Add PPDformer (ICASSP 2025)

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):
@@ -38,7 +38,8 @@ def __init__(self, args):
             'PAttn': PAttn,
             'TimeXer': TimeXer,
             'WPMixer': WPMixer,
-            'MultiPatchFormer': MultiPatchFormer
+            'MultiPatchFormer': MultiPatchFormer,
+            'PPDformer': PPDformer
         }
         if args.model == 'Mamba':
             print('Please make sure you have successfully installed mamba_ssm')

models/PPDformer.py

@@ -0,0 +1,351 @@
+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
+
+    def apply_fft_denoising(self, signal, threshold_factor=0.1):
+        fft_result = torch.fft.fft(signal, dim=1)
+        threshold = threshold_factor * torch.max(torch.abs(fft_result), dim=1, keepdim=True)[0]
+        fft_result[torch.abs(fft_result) < threshold] = 0
+        denoised_signal = torch.fft.ifft(fft_result, dim=1)
+        return denoised_signal.real
+
+    def forward(self, original_signal):
+        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.filter_order = filter_order
+        self.mu = mu
+        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)
+
+        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
+
+        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
+        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
+
+def optimized_linear_interpolation_padding(x, seq_len, period,mode='linear'):
+
+    B,C,T = x.shape
+    if seq_len % period != 0:
+        length = ((seq_len // period) + 1) * period
+    else:
+        length = seq_len
+        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':
+        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) 
+        if mode =='bicubic':
+            interpolated_data = F.interpolate(last_period_data_reshaped, size=(C, target), mode='bicubic', align_corners=True)
+        elif mode =='bilinear':
+            interpolated_data = F.interpolate(last_period_data_reshaped, size=(C, target), mode='bilinear', align_corners=True)
+        elif mode =='nearest2':
+            interpolated_data = F.interpolate(last_period_data_reshaped, size=(C, target), mode='nearest')        
+        interpolated_data = interpolated_data.squeeze(1)
+    return interpolated_data,length
+
+
+class ComplexWeightedSumModel(nn.Module):
+    def __init__(self, channels, d_model, normal, dropout):
+        super(ComplexWeightedSumModel, self).__init__()
+
+        self.weights_inner = nn.Parameter(torch.randn(1, channels, 1))
+        self.weights_whole = nn.Parameter(torch.randn(1, channels, 1))
+
+        self.norm_inner = nn.LayerNorm([channels, d_model])
+        self.norm_whole = nn.LayerNorm([channels, d_model])
+
+        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):
+        if self.normal:
+            inner = self.norm_inner(inner)
+            whole = self.norm_whole(whole)
+
+        weighted_inner = self.weights_inner * inner
+        weighted_whole = self.weights_whole * whole
+
+        sum_result = weighted_inner + weighted_whole
+
+        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__()
+        self.encoder1 = Encoder(
+            [
+                EncoderLayer(
+                    AttentionLayer(
+                        FullAttention(False, configs.factor, attention_dropout=configs.dropout,
+                                      output_attention=False), 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, 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)
+        x = self.conv1(x)
+        x = self.point_conv(x)
+        x = self.point(x)
+
+        if self.attention :
+          x = x.transpose(1, 2) 
+          x = self.layer_norm1(x) 
+          x, attns = self.encoder1(x, attn_mask=None)
+          x = x.transpose(1, 2)
+
+        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'):
+    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 = []
+
+    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)
+    frequency_list = xf_magnitude.permute(0,2,1)
+    frequency_list[:,:, 0] = 0 
+    k_amplitude_all = []
+    k_index_all = []
+    for c in range(C):
+        top_values = []
+        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]
+            else:
+              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) 
+        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.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)
+        self.encoder = Encoder(
+            [
+                EncoderLayer(
+                    AttentionLayer(
+                        FullAttention(False, configs.factor, attention_dropout=configs.dropout,
+                                      output_attention=False), 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)
+
+    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
+
+        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)
+        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():
+            x_selected = x[:, channels, :]
+            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, :] 
+        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

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))