timesnet.py

# AUTOGENERATED! DO NOT EDIT! File to edit: ../../nbs/models.timesnet.ipynb.

# %% auto 0
__all__ = ['Inception_Block_V1', 'FFT_for_Period', 'TimesBlock', 'TimesNet']

# %% ../../nbs/models.timesnet.ipynb 4
from typing import Optional

import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.fft

from neuralforecast.common._modules import DataEmbedding, TimeFeatureEmbedding
from neuralforecast.common._base_windows import BaseWindows

from neuralforecast.losses.pytorch import MAE

# %% ../../nbs/models.timesnet.ipynb 7
class Inception_Block_V1(nn.Module):
    def __init__(self, in_channels, out_channels, num_kernels=6, init_weight=True):
        super(Inception_Block_V1, self).__init__()
        self.in_channels = in_channels
        self.out_channels = out_channels
        self.num_kernels = num_kernels
        kernels = []
        for i in range(self.num_kernels):
            kernels.append(
                nn.Conv2d(in_channels, out_channels, kernel_size=2 * i + 1, padding=i)
            )
        self.kernels = nn.ModuleList(kernels)
        if init_weight:
            self._initialize_weights()

    def _initialize_weights(self):
        for m in self.modules():
            if isinstance(m, nn.Conv2d):
                nn.init.kaiming_normal_(m.weight, mode="fan_out", nonlinearity="relu")
                if m.bias is not None:
                    nn.init.constant_(m.bias, 0)

    def forward(self, x):
        res_list = []
        for i in range(self.num_kernels):
            res_list.append(self.kernels[i](x))
        res = torch.stack(res_list, dim=-1).mean(-1)
        return res

# %% ../../nbs/models.timesnet.ipynb 8
def FFT_for_Period(x, k=2):
    # [B, T, C]
    xf = torch.fft.rfft(x, dim=1)
    # find period by amplitudes
    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 TimesBlock(nn.Module):
    def __init__(self, input_size, h, k, hidden_size, conv_hidden_size, num_kernels):
        super(TimesBlock, self).__init__()
        self.input_size = input_size
        self.h = h
        self.k = k
        # parameter-efficient design
        self.conv = nn.Sequential(
            Inception_Block_V1(hidden_size, conv_hidden_size, num_kernels=num_kernels),
            nn.GELU(),
            Inception_Block_V1(conv_hidden_size, hidden_size, num_kernels=num_kernels),
        )

    def forward(self, x):
        B, T, N = x.size()
        period_list, period_weight = FFT_for_Period(x, self.k)

        res = []
        for i in range(self.k):
            period = period_list[i]
            # padding
            if (self.input_size + self.h) % period != 0:
                length = (((self.input_size + self.h) // period) + 1) * period
                padding = torch.zeros(
                    [x.shape[0], (length - (self.input_size + self.h)), x.shape[2]],
                    device=x.device,
                )
                out = torch.cat([x, padding], dim=1)
            else:
                length = self.input_size + self.h
                out = x
            # reshape
            out = (
                out.reshape(B, length // period, period, N)
                .permute(0, 3, 1, 2)
                .contiguous()
            )
            # 2D conv: from 1d Variation to 2d Variation
            out = self.conv(out)
            # reshape back
            out = out.permute(0, 2, 3, 1).reshape(B, -1, N)
            res.append(out[:, : (self.input_size + self.h), :])
        res = torch.stack(res, dim=-1)
        # adaptive aggregation
        period_weight = F.softmax(period_weight, dim=1)
        period_weight = period_weight.unsqueeze(1).unsqueeze(1).repeat(1, T, N, 1)
        res = torch.sum(res * period_weight, -1)
        # residual connection
        res = res + x
        return res

# %% ../../nbs/models.timesnet.ipynb 10
class TimesNet(BaseWindows):
    """TimesNet

    The TimesNet univariate model tackles the challenge of modeling multiple intraperiod and interperiod temporal variations.

    Parameters
    ----------
    h : int
        Forecast horizon.
    input_size : int
        Length of input window (lags).
    futr_exog_list : list of str, optional (default=None)
        Future exogenous columns.
    hist_exog_list : list of str, optional (default=None)
        Historic exogenous columns.
    stat_exog_list : list of str, optional (default=None)
        Static exogenous columns.
    exclude_insample_y : bool (default=False)
        The model skips the autoregressive features y[t-input_size:t] if True
    hidden_size : int (default=64)
        Size of embedding for embedding and encoders.
    dropout : float between [0, 1) (default=0.1)
        Dropout for embeddings.
        conv_hidden_size: int (default=64)
        Channels of the Inception block.
    top_k: int (default=5)
        Number of periods.
    num_kernels: int (default=6)
        Number of kernels for the Inception block.
    encoder_layers : int, (default=2)
        Number of encoder layers.
    loss: PyTorch module (default=MAE())
        Instantiated train loss class from [losses collection](https://nixtla.github.io/neuralforecast/losses.pytorch.html).
    valid_loss: PyTorch module (default=None, uses loss)
        Instantiated validation loss class from [losses collection](https://nixtla.github.io/neuralforecast/losses.pytorch.html).
    max_steps: int (default=1000)
        Maximum number of training steps.
    learning_rate : float (default=1e-4)
        Learning rate.
    num_lr_decays`: int (default=-1)
        Number of learning rate decays, evenly distributed across max_steps. If -1, no learning rate decay is performed.
    early_stop_patience_steps : int (default=-1)
        Number of validation iterations before early stopping. If -1, no early stopping is performed.
    val_check_steps : int (default=100)
        Number of training steps between every validation loss check.
    batch_size : int (default=32)
        Number of different series in each batch.
    valid_batch_size : int (default=None)
        Number of different series in each validation and test batch, if None uses batch_size.
    windows_batch_size : int (default=64)
        Number of windows to sample in each training batch.
    inference_windows_batch_size : int (default=256)
        Number of windows to sample in each inference batch.
    start_padding_enabled : bool (default=False)
        If True, the model will pad the time series with zeros at the beginning by input size.
    scaler_type : str (default='standard')
        Type of scaler for temporal inputs normalization see [temporal scalers](https://nixtla.github.io/neuralforecast/common.scalers.html).<br>
    random_seed : int (default=1)
        Random_seed for pytorch initializer and numpy generators.
    num_workers_loader : int (default=0)
        Workers to be used by `TimeSeriesDataLoader`.
    drop_last_loader : bool (default=False)
        If True `TimeSeriesDataLoader` drops last non-full batch.
    `optimizer`: Subclass of 'torch.optim.Optimizer', optional (default=None)
        User specified optimizer instead of the default choice (Adam).
    `optimizer_kwargs`: dict, optional (defualt=None)
        List of parameters used by the user specified `optimizer`.
    **trainer_kwargs
        Keyword trainer arguments inherited from [PyTorch Lighning's trainer](https://pytorch-lightning.readthedocs.io/en/stable/api/pytorch_lightning.trainer.trainer.Trainer.html?highlight=trainer)

        References
        ----------
    Haixu Wu and Tengge Hu and Yong Liu and Hang Zhou and Jianmin Wang and Mingsheng Long. TimesNet: Temporal 2D-Variation Modeling for General Time Series Analysis. https://openreview.net/pdf?id=ju_Uqw384Oq
    """

    # Class attributes
    SAMPLING_TYPE = "windows"

    def __init__(
        self,
        h: int,
        input_size: int,
        stat_exog_list=None,
        hist_exog_list=None,
        futr_exog_list=None,
        exclude_insample_y=False,
        hidden_size: int = 64,
        dropout: float = 0.1,
        conv_hidden_size: int = 64,
        top_k: int = 5,
        num_kernels: int = 6,
        encoder_layers: int = 2,
        loss=MAE(),
        valid_loss=None,
        max_steps: int = 1000,
        learning_rate: float = 1e-4,
        num_lr_decays: int = -1,
        early_stop_patience_steps: int = -1,
        val_check_steps: int = 100,
        batch_size: int = 32,
        valid_batch_size: Optional[int] = None,
        windows_batch_size=64,
        inference_windows_batch_size=256,
        start_padding_enabled=False,
        step_size: int = 1,
        scaler_type: str = "standard",
        random_seed: int = 1,
        num_workers_loader: int = 0,
        drop_last_loader: bool = False,
        optimizer=None,
        optimizer_kwargs=None,
        **trainer_kwargs
    ):
        super(TimesNet, self).__init__(
            h=h,
            input_size=input_size,
            hist_exog_list=hist_exog_list,
            stat_exog_list=stat_exog_list,
            futr_exog_list=futr_exog_list,
            exclude_insample_y=exclude_insample_y,
            loss=loss,
            valid_loss=valid_loss,
            max_steps=max_steps,
            learning_rate=learning_rate,
            num_lr_decays=num_lr_decays,
            early_stop_patience_steps=early_stop_patience_steps,
            val_check_steps=val_check_steps,
            batch_size=batch_size,
            windows_batch_size=windows_batch_size,
            valid_batch_size=valid_batch_size,
            inference_windows_batch_size=inference_windows_batch_size,
            start_padding_enabled=start_padding_enabled,
            step_size=step_size,
            scaler_type=scaler_type,
            num_workers_loader=num_workers_loader,
            drop_last_loader=drop_last_loader,
            random_seed=random_seed,
            optimizer=optimizer,
            optimizer_kwargs=optimizer_kwargs,
            **trainer_kwargs
        )

        # Architecture
        self.futr_input_size = len(self.futr_exog_list)
        self.hist_input_size = len(self.hist_exog_list)
        self.stat_input_size = len(self.stat_exog_list)

        if self.stat_input_size > 0:
            raise Exception("TimesNet does not support static variables yet")
        # if self.hist_input_size > 0:
        #     raise Exception("TimesNet does not support historical variables yet")

        self.c_out = self.loss.outputsize_multiplier
        self.enc_in = 1
        self.dec_in = 1

        self.model = nn.ModuleList(
            [
                TimesBlock(
                    input_size=input_size,
                    h=h,
                    k=top_k,
                    hidden_size=hidden_size,
                    conv_hidden_size=conv_hidden_size,
                    num_kernels=num_kernels,
                )
                for _ in range(encoder_layers)
            ]
        )

        self.enc_embedding = DataEmbedding(
            c_in=self.enc_in,
            exog_hist_input_size=self.hist_input_size,
            exog_futr_input_size=self.futr_input_size,
            hidden_size=hidden_size,
            pos_embedding=True,  # Original implementation uses true
            dropout=dropout,
        )
        self.encoder_layers = encoder_layers
        self.layer_norm = nn.LayerNorm(hidden_size)
        self.predict_linear = nn.Linear(self.input_size, self.h + self.input_size)
        self.projection = nn.Linear(hidden_size, self.c_out, bias=True)

        if self.futr_input_size > 0:
            self.temporal_embedding_futr = TimeFeatureEmbedding(
                input_size=self.futr_input_size, hidden_size=hidden_size
            )
            self.predict_linear_futr = nn.Linear(self.h, self.h)
        else:
            self.temporal_embedding_futr = None
            self.predict_linear_futr = None

    def forward(self, windows_batch):

        # Parse windows_batch
        insample_y = windows_batch["insample_y"]
        # insample_mask = windows_batch['insample_mask']
        hist_exog     = windows_batch['hist_exog']
        # stat_exog     = windows_batch['stat_exog']
        futr_exog = windows_batch["futr_exog"]

        # Parse inputs
        insample_y = insample_y.unsqueeze(-1)  # [Ws,L,1]
        if self.hist_input_size > 0:
            x_mark_hist = futr_exog[:, : self.input_size, :]
        else:
            x_mark_hist = None
        if self.futr_input_size > 0:
            x_mark_futr = futr_exog[:, : self.input_size, :]
            x_futr = futr_exog[:, self.input_size : , :]
        else:
            x_mark_futr = None
            x_futr = None

        # embedding
        enc_out = self.enc_embedding(insample_y, x_mark_hist, x_mark_futr)
        enc_out = self.predict_linear(enc_out.permute(0, 2, 1)).permute(
            0, 2, 1
        )  # align temporal dimension
        if x_futr is not None:
            enc_futr = self.temporal_embedding_futr(x_futr)
            enc_futr = self.predict_linear_futr(enc_futr.permute(0, 2, 1)).permute(
            0, 2, 1
        )
            enc_futr = F.pad(input=enc_futr, pad=(0, 0, self.input_size, 0), mode='constant', value=0)

        enc_out += enc_futr
        
        # TimesNet
        for i in range(self.encoder_layers):
            enc_out = self.layer_norm(self.model[i](enc_out))
        # porject back
        dec_out = self.projection(enc_out)

        forecast = self.loss.domain_map(dec_out[:, -self.h :])
        return forecast