1{
2  "op": "replace",
3  "filename": "Time-Series-Library/models/Custom.py",
4  "start_line": 1,
5  "end_line": 41,
6  "content": "import torch\nimport torch.nn as nn\nimport torch.nn.functional as F\nfrom layers.Autoformer_EncDec import series_decomp\n\n\nclass Model(nn.Module):\n    \"\"\"\n    Custom model for short-term time series forecasting (M4 dataset).\n    Fourier-enhanced Seasonal-Trend Decomposition (FourierDecomp).\n    \n    Forward signature: forward(x_enc, x_mark_enc, x_dec, x_mark_dec, mask=None)\n    - x_enc: [batch, seq_len, enc_in] — input time series (enc_in=1 for M4)\n    - x_mark_enc: [batch, seq_len, time_features] — time feature encoding\n    - x_dec: [batch, label_len+pred_len, dec_in] — decoder input\n    - x_mark_dec: [batch, label_len+pred_len, time_features] — decoder time features\n\n    Must return: [batch, pred_len, c_out] for forecasting\n    \"\"\"\n\n    def __init__(self, configs):\n        super(Model, self).__init__()\n        self.task_name = configs.task_name\n        self.seq_len = configs.seq_len\n        self.pred_len = configs.pred_len\n        self.enc_in = configs.enc_in\n        self.c_out = configs.c_out\n        \n        # Series decomposition\n        moving_avg = getattr(configs, 'moving_avg', 25)\n        self.decomposition = series_decomp(moving_avg)\n        \n        # Fourier-enhanced seasonal branch\n        # Use d_model if available, else default to 64\n        self.d_model = getattr(configs, 'd_model', 64)\n        # Number of frequency modes to keep\n        self.modes = min(getattr(configs, 'modes', 16), self.seq_len // 2)\n        \n        # Learnable complex weight for frequency domain linear transformation\n        self.weights_real = nn.Parameter(torch.randn(self.enc_in, self.modes, dtype=torch.float32) * 0.01)\n        self.weights_imag = nn.Parameter(torch.randn(self.enc_in, self.modes, dtype=torch.float32) * 0.01)\n        \n        # Linear projection for trend\n        self.linear_trend = nn.Linear(self.seq_len, self.pred_len)\n        # Initialize trend linear weight with mean filter\n        self.linear_trend.weight = nn.Parameter((1 / self.seq_len) * torch.ones([self.pred_len, self.seq_len]))\n        \n        # Output projection to combine seasonal and trend\n        self.output_proj = nn.Linear(2 * self.enc_in, self.c_out) if self.enc_in != self.c_out else None\n        \n    def fourier_seasonal(self, x_seasonal):\n        \"\"\"\n        x_seasonal: [batch, seq_len, enc_in]\n        Apply FFT, learnable frequency weighting, iFFT.\n        \"\"\"\n        batch, seq_len, enc_in = x_seasonal.shape\n        # FFT along time dimension\n        x_fft = torch.fft.rfft(x_seasonal, dim=1)  # [batch, seq_len//2+1, enc_in]\n        # Keep only first modes\n        x_fft_modes = x_fft[:, :self.modes, :]  # [batch, modes, enc_in]\n        \n        # Complex multiplication with learned weights (broadcast over batch)\n        # weights: [enc_in, modes] -> reshape to [1, modes, enc_in] for element-wise multiply\n        weights = torch.complex(self.weights_real, self.weights_imag).T.unsqueeze(0)  # [1, modes, enc_in]\n        # Multiply each channel independently\n        x_fft_weighted = x_fft_modes * weights\n        \n        # Pad back to original frequency size\n        full_freq = x_fft.shape[1]\n        x_fft_full = torch.zeros(batch, full_freq, enc_in, dtype=torch.complex64, device=x_seasonal.device)\n        x_fft_full[:, :self.modes, :] = x_fft_weighted\n        \n        # Inverse FFT\n        x_seasonal_out = torch.fft.irfft(x_fft_full, n=seq_len, dim=1)  # [batch, seq_len, enc_in]\n        \n        # Project to prediction length via linear (take last pred_len? or linear projection?)\n        # We'll use linear projection across time dimension\n        x_seasonal_out = x_seasonal_out.permute(0, 2, 1)  # [batch, enc_in, seq_len]\n        seasonal_proj = self.linear_seasonal(x_seasonal_out)  # [batch, enc_in, pred_len]\n        seasonal_proj = seasonal_proj.permute(0, 2, 1)  # [batch, pred_len, enc_in]\n        return seasonal_proj\n        \n    def forecast(self, x_enc, x_mark_enc, x_dec, x_mark_dec):\n        \"\"\"\n        Forecasting: given input sequence, predict future values.\n        Input: x_enc [batch, seq_len, 1]\n        Output: [batch, pred_len, 1]\n        \"\"\"\n        # Normalization as in Non-stationary Transformer\n        means = x_enc.mean(1, keepdim=True).detach()\n        x_enc = x_enc - means\n        stdev = torch.sqrt(torch.var(x_enc, dim=1, keepdim=True, unbiased=False) + 1e-5)\n        x_enc = x_enc / stdev\n        \n        # Decompose into seasonal and trend\n        seasonal, trend = self.decomposition(x_enc)  # each [batch, seq_len, enc_in]\n        \n        # Trend branch: linear projection\n        trend = trend.permute(0, 2, 1)  # [batch, enc_in, seq_len]\n        trend_out = self.linear_trend(trend)  # [batch, enc_in, pred_len]\n        trend_out = trend_out.permute(0, 2, 1)  # [batch, pred_len, enc_in]\n        \n        # Seasonal branch: Fourier-enhanced\n        seasonal = seasonal.permute(0, 2, 1)  # [batch, enc_in, seq_len]\n        # We'll use a simple linear projection for seasonal as well (for now)\n        seasonal_out = self.linear_seasonal(seasonal)  # [batch, enc_in, pred_len]\n        seasonal_out = seasonal_out.permute(0, 2, 1)\n        \n        # Combine seasonal and trend\n        combined = seasonal_out + trend_out  # [batch, pred_len, enc_in]\n        \n        # Denormalize\n        combined = combined * stdev[:, 0, :].unsqueeze(1).repeat(1, self.pred_len, 1)\n        combined = combined + means[:, 0, :].unsqueeze(1).repeat(1, self.pred_len, 1)\n        \n        # Output projection if needed\n        if self.output_proj is not None:\n            combined = self.output_proj(combined)\n        return combined\n        \n    def forward(self, x_enc, x_mark_enc, x_dec, x_mark_dec, mask=None):\n        if self.task_name == 'long_term_forecast' or self.task_name == 'short_term_forecast':\n            dec_out = self.forecast(x_enc, x_mark_enc, x_dec, x_mark_dec)\n            return dec_out[:, -self.pred_len:, :]\n        return None"
7}

1{
2  "n": 1
3}

1{
2  "op": "replace",
3  "filename": "Time-Series-Library/models/Custom.py",
4  "start_line": 1,
5  "end_line": 41,
6  "content": "import torch\nimport torch.nn as nn\nimport torch.nn.functional as F\nfrom layers.Autoformer_EncDec import series_decomp\n\n\nclass FrequencyAttention(nn.Module):\n    \"\"\"Frequency attention module for seasonal component.\"\"\"\n    def __init__(self, seq_len, pred_len, enc_in, modes=16):\n        super(FrequencyAttention, self).__init__()\n        self.seq_len = seq_len\n        self.pred_len = pred_len\n        self.enc_in = enc_in\n        self.modes = min(modes, seq_len // 2)\n        \n        # Attention weights over frequencies\n        self.attn = nn.Parameter(torch.ones(enc_in, self.modes) * 0.5)\n        # Learnable scaling factor\n        self.scale = nn.Parameter(torch.ones(1))\n        \n        # Linear projection after frequency filtering\n        self.linear = nn.Linear(seq_len, pred_len)\n        # Initialize linear weight with mean filter\n        self.linear.weight = nn.Parameter((1 / seq_len) * torch.ones(pred_len, seq_len))\n        \n    def forward(self, x):\n        # x: [batch, seq_len, enc_in]\n        batch, seq_len, enc_in = x.shape\n        \n        # FFT along time dimension\n        x_fft = torch.fft.rfft(x, dim=1)  # [batch, seq_len//2+1, enc_in]\n        # Keep only first modes\n        x_fft_modes = x_fft[:, :self.modes, :]  # [batch, modes, enc_in]\n        \n        # Apply attention weights (softmax over frequencies per channel)\n        attn_weights = F.softmax(self.attn.T, dim=0)  # [modes, enc_in]\n        attn_weights = attn_weights.permute(1, 0).unsqueeze(0)  # [1, enc_in, modes]\n        # Multiply each channel independently\n        x_fft_weighted = x_fft_modes * attn_weights * self.scale\n        \n        # Pad back to original frequency size\n        full_freq = x_fft.shape[1]\n        x_fft_full = torch.zeros(batch, full_freq, enc_in, dtype=torch.complex64, device=x.device)\n        x_fft_full[:, :self.modes, :] = x_fft_weighted\n        \n        # Inverse FFT\n        x_filtered = torch.fft.irfft(x_fft_full, n=seq_len, dim=1)  # [batch, seq_len, enc_in]\n        \n        # Linear projection to prediction length\n        x_filtered = x_filtered.permute(0, 2, 1)  # [batch, enc_in, seq_len]\n        out = self.linear(x_filtered)  # [batch, enc_in, pred_len]\n        out = out.permute(0, 2, 1)  # [batch, pred_len, enc_in]\n        return out\n\n\nclass Model(nn.Module):\n    \"\"\"\n    Custom model for short-term time series forecasting (M4 dataset).\n    Frequency-Attention Decomposition Linear (FADLinear).\n    \n    Forward signature: forward(x_enc, x_mark_enc, x_dec, x_mark_dec, mask=None)\n    - x_enc: [batch, seq_len, enc_in] — input time series (enc_in=1 for M4)\n    - x_mark_enc: [batch, seq_len, time_features] — time feature encoding\n    - x_dec: [batch, label_len+pred_len, dec_in] — decoder input\n    - x_mark_dec: [batch, label_len+pred_len, time_features] — decoder time features\n\n    Must return: [batch, pred_len, c_out] for forecasting\n    \"\"\"\n\n    def __init__(self, configs):\n        super(Model, self).__init__()\n        self.task_name = configs.task_name\n        self.seq_len = configs.seq_len\n        self.pred_len = configs.pred_len\n        self.enc_in = configs.enc_in\n        self.c_out = configs.c_out\n        \n        # Series decomposition\n        moving_avg = getattr(configs, 'moving_avg', 25)\n        self.decomposition = series_decomp(moving_avg)\n        \n        # Frequency attention for seasonal component\n        modes = getattr(configs, 'modes', 16)\n        self.freq_attn = FrequencyAttention(self.seq_len, self.pred_len, self.enc_in, modes)\n        \n        # Linear projection for trend component\n        self.linear_trend = nn.Linear(self.seq_len, self.pred_len)\n        # Initialize trend linear weight with mean filter\n        self.linear_trend.weight = nn.Parameter((1 / self.seq_len) * torch.ones([self.pred_len, self.seq_len]))\n        \n        # Output projection if enc_in != c_out\n        if self.enc_in != self.c_out:\n            self.output_proj = nn.Linear(self.enc_in, self.c_out)\n        else:\n            self.output_proj = None\n        \n    def forecast(self, x_enc, x_mark_enc, x_dec, x_mark_dec):\n        \"\"\"\n        Forecasting: given input sequence, predict future values.\n        Input: x_enc [batch, seq_len, 1]\n        Output: [batch, pred_len, 1]\n        \"\"\"\n        # Normalization as in Non-stationary Transformer\n        means = x_enc.mean(1, keepdim=True).detach()\n        x_enc = x_enc - means\n        stdev = torch.sqrt(torch.var(x_enc, dim=1, keepdim=True, unbiased=False) + 1e-5)\n        x_enc = x_enc / stdev\n        \n        # Decompose into seasonal and trend\n        seasonal, trend = self.decomposition(x_enc)  # each [batch, seq_len, enc_in]\n        \n        # Trend branch: linear projection\n        trend = trend.permute(0, 2, 1)  # [batch, enc_in, seq_len]\n        trend_out = self.linear_trend(trend)  # [batch, enc_in, pred_len]\n        trend_out = trend_out.permute(0, 2, 1)  # [batch, pred_len, enc_in]\n        \n        # Seasonal branch: frequency attention + linear projection\n        seasonal_out = self.freq_attn(seasonal)  # [batch, pred_len, enc_in]\n        \n        # Combine seasonal and trend\n        combined = seasonal_out + trend_out  # [batch, pred_len, enc_in]\n        \n        # Denormalize\n        combined = combined * stdev[:, 0, :].unsqueeze(1).repeat(1, self.pred_len, 1)\n        combined = combined + means[:, 0, :].unsqueeze(1).repeat(1, self.pred_len, 1)\n        \n        # Output projection if needed\n        if self.output_proj is not None:\n            combined = self.output_proj(combined)\n        return combined\n        \n    def forward(self, x_enc, x_mark_enc, x_dec, x_mark_dec, mask=None):\n        if self.task_name == 'long_term_forecast' or self.task_name == 'short_term_forecast':\n            dec_out = self.forecast(x_enc, x_mark_enc, x_dec, x_mark_dec)\n            return dec_out[:, -self.pred_len:, :]\n        return None"
7}

1{
2  "op": "replace",
3  "filename": "Time-Series-Library/models/Custom.py",
4  "start_line": 7,
5  "end_line": 53,
6  "content": "class FrequencyAttention(nn.Module):\n    \"\"\"Frequency attention module for seasonal component.\"\"\"\n    def __init__(self, seq_len, pred_len, enc_in, modes=16):\n        super(FrequencyAttention, self).__init__()\n        self.seq_len = seq_len\n        self.pred_len = pred_len\n        self.enc_in = enc_in\n        self.modes = min(modes, seq_len // 2 + 1)\n        \n        # Attention weights over frequencies per channel\n        self.attn = nn.Parameter(torch.ones(enc_in, self.modes) * 0.5)\n        # Learnable scaling factor\n        self.scale = nn.Parameter(torch.ones(1))\n        \n        # Linear projection after frequency filtering\n        self.linear = nn.Linear(seq_len, pred_len)\n        # Initialize linear weight with mean filter\n        self.linear.weight = nn.Parameter((1 / seq_len) * torch.ones(pred_len, seq_len))\n        \n    def forward(self, x):\n        # x: [batch, seq_len, enc_in]\n        batch, seq_len, enc_in = x.shape\n        \n        # FFT along time dimension\n        x_fft = torch.fft.rfft(x, dim=1)  # [batch, seq_len//2+1, enc_in]\n        # Keep only first modes\n        x_fft_modes = x_fft[:, :self.modes, :]  # [batch, modes, enc_in]\n        \n        # Apply attention weights (softmax over frequencies per channel)\n        attn_weights = F.softmax(self.attn, dim=1)  # [enc_in, modes]\n        # Reshape to [1, modes, enc_in] for element-wise multiplication\n        attn_weights = attn_weights.T.unsqueeze(0)  # [1, modes, enc_in]\n        # Multiply each channel independently\n        x_fft_weighted = x_fft_modes * attn_weights * self.scale\n        \n        # Pad back to original frequency size\n        full_freq = x_fft.shape[1]\n        x_fft_full = torch.zeros(batch, full_freq, enc_in, dtype=torch.complex64, device=x_fft.device)\n        x_fft_full[:, :self.modes, :] = x_fft_weighted\n        \n        # Inverse FFT\n        x_filtered = torch.fft.irfft(x_fft_full, n=seq_len, dim=1)  # [batch, seq_len, enc_in]\n        \n        # Linear projection to prediction length\n        x_filtered = x_filtered.permute(0, 2, 1)  # [batch, enc_in, seq_len]\n        out = self.linear(x_filtered)  # [batch, enc_in, pred_len]\n        out = out.permute(0, 2, 1)  # [batch, pred_len, enc_in]\n        return out"
7}

1{
2  "op": "replace",
3  "filename": "Time-Series-Library/models/Custom.py",
4  "start_line": 57,
5  "end_line": 137,
6  "content": "class Model(nn.Module):\n    \"\"\"\n    Custom model for short-term time series forecasting (M4 dataset).\n    Multi-Scale Decomposition with Frequency Attention (MSDFA).\n    \n    Forward signature: forward(x_enc, x_mark_enc, x_dec, x_mark_dec, mask=None)\n    - x_enc: [batch, seq_len, enc_in] — input time series (enc_in=1 for M4)\n    - x_mark_enc: [batch, seq_len, time_features] — time feature encoding\n    - x_dec: [batch, label_len+pred_len, dec_in] — decoder input\n    - x_mark_dec: [batch, label_len+pred_len, time_features] — decoder time features\n\n    Must return: [batch, pred_len, c_out] for forecasting\n    \"\"\"\n\n    def __init__(self, configs):\n        super(Model, self).__init__()\n        self.task_name = configs.task_name\n        self.seq_len = configs.seq_len\n        self.pred_len = configs.pred_len\n        self.enc_in = configs.enc_in\n        self.c_out = configs.c_out\n        \n        # Multi-scale series decomposition\n        kernel_sizes = getattr(configs, 'kernel_sizes', [3, 5, 7])\n        self.decomposition = series_decomp_multi(kernel_sizes)\n        self.num_scales = len(kernel_sizes)\n        \n        # Frequency attention for seasonal components per scale\n        modes = getattr(configs, 'modes', 16)\n        self.freq_attns = nn.ModuleList([\n            FrequencyAttention(self.seq_len, self.pred_len, self.enc_in, modes)\n            for _ in range(self.num_scales)\n        ])\n        \n        # Linear projection for trend components per scale\n        self.linear_trends = nn.ModuleList([\n            nn.Linear(self.seq_len, self.pred_len)\n            for _ in range(self.num_scales)\n        ])\n        # Initialize trend linear weights with mean filter\n        for linear in self.linear_trends:\n            linear.weight = nn.Parameter((1 / self.seq_len) * torch.ones([self.pred_len, self.seq_len]))\n        \n        # Learnable weights for combining scales\n        self.scale_weights_seasonal = nn.Parameter(torch.ones(self.num_scales) / self.num_scales)\n        self.scale_weights_trend = nn.Parameter(torch.ones(self.num_scales) / self.num_scales)\n        \n        # Output projection if enc_in != c_out\n        if self.enc_in != self.c_out:\n            self.output_proj = nn.Linear(self.enc_in, self.c_out)\n        else:\n            self.output_proj = None\n        \n    def forecast(self, x_enc, x_mark_enc, x_dec, x_mark_dec):\n        \"\"\"\n        Forecasting: given input sequence, predict future values.\n        Input: x_enc [batch, seq_len, 1]\n        Output: [batch, pred_len, 1]\n        \"\"\"\n        # Normalization as in Non-stationary Transformer\n        means = x_enc.mean(1, keepdim=True).detach()\n        x_enc = x_enc - means\n        stdev = torch.sqrt(torch.var(x_enc, dim=1, keepdim=True, unbiased=False) + 1e-5)\n        x_enc = x_enc / stdev\n        \n        # Decompose into seasonal and trend (multi-scale)\n        seasonal_list, trend_list = self.decomposition(x_enc)  # each list of length num_scales\n        \n        # Process each scale\n        seasonal_outs = []\n        trend_outs = []\n        for i in range(self.num_scales):\n            # Seasonal branch: frequency attention + linear projection\n            seasonal_i = seasonal_list[i]  # [batch, seq_len, enc_in]\n            seasonal_out_i = self.freq_attns[i](seasonal_i)  # [batch, pred_len, enc_in]\n            seasonal_outs.append(seasonal_out_i)\n            \n            # Trend branch: linear projection\n            trend_i = trend_list[i].permute(0, 2, 1)  # [batch, enc_in, seq_len]\n            trend_out_i = self.linear_trends[i](trend_i)  # [batch, enc_in, pred_len]\n            trend_out_i = trend_out_i.permute(0, 2, 1)  # [batch, pred_len, enc_in]\n            trend_outs.append(trend_out_i)\n        \n        # Combine scales with learned weights\n        scale_weights_seasonal = F.softmax(self.scale_weights_seasonal, dim=0)\n        scale_weights_trend = F.softmax(self.scale_weights_trend, dim=0)\n        \n        seasonal_out = torch.zeros_like(seasonal_outs[0])\n        trend_out = torch.zeros_like(trend_outs[0])\n        for i in range(self.num_scales):\n            seasonal_out += scale_weights_seasonal[i] * seasonal_outs[i]\n            trend_out += scale_weights_trend[i] * trend_outs[i]\n        \n        # Combine seasonal and trend\n        combined = seasonal_out + trend_out  # [batch, pred_len, enc_in]\n        \n        # Denormalize\n        combined = combined * stdev[:, 0, :].unsqueeze(1).repeat(1, self.pred_len, 1)\n        combined = combined + means[:, 0, :].unsqueeze(1).repeat(1, self.pred_len, 1)\n        \n        # Output projection if needed\n        if self.output_proj is not None:\n            combined = self.output_proj(combined)\n        return combined\n        \n    def forward(self, x_enc, x_mark_enc, x_dec, x_mark_dec, mask=None):\n        if self.task_name == 'long_term_forecast' or self.task_name == 'short_term_forecast':\n            dec_out = self.forecast(x_enc, x_mark_enc, x_dec, x_mark_dec)\n            return dec_out[:, -self.pred_len:, :]\n        return None"
7}

1{
2  "op": "replace",
3  "filename": "Time-Series-Library/models/Custom.py",
4  "start_line": 4,
5  "end_line": 4,
6  "content": "from layers.Autoformer_EncDec import series_decomp, series_decomp_multi"
7}