时间序列（Time-Series）ETSformer_EncDec.py代码解析

import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.fft as fft
from einops import rearrange, reduce, repeat
import math, random
from scipy.fftpack import next_fast_len

#数据转换
class Transform:
    #初始化，设置噪声标准差。
    def __init__(self, sigma):
        self.sigma = sigma
    #transform 应用变换：时移、尺度变换、加噪声。
    #jitter 在输入数据上添加噪声。
    #scale 通过随机因子缩放输入数据。
    #shift 在输入数据上添加一个偏移量。
    @torch.no_grad()
    def transform(self, x):
        return self.jitter(self.shift(self.scale(x)))

    def jitter(self, x):
        return x + (torch.randn(x.shape).to(x.device) * self.sigma)

    def scale(self, x):
        return x * (torch.randn(x.size(-1)).to(x.device) * self.sigma + 1)

    def shift(self, x):
        return x + (torch.randn(x.size(-1)).to(x.device) * self.sigma)

#使用快速傅里叶变换（FFT）来计算一维卷积。
def conv1d_fft(f, g, dim=-1):
    N = f.size(dim)
    M = g.size(dim)
    #计算FFT的最快长度。
    fast_len = next_fast_len(N + M - 1)
    #对两个输入进行FFT
    F_f = fft.rfft(f, fast_len, dim=dim)
    F_g = fft.rfft(g, fast_len, dim=dim)
    #对FFT结果进行逐元素乘法。
    F_fg = F_f * F_g.conj()
    #使用逆FFT转换回时域。
    out = fft.irfft(F_fg, fast_len, dim=dim)
    #调整输出，以确保它与原始数据的长度匹配。
    out = out.roll((-1,), dims=(dim,))
    idx = torch.as_tensor(range(fast_len - N, fast_len)).to(out.device)
    out = out.index_select(dim, idx)

    return out

#用于时间序列数据的平滑
class ExponentialSmoothing(nn.Module):
    #__init__ 初始化指数平滑模型的参数。
    def __init__(self, dim, nhead, dropout=0.1, aux=False):
        super().__init__()
        self._smoothing_weight = nn.Parameter(torch.randn(nhead, 1))#定义权重
        self.v0 = nn.Parameter(torch.randn(1, 1, nhead, dim))#定义偏置
        self.dropout = nn.Dropout(dropout)#减少过拟合
        if aux:
            self.aux_dropout = nn.Dropout(dropout)#减少过拟合
    #方法应用指数平滑。
    def forward(self, values, aux_values=None):
        b, t, h, d = values.shape

        init_weight, weight = self.get_exponential_weight(t)
        output = conv1d_fft(self.dropout(values), weight, dim=1)
        output = init_weight * self.v0 + output

        if aux_values is not None:
            aux_weight = weight / (1 - self.weight) * self.weight
            aux_output = conv1d_fft(self.aux_dropout(aux_values), aux_weight)
            output = output + aux_output

        return output

    def get_exponential_weight(self, T):
        # Generate array [0, 1, ..., T-1]
        powers = torch.arange(T, dtype=torch.float, device=self.weight.device)

        # (1 - \alpha) * \alpha^t, for all t = T-1, T-2, ..., 0]
        weight = (1 - self.weight) * (self.weight ** torch.flip(powers, dims=(0,)))

        # \alpha^t for all t = 1, 2, ..., T
        init_weight = self.weight ** (powers + 1)

        return rearrange(init_weight, 'h t -> 1 t h 1'), \
               rearrange(weight, 'h t -> 1 t h 1')

    @property
    def weight(self):
        return torch.sigmoid(self._smoothing_weight)

#实现一个前馈神经网络层，包括两个线性层和一个激活函数
class Feedforward(nn.Module):
    def __init__(self, d_model, dim_feedforward, dropout=0.1, activation='sigmoid'):
        # Implementation of Feedforward model
        super().__init__()
        self.linear1 = nn.Linear(d_model, dim_feedforward, bias=False)
        self.dropout1 = nn.Dropout(dropout)
        self.linear2 = nn.Linear(dim_feedforward, d_model, bias=False)
        self.dropout2 = nn.Dropout(dropout)
        self.activation = getattr(F, activation)

    def forward(self, x):
        x = self.linear2(self.dropout1(self.activation(self.linear1(x))))
        return self.dropout2(x)

#实现一个增长层，用来模拟时间序列中的趋势部分。
#包括参数初始化、投影层、指数平滑和输出投影。
class GrowthLayer(nn.Module):

    def __init__(self, d_model, nhead, d_head=None, dropout=0.1):
        super().__init__()
        self.d_head = d_head or (d_model // nhead)
        self.d_model = d_model
        self.nhead = nhead

        self.z0 = nn.Parameter(torch.randn(self.nhead, self.d_head))
        self.in_proj = nn.Linear(self.d_model, self.d_head * self.nhead)
        self.es = ExponentialSmoothing(self.d_head, self.nhead, dropout=dropout)
        self.out_proj = nn.Linear(self.d_head * self.nhead, self.d_model)

        assert self.d_head * self.nhead == self.d_model, "d_model must be divisible by nhead"

    def forward(self, inputs):
        """
        :param inputs: shape: (batch, seq_len, dim)
        :return: shape: (batch, seq_len, dim)
        """
        b, t, d = inputs.shape
        values = self.in_proj(inputs).view(b, t, self.nhead, -1)
        values = torch.cat([repeat(self.z0, 'h d -> b 1 h d', b=b), values], dim=1)
        values = values[:, 1:] - values[:, :-1]
        out = self.es(values)
        out = torch.cat([repeat(self.es.v0, '1 1 h d -> b 1 h d', b=b), out], dim=1)
        out = rearrange(out, 'b t h d -> b t (h d)')
        return self.out_proj(out)

#实现基于傅里叶变换的周期性分量建模。
class FourierLayer(nn.Module):

    def __init__(self, d_model, pred_len, k=None, low_freq=1):
        super().__init__()
        self.d_model = d_model
        self.pred_len = pred_len
        self.k = k
        self.low_freq = low_freq

#forward 方法分解输入序列的傅里叶频率，并执行外推预测。
    def forward(self, x):
        """x: (b, t, d)"""
        b, t, d = x.shape
        x_freq = fft.rfft(x, dim=1)

        if t % 2 == 0:
            x_freq = x_freq[:, self.low_freq:-1]
            f = fft.rfftfreq(t)[self.low_freq:-1]
        else:
            x_freq = x_freq[:, self.low_freq:]
            f = fft.rfftfreq(t)[self.low_freq:]

        x_freq, index_tuple = self.topk_freq(x_freq)
        f = repeat(f, 'f -> b f d', b=x_freq.size(0), d=x_freq.size(2))
        f = rearrange(f[index_tuple], 'b f d -> b f () d').to(x_freq.device)

        return self.extrapolate(x_freq, f, t)

    def extrapolate(self, x_freq, f, t):
        x_freq = torch.cat([x_freq, x_freq.conj()], dim=1)
        f = torch.cat([f, -f], dim=1)
        t_val = rearrange(torch.arange(t + self.pred_len, dtype=torch.float),
                          't -> () () t ()').to(x_freq.device)

        amp = rearrange(x_freq.abs() / t, 'b f d -> b f () d')
        phase = rearrange(x_freq.angle(), 'b f d -> b f () d')

        x_time = amp * torch.cos(2 * math.pi * f * t_val + phase)

        return reduce(x_time, 'b f t d -> b t d', 'sum')

    def topk_freq(self, x_freq):
        values, indices = torch.topk(x_freq.abs(), self.k, dim=1, largest=True, sorted=True)
        mesh_a, mesh_b = torch.meshgrid(torch.arange(x_freq.size(0)), torch.arange(x_freq.size(2)))
        index_tuple = (mesh_a.unsqueeze(1), indices, mesh_b.unsqueeze(1))
        x_freq = x_freq[index_tuple]

        return x_freq, index_tuple

#结合水平（level）、增长（growth）和季节性（season）分量的模型层。
class LevelLayer(nn.Module):

    def __init__(self, d_model, c_out, dropout=0.1):
        super().__init__()
        self.d_model = d_model
        self.c_out = c_out

        self.es = ExponentialSmoothing(1, self.c_out, dropout=dropout, aux=True)
        self.growth_pred = nn.Linear(self.d_model, self.c_out)
        self.season_pred = nn.Linear(self.d_model, self.c_out)

    def forward(self, level, growth, season):
        b, t, _ = level.shape
        growth = self.growth_pred(growth).view(b, t, self.c_out, 1)
        season = self.season_pred(season).view(b, t, self.c_out, 1)
        growth = growth.view(b, t, self.c_out, 1)
        season = season.view(b, t, self.c_out, 1)
        level = level.view(b, t, self.c_out, 1)
        out = self.es(level - season, aux_values=growth)
        out = rearrange(out, 'b t h d -> b t (h d)')
        return out

#实现编码器层，结合了增长、季节性、水平分量以及一个前馈网络。
class EncoderLayer(nn.Module):

    def __init__(self, d_model, nhead, c_out, seq_len, pred_len, k, dim_feedforward=None, dropout=0.1,
                 activation='sigmoid', layer_norm_eps=1e-5):
        super().__init__()
        self.d_model = d_model
        self.nhead = nhead
        self.c_out = c_out
        self.seq_len = seq_len
        self.pred_len = pred_len
        dim_feedforward = dim_feedforward or 4 * d_model
        self.dim_feedforward = dim_feedforward

        self.growth_layer = GrowthLayer(d_model, nhead, dropout=dropout)
        self.seasonal_layer = FourierLayer(d_model, pred_len, k=k)
        self.level_layer = LevelLayer(d_model, c_out, dropout=dropout)

        # Implementation of Feedforward model
        self.ff = Feedforward(d_model, dim_feedforward, dropout=dropout, activation=activation)
        self.norm1 = nn.LayerNorm(d_model, eps=layer_norm_eps)
        self.norm2 = nn.LayerNorm(d_model, eps=layer_norm_eps)

        self.dropout1 = nn.Dropout(dropout)
        self.dropout2 = nn.Dropout(dropout)

    def forward(self, res, level, attn_mask=None):
        season = self._season_block(res)
        res = res - season[:, :-self.pred_len]
        growth = self._growth_block(res)
        res = self.norm1(res - growth[:, 1:])
        res = self.norm2(res + self.ff(res))

        level = self.level_layer(level, growth[:, :-1], season[:, :-self.pred_len])
        return res, level, growth, season

    def _growth_block(self, x):
        x = self.growth_layer(x)
        return self.dropout1(x)

    def _season_block(self, x):
        x = self.seasonal_layer(x)
        return self.dropout2(x)

#使用多个 EncoderLayer 层堆叠来构建编码器。
#forward 方法遍历所有的编码器层并聚合输出。
class Encoder(nn.Module):

    def __init__(self, layers):
        super().__init__()
        self.layers = nn.ModuleList(layers)

    def forward(self, res, level, attn_mask=None):
        growths = []
        seasons = []
        for layer in self.layers:
            res, level, growth, season = layer(res, level, attn_mask=None)
            growths.append(growth)
            seasons.append(season)

        return level, growths, seasons

#实现一个阻尼层，用于预测未来的增长分量。
class DampingLayer(nn.Module):

    def __init__(self, pred_len, nhead, dropout=0.1):
        super().__init__()
        self.pred_len = pred_len
        self.nhead = nhead
        self._damping_factor = nn.Parameter(torch.randn(1, nhead))
        self.dropout = nn.Dropout(dropout)

    def forward(self, x):
        x = repeat(x, 'b 1 d -> b t d', t=self.pred_len)
        b, t, d = x.shape

        powers = torch.arange(self.pred_len).to(self._damping_factor.device) + 1
        powers = powers.view(self.pred_len, 1)
        damping_factors = self.damping_factor ** powers
        damping_factors = damping_factors.cumsum(dim=0)
        x = x.view(b, t, self.nhead, -1)
        x = self.dropout(x) * damping_factors.unsqueeze(-1)
        return x.view(b, t, d)

    @property
    def damping_factor(self):
        return torch.sigmoid(self._damping_factor)

#实现解码器层，其中阻尼增长和固定季节性分量用于长期预测。
class DecoderLayer(nn.Module):

    def __init__(self, d_model, nhead, c_out, pred_len, dropout=0.1):
        super().__init__()
        self.d_model = d_model
        self.nhead = nhead
        self.c_out = c_out
        self.pred_len = pred_len

        self.growth_damping = DampingLayer(pred_len, nhead, dropout=dropout)
        self.dropout1 = nn.Dropout(dropout)

    def forward(self, growth, season):
        growth_horizon = self.growth_damping(growth[:, -1:])
        growth_horizon = self.dropout1(growth_horizon)

        seasonal_horizon = season[:, -self.pred_len:]
        return growth_horizon, seasonal_horizon

#通过 DecoderLayer 层对时间序列增长和季节性分量进行解码。
#forward 方法遍历所有的解码器层并聚合输出。
class Decoder(nn.Module):

    def __init__(self, layers):
        super().__init__()
        self.d_model = layers[0].d_model
        self.c_out = layers[0].c_out
        self.pred_len = layers[0].pred_len
        self.nhead = layers[0].nhead

        self.layers = nn.ModuleList(layers)
        self.pred = nn.Linear(self.d_model, self.c_out)

    def forward(self, growths, seasons):
        growth_repr = []
        season_repr = []

        for idx, layer in enumerate(self.layers):
            growth_horizon, season_horizon = layer(growths[idx], seasons[idx])
            growth_repr.append(growth_horizon)
            season_repr.append(season_horizon)
        growth_repr = sum(growth_repr)
        season_repr = sum(season_repr)
        return self.pred(growth_repr), self.pred(season_repr)