x, 1) + x = x.squeeze(2) + x = self.mlp(x) + x = x.reshape(B,C,P) + return x +def STFT_for_Period(x, k=3, n_fft=16, hop_length=8, win_length=16, window='hann'): + # Window setup based on user selection + if window == 'hann': + window_tensor = torch.hann_window(win_length, periodic=True).to(x.device) + elif window == 'hamming': + window_tensor = torch.hamming_window(win_length, periodic=True).to(x.device) + else: + raise ValueError(f"Unsupported window type: {window}") + + B, C, T = x.shape + + stft_results = [] + + # Perform STFT for each channel separately + for c in range(C): + single_channel_data = x[:, c, :] + stft_result = torch.stft(single_channel_data, n_fft=n_fft, hop_length=hop_length, + win_length=win_length, window=window_tensor, return_complex=True) + stft_results.append(stft_result) + + stft_results = torch.stack(stft_results, dim = 1) + xf_magnitude = torch.abs(stft_results).mean(dim=0) + + # Calculate frequency list + + frequency_list = xf_magnitude.permute(0,2,1) + + + frequency_list[:,:, 0] = 0 # Eliminate the DC component + + k_amplitude_all = [] + k_index_all = [] + + for c in range(C): # Iterate over C dimension + top_values = [] + for t in range(frequency_list.shape[1]): # Iterate over T dimension + _, top_list = torch.topk(frequency_list[c, t, :], k) + if top_list[0] == 1: + if top_list[1] < 1: + chosen_index = top_list[2] + else: + chosen_index = top_list[1]# 如果第一位为1，则取第二位 + else: + chosen_index = top_list[0] # 如果第一位不为1，则取第一位 + top_values.append(chosen_index) + # top_values.append(top_list) + + top_values = torch.tensor(top_values) + + k_amplitude, k_index = torch.topk(top_values.flatten(), 1) # Flatten to get top k values across all time bins + k_amplitude_all.append(k_amplitude) + k_index_all.append(k_index) + + + + k_amplitude_all = torch.stack(k_amplitude_all) + k_index_all = torch.stack(k_index_all) + + period_list = T // k_amplitude_all + return period_list + +class Model(nn.Module): + + def __init__(self, configs): + super(Model, self).__init__() + self.task_name = configs.task_name + self.seq_len = configs.seq_len + self.pred_len = configs.pred_len + self.output_attention = configs.output_attention + self.k = configs.top_k + self.patchH =configs.patchH + self.patchW =configs.patchW + self.strideH = configs.strideH + self.strideW = configs.strideW + self.B = configs.batch_size + # Embedding + self.enc_embedding = DataEmbedding_inverted(configs.seq_len, configs.d_model, configs.embed, configs.freq, + configs.dropout) + # Encoder + + + self.encoder = Encoder( + [ + EncoderLayer( + AttentionLayer( + FullAttention(False, configs.factor, attention_dropout=configs.dropout, + output_attention=configs.output_attention), configs.d_model, configs.n_heads), + configs.d_model, + configs.d_ff, + dropout=configs.dropout, + activation=configs.activation + ) for l in range(configs.e_layers) + ], + norm_layer=torch.nn.LayerNorm(configs.d_model) + ) + + self.value_embedding = nn.Linear(self.patchH*self.patchW, configs.d_model) + + self.CWSM = ComplexWeightedSumModel(configs.dec_in,configs.d_model,configs.normal,configs.dropout) + self.projection = nn.Linear(configs.d_model, configs.pred_len, bias=True) + self.CTP = ComplexTensorProcessor(configs.d_model,configs) + self.noise_cancellation = NoiseCancellation(filter_order=10, mu=0.01, window_size=5, batch_size = self.B, c_in =configs.dec_in) + + # 对数据进行去噪处理 + + def forecast(self, x_enc, x_mark_enc, x_dec, x_mark_dec): + # 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 + B, T, C = x_enc.shape + # Embedding + whole_out = self.enc_embedding(x_enc, x_mark_enc) + whole_out, attns = self.encoder(whole_out, attn_mask=None) + whole_out = whole_out[:, :C, :] + + x_enc = self.noise_cancellation(x_enc) + + x = x_enc.permute(0,2,1) + period_list = STFT_for_Period(x) + + patch_size = [self.patchH,self.patchW ] + stride = (self.strideH,self.strideW) +# 创建周期到通道索引的映射 + period_to_channels = defaultdict(list) + for i in range(C): + period = period_list[i].item() + period_to_channels[period].append(i) + + CI = [] + + channel_order = [] + for period, channels in period_to_channels.items(): + # 提取所有具有相同周期的通道W + x_selected = x[:, channels, :] + # 执行周期性插值和填充 + + if period < self.patchW: + period = self.patchW + + x_padded, length = optimized_linear_interpolation_padding(x_selected, self.seq_len, period) + N_per = length // period + + + out = x_padded.reshape(B, len(channels), N_per, period).contiguous() + patches = out.unfold(2, patch_size[0], stride[0]).unfold(3, patch_size[1], stride[1]) + patches_reshape = patches.contiguous().view(B, len(channels), -1, patch_size[0] * patch_size[1]) + embedfor = torch.reshape(patches_reshape, (patches_reshape.shape[0]*len(channels) , patches_reshape.shape[2], patches_reshape.shape[3])) + enc_out = self.value_embedding(embedfor) + enc_out = torch.reshape(enc_out, (B, len(channels), enc_out.shape[-2], enc_out.shape[-1])) + CI_out = self.CTP(enc_out) + + CI.append(CI_out) + channel_order.extend(channels) + + + dec_out = torch.cat(CI, dim=1) + original_order_index = torch.argsort(torch.tensor(channel_order)) + dec_out = dec_out[:, original_order_index, :] # 重新排列通道 + + dec_out = self.CWSM(dec_out,whole_out) + + dec_out = self.projection(dec_out).permute(0, 2, 1) + + # 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 + + + 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] + + return None From 541415b614bdd41cdad35f15fe6f20ac90ff0ad2 Mon Sep 17 00:00:00 2001 From: RealSuSeven <122971162+RealSuSeven@users.noreply.github.com> Date: Mon, 17 Mar 2025 15:00:24 +0800 Subject: [PATCH 05/11] Update exp_basic.py Add PPDformer --- exp/exp_basic.py | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/exp/exp_basic.py b/exp/exp_basic.py index 2f87908fd..3871f0bf3 100644 --- a/exp/exp_basic.py +++ b/exp/exp_basic.py @@ -3,7 +3,7 @@ from models import Autoformer, Transformer, TimesNet, Nonstationary_Transformer, DLinear, FEDformer, \ Informer, LightTS, Reformer, ETSformer, Pyraformer, PatchTST, MICN, Crossformer, FiLM, iTransformer, \ Koopa, TiDE, FreTS, TimeMixer, TSMixer, SegRNN, MambaSimple, TemporalFusionTransformer, SCINet, PAttn, TimeXer, \ - WPMixer, MultiPatchFormer + WPMixer, MultiPatchFormer,PPDformer class Exp_Basic(object): @@ -39,6 +39,7 @@ def __init__(self, args): 'TimeXer': TimeXer, 'WPMixer': WPMixer, 'MultiPatchFormer': MultiPatchFormer + 'PPDformer': PPDformer } if args.model == 'Mamba': print('Please make sure you have successfully installed mamba_ssm') From fd488c3c533306798451530a8d12feee98f3ef7f Mon Sep 17 00:00:00 2001 From: RealSuSeven <122971162+RealSuSeven@users.noreply.github.com> Date: Mon, 17 Mar 2025 15:26:38 +0800 Subject: [PATCH 06/11] Add PPDformer --- run.py | 8 ++++++++ 1 file changed, 8 insertions(+) diff --git a/run.py b/run.py index 793a53ac2..627141078 100644 --- a/run.py +++ b/run.py @@ -72,6 +72,7 @@ parser.add_argument('--dropout', type=float, default=0.1, help='dropout') parser.add_argument('--embed', type=str, default='timeF', help='time features encoding, options:[timeF, fixed, learned]') + parser.add_argument('--activation', type=str, default='gelu', help='activation') parser.add_argument('--channel_independence', type=int, default=1, help='0: channel dependence 1: channel independence for FreTS model') @@ -140,6 +141,13 @@ # TimeXer parser.add_argument('--patch_len', type=int, default=16, help='patch length') + # PPDformer + parser.add_argument('--patchH', type=int, default=2, help='patch length of H') + parser.add_argument('--patchW', type=int, default=8, help='patch length of W') + parser.add_argument('--strideH', type=int, default=2, help='stride length of H') + parser.add_argument('--strideW', type=int, default=8, help='stride length of W') + parser.add_argument('--normal', type=int, default=1, help='use normal for patches') + parser.add_argument('--attention', type=int, default=0, help='use attention for patches') args = parser.parse_args() if torch.cuda.is_available() and args.use_gpu: args.device = torch.device('cuda:{}'.format(args.gpu)) From 93075469eba4b5ab1f002e21d68e4d51a745cd1c Mon Sep 17 00:00:00 2001 From: RealSuSeven <122971162+RealSuSeven@users.noreply.github.com> Date: Mon, 17 Mar 2025 15:50:11 +0800 Subject: [PATCH 07/11] Update PPDformer.py --- models/PPDformer.py | 120 +++++++------------------------------------- 1 file changed, 19 insertions(+), 101 deletions(-) diff --git a/models/PPDformer.py b/models/PPDformer.py index 5826e800c..53a7686d2 100644 --- a/models/PPDformer.py +++ b/models/PPDformer.py @@ -36,20 +36,11 @@ def moving_average(self, signal): ma_signal = (cumsum_signal[:, self.window_size:] - cumsum_signal[:, :-self.window_size]) / self.window_size return ma_signal - # 设置频率阈值函数 def apply_fft_denoising(self, signal, threshold_factor=0.1): - # 对每个通道进行傅里叶变换 fft_result = torch.fft.fft(signal, dim=1) - - # 计算阈值 threshold = threshold_factor * torch.max(torch.abs(fft_result), dim=1, keepdim=True)[0] - - # 过滤噪声 fft_result[torch.abs(fft_result) < threshold] = 0 - - # 对每个通道进行逆傅里叶变换 denoised_signal = torch.fft.ifft(fft_result, dim=1) - return denoised_signal.real def forward(self, original_signal): @@ -64,19 +55,13 @@ def forward(self, original_signal): class AdaptiveNoiseCanceller(nn.Module): def __init__(self, filter_order, mu, batch, c_in): super().__init__() - # self.device = device self.filter_order = filter_order self.mu = mu - #self.mu = nn.Parameter(torch.tensor(0.01)) - # self.weights = nn.Parameter(torch.zeros((32, filter_order, 21))) - # self.weights = torch.zeros((32, 10, 21), device=device) - # self.weights = torch.zeros((32, 10, 21), device=self.device) self.weights = torch.randn((batch, filter_order, c_in)) * 0.1 def forward(self, noisy_signal, reference_signal): B, L, C = noisy_signal.size() denoised_signal = torch.zeros_like(noisy_signal) - # weights = self.weights.clone() # 复制权重以进行显式更新 for n in range(self.filter_order, L): if n < self.filter_order: @@ -84,20 +69,13 @@ def forward(self, noisy_signal, reference_signal): x = reference_signal[:, n - self.filter_order:n, :].flip(dims=[1]).to(noisy_signal.device) weights = self.weights.to(x.device) y = torch.sum(weights * x, dim=1) - e = noisy_signal[:, n, :] - y - probabilities = 2 * self.mu * e - probabilities = probabilities.unsqueeze(1).expand(-1, 10, -1).to(x.device) - weights.data += probabilities * x denoised_signal[:, n, :] = e - # 用原始数据中的点代替前几个数据点 denoised_signal[:, :self.filter_order, :] = noisy_signal[:, :self.filter_order, :] - - return denoised_signal @@ -107,99 +85,73 @@ def __init__(self, filter_order=5, mu=0.01, window_size=5, batch_size = 32, c_in super().__init__() self.device = device self.filter_order = filter_order - # mu_tensor = torch.full((B, C), mu) self.window_size = window_size self.preprocessor = Preprocessor(window_size) self.anc = AdaptiveNoiseCanceller(filter_order, mu, batch_size, c_in) def forward(self, x): reference_noise = self.preprocessor(x) - denoised_data = self.anc(x, reference_noise) - return denoised_data -##########full############# def optimized_linear_interpolation_padding(x, seq_len, period,mode='linear'): B,C,T = x.shape - # 计算需要的总长度，使其为周期的整数倍 if seq_len % period != 0: length = ((seq_len // period) + 1) * period - else: length = seq_len - return x,length # 如果已经是周期的整数倍，直接返回 - # 获取最后一个周期的数据 + return x,length last_period_start = seq_len - (seq_len % period) - remaining_length = seq_len - last_period_start - last_period_data = x - target = length - if mode =='linear': - # 目标长度为一个完整的周期长度 interpolated_data = F.interpolate(last_period_data, size=target, mode='linear', align_corners=False) elif mode =='nearest': interpolated_data = F.interpolate(last_period_data, size=target, mode='nearest') else: - last_period_data_reshaped = last_period_data.unsqueeze(1) # 添加一个维度作为channel维 + last_period_data_reshaped = last_period_data.unsqueeze(1) if mode =='bicubic': - # 使用bicubic插值将图像的大小调整到48x48 interpolated_data = F.interpolate(last_period_data_reshaped, size=(C, target), mode='bicubic', align_corners=True) elif mode =='bilinear': - # 使用bicubic插值将图像的大小调整到48x48 interpolated_data = F.interpolate(last_period_data_reshaped, size=(C, target), mode='bilinear', align_corners=True) elif mode =='nearest2': - # 使用bicubic插值将图像的大小调整到48x48 interpolated_data = F.interpolate(last_period_data_reshaped, size=(C, target), mode='nearest') interpolated_data = interpolated_data.squeeze(1) - - # 将插值后的数据重塑回原始维度 [batch_size, C, new_seq_len] - return interpolated_data,length class ComplexWeightedSumModel(nn.Module): def __init__(self, channels, d_model, normal, dropout): super(ComplexWeightedSumModel, self).__init__() - - # 权重参数 + self.weights_inner = nn.Parameter(torch.randn(1, channels, 1)) self.weights_whole = nn.Parameter(torch.randn(1, channels, 1)) - # LayerNorm self.norm_inner = nn.LayerNorm([channels, d_model]) self.norm_whole = nn.LayerNorm([channels, d_model]) - # MLP层和Dropout self.linear1 = nn.Linear(d_model, d_model*2) self.linear2 = nn.Linear(d_model*2, d_model*2) self.linear3 = nn.Linear(d_model*2, d_model) self.dropout1 = nn.Dropout(dropout) self.dropout2 = nn.Dropout(dropout) + self.normal = normal self.drop = dropout def forward(self, inner, whole): - # 应用LayerNorm if self.normal: - inner = self.norm_inner(inner) whole = self.norm_whole(whole) - # 应用权重 - weighted_inner = self.weights_inner * inner weighted_whole = self.weights_whole * whole - # 相加 sum_result = weighted_inner + weighted_whole - # MLP处理 mlp_output = F.relu(self.linear1(sum_result)) mlp_output = self.dropout1(mlp_output) mlp_output = F.relu(self.linear2(mlp_output)) @@ -212,7 +164,6 @@ def forward(self, inner, whole): class ComplexTensorProcessor(nn.Module): def __init__(self, d_model,configs): super().__init__() - # 使用不同大小的卷积核进行特征提取 self.encoder1 = Encoder( [ EncoderLayer( @@ -245,17 +196,17 @@ def forward(self, x): B,C,N,P = x.shape x = x.reshape(B*C,N,P) - x = x.transpose(1, 2) # 交换N和T的维度 + x = x.transpose(1, 2) x = self.conv1(x) x = self.point_conv(x) x = self.point(x) - # 注意力层 + if self.attention : - - x = x.transpose(1, 2) # 为自注意力准备维度 - x = self.layer_norm1(x) # 添加层归一化 + x = x.transpose(1, 2) + x = self.layer_norm1(x) x, attns = self.encoder1(x, attn_mask=None) - x = x.transpose(1, 2) # 恢复维度 + x = x.transpose(1, 2) + x = F.adaptive_avg_pool1d(x, 1) x = x.squeeze(2) x = self.mlp(x) @@ -271,10 +222,8 @@ def STFT_for_Period(x, k=3, n_fft=16, hop_length=8, win_length=16, window='hann' raise ValueError(f"Unsupported window type: {window}") B, C, T = x.shape - stft_results = [] - # Perform STFT for each channel separately for c in range(C): single_channel_data = x[:, c, :] stft_result = torch.stft(single_channel_data, n_fft=n_fft, hop_length=hop_length, @@ -283,42 +232,29 @@ def STFT_for_Period(x, k=3, n_fft=16, hop_length=8, win_length=16, window='hann' stft_results = torch.stack(stft_results, dim = 1) xf_magnitude = torch.abs(stft_results).mean(dim=0) - - # Calculate frequency list - frequency_list = xf_magnitude.permute(0,2,1) - - - frequency_list[:,:, 0] = 0 # Eliminate the DC component - + frequency_list[:,:, 0] = 0 k_amplitude_all = [] k_index_all = [] - - for c in range(C): # Iterate over C dimension + for c in range(C): top_values = [] - for t in range(frequency_list.shape[1]): # Iterate over T dimension + for t in range(frequency_list.shape[1]): _, top_list = torch.topk(frequency_list[c, t, :], k) if top_list[0] == 1: if top_list[1] < 1: chosen_index = top_list[2] else: - chosen_index = top_list[1]# 如果第一位为1，则取第二位 + chosen_index = top_list[1] else: - chosen_index = top_list[0] # 如果第一位不为1，则取第一位 + chosen_index = top_list[0] top_values.append(chosen_index) - # top_values.append(top_list) - top_values = torch.tensor(top_values) - k_amplitude, k_index = torch.topk(top_values.flatten(), 1) # Flatten to get top k values across all time bins k_amplitude_all.append(k_amplitude) k_index_all.append(k_index) - - - + k_amplitude_all = torch.stack(k_amplitude_all) k_index_all = torch.stack(k_index_all) - period_list = T // k_amplitude_all return period_list @@ -339,9 +275,6 @@ def __init__(self, configs): # Embedding self.enc_embedding = DataEmbedding_inverted(configs.seq_len, configs.d_model, configs.embed, configs.freq, configs.dropout) - # Encoder - - self.encoder = Encoder( [ EncoderLayer( @@ -363,8 +296,6 @@ def __init__(self, configs): self.projection = nn.Linear(configs.d_model, configs.pred_len, bias=True) self.CTP = ComplexTensorProcessor(configs.d_model,configs) self.noise_cancellation = NoiseCancellation(filter_order=10, mu=0.01, window_size=5, batch_size = self.B, c_in =configs.dec_in) - - # 对数据进行去噪处理 def forecast(self, x_enc, x_mark_enc, x_dec, x_mark_dec): # Normalization from Non-stationary Transformer @@ -373,19 +304,16 @@ def forecast(self, x_enc, x_mark_enc, x_dec, x_mark_dec): stdev = torch.sqrt(torch.var(x_enc, dim=1, keepdim=True, unbiased=False) + 1e-5) x_enc /= stdev B, T, C = x_enc.shape - # Embedding + whole_out = self.enc_embedding(x_enc, x_mark_enc) whole_out, attns = self.encoder(whole_out, attn_mask=None) whole_out = whole_out[:, :C, :] x_enc = self.noise_cancellation(x_enc) - x = x_enc.permute(0,2,1) - period_list = STFT_for_Period(x) - + period_list = STFT_for_Period(x) patch_size = [self.patchH,self.patchW ] stride = (self.strideH,self.strideW) -# 创建周期到通道索引的映射 period_to_channels = defaultdict(list) for i in range(C): period = period_list[i].item() @@ -395,17 +323,11 @@ def forecast(self, x_enc, x_mark_enc, x_dec, x_mark_dec): channel_order = [] for period, channels in period_to_channels.items(): - # 提取所有具有相同周期的通道W x_selected = x[:, channels, :] - # 执行周期性插值和填充 - if period < self.patchW: period = self.patchW - x_padded, length = optimized_linear_interpolation_padding(x_selected, self.seq_len, period) N_per = length // period - - out = x_padded.reshape(B, len(channels), N_per, period).contiguous() patches = out.unfold(2, patch_size[0], stride[0]).unfold(3, patch_size[1], stride[1]) patches_reshape = patches.contiguous().view(B, len(channels), -1, patch_size[0] * patch_size[1]) @@ -413,17 +335,13 @@ def forecast(self, x_enc, x_mark_enc, x_dec, x_mark_dec): enc_out = self.value_embedding(embedfor) enc_out = torch.reshape(enc_out, (B, len(channels), enc_out.shape[-2], enc_out.shape[-1])) CI_out = self.CTP(enc_out) - CI.append(CI_out) channel_order.extend(channels) - dec_out = torch.cat(CI, dim=1) original_order_index = torch.argsort(torch.tensor(channel_order)) - dec_out = dec_out[:, original_order_index, :] # 重新排列通道 - + dec_out = dec_out[:, original_order_index, :] dec_out = self.CWSM(dec_out,whole_out) - dec_out = self.projection(dec_out).permute(0, 2, 1) # De-Normalization from Non-stationary Transformer From 7e7267b69ff4372ccb019a02f827ec630f826321 Mon Sep 17 00:00:00 2001 From: RealSuSeven <122971162+RealSuSeven@users.noreply.github.com> Date: Mon, 17 Mar 2025 15:52:16 +0800 Subject: [PATCH 08/11] Update PPDformer.py --- models/PPDformer.py | 8 +------- 1 file changed, 1 insertion(+), 7 deletions(-) diff --git a/models/PPDformer.py b/models/PPDformer.py index 53a7686d2..e9068884a 100644 --- a/models/PPDformer.py +++ b/models/PPDformer.py @@ -44,9 +44,6 @@ def apply_fft_denoising(self, signal, threshold_factor=0.1): return denoised_signal.real def forward(self, original_signal): - # ma_signal = self.moving_average(original_signal) - # ma_signal = torch.nn.functional.pad(ma_signal, (0, 0, self.window_size - 1, 0), mode='reflect')[:, - # :original_signal.size(1)] ma_signal = self.apply_fft_denoising(original_signal) reference_noise = original_signal - ma_signal return reference_noise @@ -183,8 +180,6 @@ def __init__(self, d_model,configs): self.point = nn.Conv1d(d_model*2, d_model, kernel_size=1) self.mlp = nn.Sequential( nn.Linear(d_model, d_model), - # nn.ReLU(), - # nn.Linear(d_model*2, d_model*2), nn.ReLU(), nn.Linear(d_model, d_model) @@ -213,7 +208,6 @@ def forward(self, x): x = x.reshape(B,C,P) return x def STFT_for_Period(x, k=3, n_fft=16, hop_length=8, win_length=16, window='hann'): - # Window setup based on user selection if window == 'hann': window_tensor = torch.hann_window(win_length, periodic=True).to(x.device) elif window == 'hamming': @@ -249,7 +243,7 @@ def STFT_for_Period(x, k=3, n_fft=16, hop_length=8, win_length=16, window='hann' chosen_index = top_list[0] top_values.append(chosen_index) top_values = torch.tensor(top_values) - k_amplitude, k_index = torch.topk(top_values.flatten(), 1) # Flatten to get top k values across all time bins + k_amplitude, k_index = torch.topk(top_values.flatten(), 1) k_amplitude_all.append(k_amplitude) k_index_all.append(k_index) From 2b33f493102253fface680429561223fa44b563a Mon Sep 17 00:00:00 2001 From: RealSuSeven <122971162+RealSuSeven@users.noreply.github.com> Date: Mon, 17 Mar 2025 15:52:53 +0800 Subject: [PATCH 09/11] Update PPDformer_ETTm2.sh --- scripts/long_term_forecast/ETT_script/PPDformer_ETTm2.sh | 8 +++++--- 1 file changed, 5 insertions(+), 3 deletions(-) diff --git a/scripts/long_term_forecast/ETT_script/PPDformer_ETTm2.sh b/scripts/long_term_forecast/ETT_script/PPDformer_ETTm2.sh index af581d0f2..ca49a4200 100644 --- a/scripts/long_term_forecast/ETT_script/PPDformer_ETTm2.sh +++ b/scripts/long_term_forecast/ETT_script/PPDformer_ETTm2.sh @@ -1,6 +1,4 @@ export CUDA_VISIBLE_DEVICES=2 - - model_name=PPDformer patchH=2 # patchW=2 @@ -10,6 +8,7 @@ seq_len=96 python -u run.py \ + --task_name long_term_forecast \ --is_training 1 \ --root_path ./dataset/ETT-small/ \ --data_path ETTm2.csv \ @@ -37,6 +36,7 @@ python -u run.py \ --itr 1 python -u run.py \ + --task_name long_term_forecast \ --is_training 1 \ --root_path ./dataset/ETT-small/ \ --data_path ETTm2.csv \ @@ -66,6 +66,7 @@ python -u run.py \ --itr 1 python -u run.py \ + --task_name long_term_forecast \ --is_training 1 \ --root_path ./dataset/ETT-small/ \ --data_path ETTm2.csv \ @@ -95,6 +96,7 @@ python -u run.py \ --itr 1 python -u run.py \ + --task_name long_term_forecast \ --is_training 1 \ --root_path ./dataset/ETT-small/ \ --data_path ETTm2.csv \ @@ -121,4 +123,4 @@ python -u run.py \ --strideH $strideH\ --strideW 4\ --n_heads 16 \ - --itr 1 \ No newline at end of file + --itr 1 From f00732171b7cf42b6a9997f33d2b943d3f2d298f Mon Sep 17 00:00:00 2001 From: RealSuSeven <122971162+RealSuSeven@users.noreply.github.com> Date: Mon, 17 Mar 2025 15:55:58 +0800 Subject: [PATCH 10/11] Add PPDformer --- exp/exp_basic.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/exp/exp_basic.py b/exp/exp_basic.py index 3871f0bf3..dda2489aa 100644 --- a/exp/exp_basic.py +++ b/exp/exp_basic.py @@ -38,7 +38,7 @@ def __init__(self, args): 'PAttn': PAttn, 'TimeXer': TimeXer, 'WPMixer': WPMixer, - 'MultiPatchFormer': MultiPatchFormer + 'MultiPatchFormer': MultiPatchFormer, 'PPDformer': PPDformer } if args.model == 'Mamba': From 1f419170aaf4d230ccb422919318ac9e8155c6b7 Mon Sep 17 00:00:00 2001 From: RealSuSeven <122971162+RealSuSeven@users.noreply.github.com> Date: Mon, 17 Mar 2025 15:59:13 +0800 Subject: [PATCH 11/11] Update PPDformer.py --- models/PPDformer.py | 5 ++--- 1 file changed, 2 insertions(+), 3 deletions(-) diff --git a/models/PPDformer.py b/models/PPDformer.py index e9068884a..3351f7ff6 100644 --- a/models/PPDformer.py +++ b/models/PPDformer.py @@ -166,7 +166,7 @@ def __init__(self, d_model,configs): EncoderLayer( AttentionLayer( FullAttention(False, configs.factor, attention_dropout=configs.dropout, - output_attention=configs.output_attention), configs.d_model, configs.n_heads), + output_attention=False), configs.d_model, configs.n_heads), configs.d_model, configs.d_ff, dropout=configs.dropout, @@ -259,7 +259,6 @@ def __init__(self, configs): self.task_name = configs.task_name self.seq_len = configs.seq_len self.pred_len = configs.pred_len - self.output_attention = configs.output_attention self.k = configs.top_k self.patchH =configs.patchH self.patchW =configs.patchW @@ -274,7 +273,7 @@ def __init__(self, configs): EncoderLayer( AttentionLayer( FullAttention(False, configs.factor, attention_dropout=configs.dropout, - output_attention=configs.output_attention), configs.d_model, configs.n_heads), + output_attention=False), configs.d_model, configs.n_heads), configs.d_model, configs.d_ff, dropout=configs.dropout,