时间序列模型SCINet（代码解析）

前言

• SCINet模型，精度仅次于NLinear的时间序列模型，在ETTh2数据集上单变量预测结果甚至比NLinear模型还要好。
• 在这里还是建议大家去读一读论文，论文写的很规范，很值得学习，论文地址
• SCINet模型Github项目地址，下载项目文件，需要注意的是该项目仅支持在GPU上运行，如果没有GPU会报错。
• 关于该模型的理论部分，本来准备自己写的，但是看到已经有很多很优秀的帖子了，这里给大家推荐几篇：
  • SCINet学习记录
  • SCONet论文阅读笔记
• SCINet学习记录中有一副思维导图画的很好，这里搬运过来方便大家在阅读代码时对照模型架构。
  
• 由于理论部分已经有了，这里我仅对项目中各代码以及框架做注释说明，方便大家理解代码，后面如果有需要，可以再写一篇，对于自定义数据如何使用SCINet模型。

参数设定模块(run_ETTh)

• 因为作者在做对比实验时用了很多公共数据集，所以文件夹中有run_ETTh.py、run_financial.py、run_pems.py3个文件，分别对应3大主要公共数据集，这里选用ETTh数据集作为示范。所以首先打开run_ETTh.py文件
• ETTh数据集需要自行下载，如果是在Linux系统中可以直接运行项目文件下prepare_data.sh文件，下载全部数据集。如果是win系统，则需要自己下载.csv文件，并在项目文件夹下创建datasets文件夹，并将数据放入该文件夹。
• 我下载了ETTh1.csv文件，后面的示范均在该数据集上进行

参数含义

下面是各参数含义（注释）

# 模型名称
parser.add_argument('--model', type=str, required=False, default='SCINet', help='model of the experiment')

### -------  dataset settings --------------
# 数据名称
parser.add_argument('--data', type=str, required=False, default='ETTh1', choices=['ETTh1', 'ETTh2', 'ETTm1'], help='name of dataset')
# 数据路径
parser.add_argument('--root_path', type=str, default='./datasets/', help='root path of the data file')
# 数据文件
parser.add_argument('--data_path', type=str, default='ETTh1.csv', help='location of the data file')
# 预测方式(S:单变量预测,M：多变量预测)
parser.add_argument('--features', type=str, default='M', choices=['S', 'M'], help='features S is univariate, M is multivariate')
# 需要预测列的列名
parser.add_argument('--target', type=str, default='OT', help='target feature')
# 时间采样格式
parser.add_argument('--freq', type=str, default='h', help='freq for time features encoding, options:[s:secondly, t:minutely, h:hourly, d:daily, b:business days, w:weekly, m:monthly], you can also use more detailed freq like 15min or 3h')
# 模型存储路径
parser.add_argument('--checkpoints', type=str, default='exp/ETT_checkpoints/', help='location of model checkpoints')
# 是否翻转序列
parser.add_argument('--inverse', type=bool, default =False, help='denorm the output data')
# 时间特征编码方式
parser.add_argument('--embed', type=str, default='timeF', help='time features encoding, options:[timeF, fixed, learned]')


### -------  device settings --------------
# 是否使用GPU(实测这个参数并没什么作用，即使填写False也无法使用CPU训练模型)
parser.add_argument('--use_gpu', type=bool, default=True, help='use gpu')
# 使用GPU设备ID
parser.add_argument('--gpu', type=int, default=0, help='gpu')
# 是否多GPU并行
parser.add_argument('--use_multi_gpu', action='store_true', help='use multiple gpus', default=False)
# 选用GPU设备ID
parser.add_argument('--devices', type=str, default='0',help='device ids of multile gpus')
                                                                                  
### -------  input/output length settings --------------
# 回视窗口大小
parser.add_argument('--seq_len', type=int, default=96, help='input sequence length of SCINet encoder, look back window')
# 先验窗口大小
parser.add_argument('--label_len', type=int, default=48, help='start token length of Informer decoder')
# 需要预测序列长度
parser.add_argument('--pred_len', type=int, default=48, help='prediction sequence length, horizon')
# 丢弃数据长度
parser.add_argument('--concat_len', type=int, default=0)
parser.add_argument('--single_step', type=int, default=0)
parser.add_argument('--single_step_output_One', type=int, default=0)
# 最后一层损失权重
parser.add_argument('--lastWeight', type=float, default=1.0)
                                                              
### -------  training settings --------------
# 多文件并列
parser.add_argument('--cols', type=str, nargs='+', help='file list')
# 多线程训练(win系统下该参数置0)
parser.add_argument('--num_workers', type=int, default=0, help='data loader num workers')
# 实验次数
parser.add_argument('--itr', type=int, default=0, help='experiments times')
# 训练迭代次数
parser.add_argument('--train_epochs', type=int, default=100, help='train epochs')
# mini_batch_size
parser.add_argument('--batch_size', type=int, default=32, help='batch size of train input data')
# 早停策略检测轮数
parser.add_argument('--patience', type=int, default=5, help='early stopping patience')
# 学习率
parser.add_argument('--lr', type=float, default=0.0001, help='optimizer learning rate')
# 损失函数
parser.add_argument('--loss', type=str, default='mae',help='loss function')
# 学习率更新策略
parser.add_argument('--lradj', type=int, default=1,help='adjust learning rate')
# 是否使用半精度加快训练速度
parser.add_argument('--use_amp', action='store_true', help='use automatic mixed precision training', default=False)
# 是否保存结果(如果你想要保存预测结果，请将该参数改为True)
parser.add_argument('--save', type=bool, default =False, help='save the output results')
# 模型名称
parser.add_argument('--model_name', type=str, default='SCINet')
# 是否断续训练
parser.add_argument('--resume', type=bool, default=False)
# 是否评估模型
parser.add_argument('--evaluate', type=bool, default=False)

### -------  model settings --------------
# 隐藏通道数
parser.add_argument('--hidden-size', default=1, type=float, help='hidden channel of module')
# 使用交互学习或基本学习策略
parser.add_argument('--INN', default=1, type=int, help='use INN or basic strategy')
# kernel size
parser.add_argument('--kernel', default=5, type=int, help='kernel size, 3, 5, 7')
# 是否扩张
parser.add_argument('--dilation', default=1, type=int, help='dilation')
# 回视窗口
parser.add_argument('--window_size', default=12, type=int, help='input size')
# dropout率
parser.add_argument('--dropout', type=float, default=0.5, help='dropout')
# 位置编码
parser.add_argument('--positionalEcoding', type=bool, default=False)
parser.add_argument('--groups', type=int, default=1)
# SCINet block
parser.add_argument('--levels', type=int, default=3)
# SCINet blocks层数
parser.add_argument('--stacks', type=int, default=1, help='1 stack or 2 stacks')
# 解码器层数
parser.add_argument('--num_decoder_layer', type=int, default=1)
parser.add_argument('--RIN', type=bool, default=False)
parser.add_argument('--decompose', type=bool,default=False)

数据文件参数

data_parser = {
	# data：数据文件名，T：预测列列名，M(多变量预测)，S(单变量预测)，MS(多特征预测单变量)
    'ETTh1': {'data': 'ETTh1.csv', 'T': 'OT', 'M': [7, 7, 7], 'S': [1, 1, 1], 'MS': [7, 7, 1]},
    'ETTh2': {'data': 'ETTh2.csv', 'T': 'OT', 'M': [7, 7, 7], 'S': [1, 1, 1], 'MS': [7, 7, 1]},
    'ETTm1': {'data': 'ETTm1.csv', 'T': 'OT', 'M': [7, 7, 7], 'S': [1, 1, 1], 'MS': [7, 7, 1]},
    'ETTm2': {'data': 'ETTm2.csv', 'T': 'OT', 'M': [7, 7, 7], 'S': [1, 1, 1], 'MS': [7, 7, 1]},
    'WTH': {'data': 'WTH.csv', 'T': 'WetBulbCelsius', 'M': [12, 12, 12], 'S': [1, 1, 1], 'MS': [12, 12, 1]},
    'ECL': {'data': 'ECL.csv', 'T': 'MT_320', 'M': [321, 321, 321], 'S': [1, 1, 1], 'MS': [321, 321, 1]},
    'Solar': {'data': 'solar_AL.csv', 'T': 'POWER_136', 'M': [137, 137, 137], 'S': [1, 1, 1], 'MS': [137, 137, 1]},
}

• 下面是模型训练函数，这里不进行注释了

数据处理模块（etth_data_loader）

• 从run_ETTh.py文件中exp.train(setting)，train方法进入exp_ETTh.py文件，在_get_data中找到ETTh1数据处理方法

data_dict = {'ETTh1':Dataset_ETT_hour,
             'ETTh2':Dataset_ETT_hour,
             'ETTm1':Dataset_ETT_minute,
             'ETTm2':Dataset_ETT_minute,
             'WTH':Dataset_Custom,
             'ECL':Dataset_Custom,
             'Solar':Dataset_Custom,}

• 可以看到ETTh1数据处理方法为Dataset_ETT_hour，我们进入etth_data_loader.py文件，找到Dataset_ETT_hour类
• __init__主要用于传各类参数，这里不过多赘述，主要对__read_data__和__getitem__进行说明

    def __read_data__(self):
        # 实例化归一化
        self.scaler = StandardScaler()
        # 读取CSV文件
        df_raw = pd.read_csv(os.path.join(self.root_path,
                                          self.data_path))
        # [0,训练序列长度-回视窗口,全部序列长度-测试序列长度-回视窗口]
        border1s = [0, 12*30*24 - self.seq_len, 12*30*24+4*30*24 - self.seq_len]
        # [训练序列长度,全部序列长度-测试序列长度,全部序列长度]
        border2s = [12*30*24, 12*30*24+4*30*24, 12*30*24+8*30*24]
        # train:[0,训练数据长度]
        # val:[训练序列长度-回视窗口,全部序列长度-测试序列长度]
        # test:[全部序列长度-测试序列长度-回视窗口,全部序列长度]
        border1 = border1s[self.set_type]
        border2 = border2s[self.set_type]

        # 若采用多变量预测(M或MS)
        if self.features=='M' or self.features=='MS':
            # 取出特征列列名
            cols_data = df_raw.columns[1:]
            # 取出特征列
            df_data = df_raw[cols_data]
        # 若采用单变量预测
        elif self.features=='S':
            # 取出预测列
            df_data = df_raw[[self.target]]
        # 若需要进行归一化
        if self.scale:
            # 取出[0,训练序列长度]区间数据
            train_data = df_data[border1s[0]:border2s[0]]
            # 归一化
            self.scaler.fit(train_data.values)
            data = self.scaler.transform(df_data.values)
            # data = self.scaler.fit_transform(df_data.values)
        # 否则将预测列变为数组
        else:
            data = df_data.values
        # 取对应区间时间列
        df_stamp = df_raw[['date']][border1:border2]
        # 将时间转换为标准格式
        df_stamp['date'] = pd.to_datetime(df_stamp.date)
        # 构建时间特征
        data_stamp = time_features(df_stamp, timeenc=self.timeenc, freq=self.freq)

        # 取对应数据区间(train、val、test)
        self.data_x = data[border1:border2]
        # 如果需要翻转时间序列
        if self.inverse:
            self.data_y = df_data.values[border1:border2]
        # 否则取数据区间(train、val、test)
        else:
            self.data_y = data[border1:border2]
        self.data_stamp = data_stamp

• 需要注意的是time_features函数，用来提取日期特征，比如't':['month','day','weekday','hour','minute']，表示提月，天，周，小时，分钟。可以打开timefeatures.py文件进行查阅
• 同样的，对__getitem__进行说明

    def __getitem__(self, index):
        # 起点
        s_begin = index
        # 终点(起点 + 回视窗口)
        s_end = s_begin + self.seq_len
        # (终点 - 先验序列窗口)
        r_begin = s_end - self.label_len
        # (终点 + 预测序列长度)
        r_end = r_begin + self.label_len + self.pred_len

        # seq_x = [起点,起点 + 回视窗口]
        seq_x = self.data_x[s_begin:s_end]  # 0 - 24
        # seq_y = [终点 - 先验序列窗口,终点 + 预测序列长度]
        seq_y = self.data_y[r_begin:r_end] # 0 - 48
        # 取对应时间特征
        seq_x_mark = self.data_stamp[s_begin:s_end]
        seq_y_mark = self.data_stamp[r_begin:r_end]

        return seq_x, seq_y, seq_x_mark, seq_y_mark

• 光看注释可能对各区间划分不那么清楚，这里我画了一幅示意图，希望能帮大家理解
  

SCINet模型架构（SCINet）

• 打开model文件夹，找到SCINet类，先定位到main()函数，可以看到main()函数这里实例化了一个SCINet类，并将参数传入其中

if __name__ == '__main__':
    parser = argparse.ArgumentParser()

    parser.add_argument('--window_size', type=int, default=96)
    parser.add_argument('--horizon', type=int, default=12)

    parser.add_argument('--dropout', type=float, default=0.5)
    parser.add_argument('--groups', type=int, default=1)

    parser.add_argument('--hidden-size', default=1, type=int, help='hidden channel of module')
    parser.add_argument('--INN', default=1, type=int, help='use INN or basic strategy')
    parser.add_argument('--kernel', default=3, type=int, help='kernel size')
    parser.add_argument('--dilation', default=1, type=int, help='dilation')
    parser.add_argument('--positionalEcoding', type=bool, default=True)

    parser.add_argument('--single_step_output_One', type=int, default=0)

    args = parser.parse_args()
    # 实例化SCINet类
    model = SCINet(output_len = args.horizon, input_len= args.window_size, input_dim = 9, hid_size = args.hidden_size, num_stacks = 1,
                num_levels = 3, concat_len = 0, groups = args.groups, kernel = args.kernel, dropout = args.dropout,
                 single_step_output_One = args.single_step_output_One, positionalE =  args.positionalEcoding, modified = True).cuda()
    x = torch.randn(32, 96, 9).cuda()
    y = model(x)
    print(y.shape)

• 下面我们从头开始结合论文中的架构图讲解代码。

Splitting类（奇偶序列分离）

• 这部分比较简单，就是通过数据下标将序列分为奇序列与偶序列

class Splitting(nn.Module):
    def __init__(self):
        super(Splitting, self).__init__()

    def even(self, x):
        # 将奇序列分离
        return x[:, ::2, :]

    def odd(self, x):
        # 将偶序列分离
        return x[:, 1::2, :]

    def forward(self, x):
        return (self.even(x), self.odd(x))

Interactor类（下采样与交互学习）

• 这一部分将奇、偶序列分别使用不同分辨率的卷积捕捉时间信息，然后两序列分别进行加减运算，模型架构图
  

• 注释写的非常清楚，这一部分建议多琢磨

class Interactor(nn.Module):
    def __init__(self, in_planes, splitting=True,
                 kernel = 5, dropout=0.5, groups = 1, hidden_size = 1, INN = True):
        super(Interactor, self).__init__()
        self.modified = INN
        self.kernel_size = kernel
        self.dilation = 1
        self.dropout = dropout
        self.hidden_size = hidden_size
        self.groups = groups

        # 如果通道数为偶数
        if self.kernel_size % 2 == 0:
            # 1 * (kernel -2) // 2 + 1
            pad_l = self.dilation * (self.kernel_size - 2) // 2 + 1 #by default: stride==1
            # 1 * kernel // 2 + 1
            pad_r = self.dilation * (self.kernel_size) // 2 + 1 #by default: stride==1
            # 如果kernel_size = 4, pda_l = 2,pad_r = 3

        # 如果通道数为奇数
        else:
            pad_l = self.dilation * (self.kernel_size - 1) // 2 + 1 # we fix the kernel size of the second layer as 3.
            pad_r = self.dilation * (self.kernel_size - 1) // 2 + 1
            # 如果kernel_size = 3, pda_l = 2,pad_r = 2

        self.splitting = splitting
        self.split = Splitting()

        modules_P = []
        modules_U = []
        modules_psi = []
        modules_phi = []
        prev_size = 1

        size_hidden = self.hidden_size
        modules_P += [
            # ReplicationPad1d用输入边界的反射来填充输入张量
            nn.ReplicationPad1d((pad_l, pad_r)),

            # 1维卷积(in_channels,out_channels,kernel_size)-->(7,7,5)
            nn.Conv1d(in_planes * prev_size, int(in_planes * size_hidden),
                      kernel_size=self.kernel_size, dilation=self.dilation, stride=1, groups= self.groups),
            # LeakyReLU激活层
            nn.LeakyReLU(negative_slope=0.01, inplace=True),

            # Dropout层
            nn.Dropout(self.dropout),
            # 1维卷积(in_channels,out_channels,kernel_size)-->(7,7,3)
            nn.Conv1d(int(in_planes * size_hidden), in_planes,
                      kernel_size=3, stride=1, groups= self.groups),
            # Tanh激活层
            nn.Tanh()
        ]
        modules_U += [
            # ReplicationPad1d用输入边界的反射来填充输入张量
            nn.ReplicationPad1d((pad_l, pad_r)),
            # 1维卷积(in_channels, out_channels,kernel_size)-->(7,7,5)
            nn.Conv1d(in_planes * prev_size, int(in_planes * size_hidden),
                      kernel_size=self.kernel_size, dilation=self.dilation, stride=1, groups= self.groups),
            # LeakyReLu激活层
            nn.LeakyReLU(negative_slope=0.01, inplace=True),
            # Dropout层
            nn.Dropout(self.dropout),
            # 1维卷积(in_channels, out_channels,kernel_size)-->(7,7,3)
            nn.Conv1d(int(in_planes * size_hidden), in_planes,
                      kernel_size=3, stride=1, groups= self.groups),
            # Tanh激活层
            nn.Tanh()
        ]

        modules_phi += [
            # ReplicationPad1d用输入边界的反射来填充输入张量
            nn.ReplicationPad1d((pad_l, pad_r)),
            # 1维卷积(in_channels, out_channels,kernel_size)-->(7,7,5)
            nn.Conv1d(in_planes * prev_size, int(in_planes * size_hidden),
                      kernel_size=self.kernel_size, dilation=self.dilation, stride=1, groups= self.groups),
            # LeakyReLU激活层
            nn.LeakyReLU(negative_slope=0.01, inplace=True),
            # Dropout层
            nn.Dropout(self.dropout),
            # 1维卷积(in_channels, out_channels,kernel_size)-->(7,7,3)
            nn.Conv1d(int(in_planes * size_hidden), in_planes,
                      kernel_size=3, stride=1, groups= self.groups),
            # Tanh激活层
            nn.Tanh()
        ]
        modules_psi += [
            # ReplicationPad1d用输入边界的反射来填充输入张量
            nn.ReplicationPad1d((pad_l, pad_r)),
            # 一维卷积(in_channels, out_channels,kernel_size)-->(7,7,5)
            nn.Conv1d(in_planes * prev_size, int(in_planes * size_hidden),
                      kernel_size=self.kernel_size, dilation=self.dilation, stride=1, groups= self.groups),
            # LeakyReLU激活层
            nn.LeakyReLU(negative_slope=0.01, inplace=True),
            # Dropout层
            nn.Dropout(self.dropout),
            # 1维卷积(in_channels, out_channels,kernel_size)-->(7,7,3)
            nn.Conv1d(int(in_planes * size_hidden), in_planes,
                      kernel_size=3, stride=1, groups= self.groups),
            # Tanh激活层
            nn.Tanh()
        ]
        self.phi = nn.Sequential(*modules_phi)
        self.psi = nn.Sequential(*modules_psi)
        self.P = nn.Sequential(*modules_P)
        self.U = nn.Sequential(*modules_U)

    def forward(self, x):
        # 将奇偶序列分隔
        if self.splitting:
            (x_even, x_odd) = self.split(x)
        else:
            (x_even, x_odd) = x

        # 如果INN不为0
        if self.modified:
            # 交换奇、偶序列维度[B,L,D] --> [B,D,L]
            x_even = x_even.permute(0, 2, 1)
            x_odd = x_odd.permute(0, 2, 1)

            # mul()函数矩阵点乘,计算经过phi层的指数值
            d = x_odd.mul(torch.exp(self.phi(x_even)))
            c = x_even.mul(torch.exp(self.psi(x_odd)))

            # 更新奇序列(奇序列 + 经过U层的偶序列)
            x_even_update = c + self.U(d)
            # 更新偶序列(偶序列 - 经过P层的奇序列)
            x_odd_update = d - self.P(c)

            return (x_even_update, x_odd_update)

        else:
            # 不计算指数值
            x_even = x_even.permute(0, 2, 1)
            x_odd = x_odd.permute(0, 2, 1)

            d = x_odd - self.P(x_even)
            c = x_even + self.U(d)

            return (c, d)

InteractorLevel类

• 该类主要实例化Interactor类，并得到奇、偶序列特征

class InteractorLevel(nn.Module):
    def __init__(self, in_planes, kernel, dropout, groups , hidden_size, INN):
        super(InteractorLevel, self).__init__()
        self.level = Interactor(in_planes = in_planes, splitting=True,
                 kernel = kernel, dropout=dropout, groups = groups, hidden_size = hidden_size, INN = INN)

    def forward(self, x):
        (x_even_update, x_odd_update) = self.level(x)
        return (x_even_update, x_odd_update)

LevelSCINet类

• 该类主要实例化InteractorLevel类，并将得到的奇、偶序列特征进行维度交换方便SCINet_Tree框架运算

class LevelSCINet(nn.Module):
    def __init__(self,in_planes, kernel_size, dropout, groups, hidden_size, INN):
        super(LevelSCINet, self).__init__()
        self.interact = InteractorLevel(in_planes= in_planes, kernel = kernel_size, dropout = dropout, groups =groups , hidden_size = hidden_size, INN = INN)

    def forward(self, x):
        (x_even_update, x_odd_update) = self.interact(x)
        # 交换奇、偶序列维度[B,D,L] --> [B,T,D]
        return x_even_update.permute(0, 2, 1), x_odd_update.permute(0, 2, 1)

SCINet_Tree类

• 这就是论文中提到的二叉树结构，可以更有效的捕捉时间序列的长短期依赖，网络框架图：
  

• 这部分框架为SCINet的核心框架，建议认真阅读

class SCINet_Tree(nn.Module):
    def __init__(self, in_planes, current_level, kernel_size, dropout, groups, hidden_size, INN):
        super().__init__()
        self.current_level = current_level


        self.workingblock = LevelSCINet(
            in_planes = in_planes,
            kernel_size = kernel_size,
            dropout = dropout,
            groups= groups,
            hidden_size = hidden_size,
            INN = INN)

        # 如果current_level不为0
        if current_level!=0:
            self.SCINet_Tree_odd=SCINet_Tree(in_planes, current_level-1, kernel_size, dropout, groups, hidden_size, INN)
            self.SCINet_Tree_even=SCINet_Tree(in_planes, current_level-1, kernel_size, dropout, groups, hidden_size, INN)
    
    def zip_up_the_pants(self, even, odd):
        # 交换奇数据下标(B,L,D) --> (L,B,D)
        even = even.permute(1, 0, 2)
        odd = odd.permute(1, 0, 2) #L, B, D

        # 取序列长度
        even_len = even.shape[0]
        odd_len = odd.shape[0]

        # 取奇、偶数据序列长度小值
        mlen = min((odd_len, even_len))
        _ = []
        for i in range(mlen):
            # 在第1维度前增加1个维度
            # _.shape:[12],even.shape:[12,32,7],odd.shape:[12,32,7]
            _.append(even[i].unsqueeze(0))
            _.append(odd[i].unsqueeze(0))

        # 如果偶序列长度 < 奇序列长度
        if odd_len < even_len: 
            _.append(even[-1].unsqueeze(0))
        # 将张量按照第1维度拼接
        return torch.cat(_,0).permute(1,0,2) #B, L, D
        
    def forward(self, x):
        # 取得更新后的奇、偶序列
        x_even_update, x_odd_update= self.workingblock(x)
        # We recursively reordered these sub-series. You can run the ./utils/recursive_demo.py to emulate this procedure. 
        if self.current_level == 0:
            return self.zip_up_the_pants(x_even_update, x_odd_update)
        else:
            return self.zip_up_the_pants(self.SCINet_Tree_even(x_even_update), self.SCINet_Tree_odd(x_odd_update))

EncoderTree类（编码器）

• 实例化SCINet_Tree类，编码器，让输入进入SCINet_Tree模块

class EncoderTree(nn.Module):
    def __init__(self, in_planes,  num_levels, kernel_size, dropout, groups, hidden_size, INN):
        super().__init__()
        self.levels=num_levels
        self.SCINet_Tree = SCINet_Tree(
            in_planes = in_planes,
            current_level = num_levels-1,
            kernel_size = kernel_size,
            dropout =dropout ,
            groups = groups,
            hidden_size = hidden_size,
            INN = INN)
        
    def forward(self, x):

        # 编码器，让输入进入SCINet_Tree模块
        x= self.SCINet_Tree(x)

        return x

SCINet类（堆叠模型整体架构）

• 在该类中实现了整个模型的搭建，当然也包含架构图的最后一张，stacked堆叠、解码器、RIN激活等等
  

class SCINet(nn.Module):
    def __init__(self, output_len, input_len, input_dim = 9, hid_size = 1, num_stacks = 1,
                num_levels = 3, num_decoder_layer = 1, concat_len = 0, groups = 1, kernel = 5, dropout = 0.5,
                 single_step_output_One = 0, input_len_seg = 0, positionalE = False, modified = True, RIN=False):
        super(SCINet, self).__init__()

        self.input_dim = input_dim
        self.input_len = input_len
        self.output_len = output_len
        self.hidden_size = hid_size
        self.num_levels = num_levels
        self.groups = groups
        self.modified = modified
        self.kernel_size = kernel
        self.dropout = dropout
        self.single_step_output_One = single_step_output_One
        self.concat_len = concat_len
        self.pe = positionalE
        self.RIN=RIN
        self.num_decoder_layer = num_decoder_layer

        self.blocks1 = EncoderTree(
            in_planes=self.input_dim,
            num_levels = self.num_levels,
            kernel_size = self.kernel_size,
            dropout = self.dropout,
            groups = self.groups,
            hidden_size = self.hidden_size,
            INN =  modified)

        if num_stacks == 2: # we only implement two stacks at most.
            self.blocks2 = EncoderTree(
                in_planes=self.input_dim,
            num_levels = self.num_levels,
            kernel_size = self.kernel_size,
            dropout = self.dropout,
            groups = self.groups,
            hidden_size = self.hidden_size,
            INN =  modified)

        self.stacks = num_stacks

        for m in self.modules():
            # 如果m为2维卷积层
            if isinstance(m, nn.Conv2d):
                # 初始化权重
                n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
                m.weight.data.normal_(0, math.sqrt(2. / n))
            elif isinstance(m, nn.BatchNorm2d):
                m.weight.data.fill_(1)
                m.bias.data.zero_()
            elif isinstance(m, nn.Linear):
                m.bias.data.zero_()
        self.projection1 = nn.Conv1d(self.input_len, self.output_len, kernel_size=1, stride=1, bias=False)
        self.div_projection = nn.ModuleList()
        self.overlap_len = self.input_len//4
        self.div_len = self.input_len//6

        # 若解码层大于1
        if self.num_decoder_layer > 1:
            # pro1层变为线性层
            self.projection1 = nn.Linear(self.input_len, self.output_len)
            # 循环range(解码层-1)
            for layer_idx in range(self.num_decoder_layer-1):
                # 创建子模块列表
                div_projection = nn.ModuleList()
                for i in range(6):
                    # 计算全连接层输出维度
                    # 若input_len = 96 --> div_len = 16,overlap_len = 24
                    # len = 24 --> 24 --> 24 --> 24 --> 24 --> 16
                    lens = min(i*self.div_len+self.overlap_len,self.input_len) - i*self.div_len
                    # (24,16) --> (24,16) --> (24,16) --> (24,16) --> (24,16) --> (16,16)
                    div_projection.append(nn.Linear(lens, self.div_len))
                self.div_projection.append(div_projection)

        if self.single_step_output_One: # only output the N_th timestep.
            if self.stacks == 2:
                if self.concat_len:
                    self.projection2 = nn.Conv1d(self.concat_len + self.output_len, 1,
                                                kernel_size = 1, bias = False)
                else:
                    self.projection2 = nn.Conv1d(self.input_len + self.output_len, 1,
                                                kernel_size = 1, bias = False)
        else: # output the N timesteps.
            if self.stacks == 2:
                if self.concat_len:
                    self.projection2 = nn.Conv1d(self.concat_len + self.output_len, self.output_len,
                                                kernel_size = 1, bias = False)
                else:
                    self.projection2 = nn.Conv1d(self.input_len + self.output_len, self.output_len,
                                                kernel_size = 1, bias = False)

        # For positional encoding
        self.pe_hidden_size = input_dim
        if self.pe_hidden_size % 2 == 1:
            self.pe_hidden_size += 1
    
        num_timescales = self.pe_hidden_size // 2
        max_timescale = 10000.0
        min_timescale = 1.0

        log_timescale_increment = (
                math.log(float(max_timescale) / float(min_timescale)) /
                max(num_timescales - 1, 1))
        temp = torch.arange(num_timescales, dtype=torch.float32)
        inv_timescales = min_timescale * torch.exp(
            torch.arange(num_timescales, dtype=torch.float32) *
            -log_timescale_increment)
        self.register_buffer('inv_timescales', inv_timescales)

        ### RIN Parameters ###
        if self.RIN:
            self.affine_weight = nn.Parameter(torch.ones(1, 1, input_dim))
            self.affine_bias = nn.Parameter(torch.zeros(1, 1, input_dim))
    
    def get_position_encoding(self, x):
        # 取数据第2个维度
        max_length = x.size()[1]
        # 位置编码
        position = torch.arange(max_length, dtype=torch.float32, device=x.device)
        # 在第2个维度前面再添加一个维度
        temp1 = position.unsqueeze(1)  # 5 1
        temp2 = self.inv_timescales.unsqueeze(0)  # 1 256
        # 矩阵乘法
        scaled_time = position.unsqueeze(1) * self.inv_timescales.unsqueeze(0)  # 5 256
        # 拼接sin(特征)和cos(特征)
        signal = torch.cat([torch.sin(scaled_time), torch.cos(scaled_time)], dim=1)  #[T, C]
        # pad操作
        signal = F.pad(signal, (0, 0, 0, self.pe_hidden_size % 2))
        # 改变数组维度,并使其称为视图
        signal = signal.view(1, max_length, self.pe_hidden_size)
    
        return signal

    def forward(self, x):
        # 判断输出序列长度合理性
        assert self.input_len % (np.power(2, self.num_levels)) == 0
        # 如果需要位置编码
        if self.pe:
            pe = self.get_position_encoding(x)
            if pe.shape[2] > x.shape[2]:
                x += pe[:, :, :-1]
            else:
                x += self.get_position_encoding(x)

        # 若使用RIN激活
        if self.RIN:
            print('/// RIN ACTIVATED ///\r',end='')
            means = x.mean(1, keepdim=True).detach()
            #mean
            x = x - means
            #var
            stdev = torch.sqrt(torch.var(x, dim=1, keepdim=True, unbiased=False) + 1e-5)
            x /= stdev
            # affine
            # print(x.shape,self.affine_weight.shape,self.affine_bias.shape)
            x = x * self.affine_weight + self.affine_bias

        # 第一层stack
        res1 = x
        # 进入编码器
        x = self.blocks1(x)
        # 相加操作
        x += res1
        # 如果解码层为1
        if self.num_decoder_layer == 1:
            # 经过1维卷积层Conv1d(input_len, output_len, kernel_size = 1),得到结果
            x = self.projection1(x)
        else:
            # 交换维度(B,L,D) --> (B,D,L)
            x = x.permute(0,2,1)

            for div_projection in self.div_projection:
                # 创建与x相同的全0矩阵
                output = torch.zeros(x.shape,dtype=x.dtype).cuda()
                # 取出下标和对应层
                for i, div_layer in enumerate(div_projection):
                    # 赋值对应维度
                    div_x = x[:,:,i*self.div_len:min(i*self.div_len+self.overlap_len,self.input_len)]
                    output[:,:,i*self.div_len:(i+1)*self.div_len] = div_layer(div_x)
                x = output
            # 经过1维卷积层Conv1d(input_len, output_len, kernel_size = 1),得到结果
            x = self.projection1(x)
            # 交换维度(B,L,D) --> (B,D,L)
            x = x.permute(0,2,1)

        # 如果stacks为1
        if self.stacks == 1:
            # 反转RIN激活
            if self.RIN:
                # x - 偏置
                x = x - self.affine_bias
                # x / 权值
                x = x / (self.affine_weight + 1e-10)
                # x * 标准差
                x = x * stdev
                # x + 平均值
                x = x + means
            return x

        # 若stacks为2
        elif self.stacks == 2:
            # 赋值中间层输出
            MidOutPut = x
            # 若concat_len不为0
            if self.concat_len:
                # 将res1(部分)和x在沿1维度进行拼接
                x = torch.cat((res1[:, -self.concat_len:,:], x), dim=1)
            else:
                # 将res1(部分)和x在沿1维度进行拼接
                x = torch.cat((res1, x), dim=1)

            # 第2层stacks
            res2 = x

            # 进入编码层
            x = self.blocks2(x)
            # 加法操作
            x += res2
            # 进入1维卷积Conv1d(output_len, output_len, kernel_size = 1)
            x = self.projection2(x)
            
            # 反转RIN激活
            if self.RIN:
                MidOutPut = MidOutPut - self.affine_bias
                MidOutPut = MidOutPut / (self.affine_weight + 1e-10)
                MidOutPut = MidOutPut * stdev
                MidOutPut = MidOutPut + means

            # 反转RIN激活
            if self.RIN:
                x = x - self.affine_bias
                x = x / (self.affine_weight + 1e-10)
                x = x * stdev
                x = x + means

            # 输出结果以及中间层特征输出
            return x, MidOutPut


def get_variable(x):
    x = Variable(x)
    return x.cuda() if torch.cuda.is_available() else x

• 有一点奇怪的是，在论文中stack可以达到3，但是在该代码中只要stack大于2就会报错，但其实当你读完模型架构以后，你完全可以将这个约束解除，因为我们不需要做实验，所以3层中间的2层不需要输出特征，只要最后一层结果就行。

模型训练（exp_ETTh）

• 这里我主要注释一下train函数，valid和test函数都差不多，只是有些操作不需要删减了而已。

    def train(self, setting):
        # 取得训练、验证、测试数据及数据加载器
        train_data, train_loader = self._get_data(flag = 'train')
        valid_data, valid_loader = self._get_data(flag = 'val')
        test_data, test_loader = self._get_data(flag = 'test')
        path = os.path.join(self.args.checkpoints, setting)

        # 创建模型保存路径
        if not os.path.exists(path):
            os.makedirs(path)

        # 绘制模型训练信息曲线
        writer = SummaryWriter('event/run_ETTh/{}'.format(self.args.model_name))

        # 获取当前时间
        time_now = time.time()

        # 取训练步数
        train_steps = len(train_loader)
        # 设置早停参数
        early_stopping = EarlyStopping(patience=self.args.patience, verbose=True)

        # 选择优化器
        model_optim = self._select_optimizer()
        # 选择损失函数
        criterion =  self._select_criterion(self.args.loss)

        # 如果多GPU并行
        if self.args.use_amp:
            scaler = torch.cuda.amp.GradScaler()

        # 如果断点续传训练
        if self.args.resume:
            self.model, lr, epoch_start = load_model(self.model, path, model_name=self.args.data, horizon=self.args.horizon)
        else:
            epoch_start = 0

        for epoch in range(epoch_start, self.args.train_epochs):
            iter_count = 0
            train_loss = []
            
            self.model.train()
            epoch_time = time.time()
            for i, (batch_x,batch_y,batch_x_mark,batch_y_mark) in enumerate(train_loader):
                iter_count += 1
                
                model_optim.zero_grad()

                # 得到预测值、反归一化预测值、中间层输出、反归一化中间层输出、真实值、反归一化真实值
                pred, pred_scale, mid, mid_scale, true, true_scale = self._process_one_batch_SCINet(
                    train_data, batch_x, batch_y)

                # stacks为1
                if self.args.stacks == 1:
                    # loss损失为mae(真实值+预测值)
                    loss = criterion(pred, true)

                # stacks为2
                elif self.args.stacks == 2:
                    # loss损失为mae(真实值,预测值) + mae(中间层输出,预测值)
                    loss = criterion(pred, true) + criterion(mid, true)
                else:
                    print('Error!')

                # 将loss信息记录到train_loss列表中
                train_loss.append(loss.item())

                # 100个训练步数输出一次训练、验证、测试损失信息
                if (i+1) % 100==0:
                    print("\titers: {0}, epoch: {1} | loss: {2:.7f}".format(i + 1, epoch + 1, loss.item()))
                    speed = (time.time()-time_now)/iter_count
                    left_time = speed*((self.args.train_epochs - epoch)*train_steps - i)
                    print('\tspeed: {:.4f}s/iter; left time: {:.4f}s'.format(speed, left_time))
                    iter_count = 0
                    time_now = time.time()

                # 如果有分布式计算
                if self.args.use_amp:
                    print('use amp')    
                    scaler.scale(loss).backward()
                    scaler.step(model_optim)
                    scaler.update()
                else:
                    # 反向传播
                    loss.backward()
                    # 更新优化器
                    model_optim.step()

            # 打印关键信息
            print("Epoch: {} cost time: {}".format(epoch+1, time.time()-epoch_time))
            train_loss = np.average(train_loss)
            print('--------start to validate-----------')
            valid_loss = self.valid(valid_data, valid_loader, criterion)
            print('--------start to test-----------')
            test_loss = self.valid(test_data, test_loader, criterion)

            print("Epoch: {0}, Steps: {1} | Train Loss: {2:.7f} valid Loss: {3:.7f} Test Loss: {4:.7f}".format(
                epoch + 1, train_steps, train_loss, valid_loss, test_loss))

            # 记录训练、测试、验证集损失下降情况
            writer.add_scalar('train_loss', train_loss, global_step=epoch)
            writer.add_scalar('valid_loss', valid_loss, global_step=epoch)
            writer.add_scalar('test_loss', test_loss, global_step=epoch)

            # 测算早停策略
            early_stopping(valid_loss, self.model, path)
            # 若达到早停标准
            if early_stopping.early_stop:
                print("Early stopping")
                break
            # 更新学习率
            lr = adjust_learning_rate(model_optim, epoch+1, self.args)
        # 保存模型
        save_model(epoch, lr, self.model, path, model_name=self.args.data, horizon=self.args.pred_len)
        # 保存表现最好模型
        best_model_path = path+'/'+'checkpoint.pth'
        # 加载表现最好模型
        self.model.load_state_dict(torch.load(best_model_path))
        # 返回模型
        return self.model

结果展示

• 我用kaggle上的GPU(P100)跑的，时间很短，跑这个ETTh这个数据集需要40分钟左右

>>>>>>>start training : SCINet_ETTh1_ftM_sl96_ll48_pl48_lr0.0001_bs32_hid1_s1_l3_dp0.5_invFalse_itr0>>>>>>>>>>>>>>>>>>>>>>>>>>
train 8497
val 2833
test 2833

	iters: 100, epoch: 41 | loss: 0.3506456
	speed: 0.2028s/iter; left time: 3204.9921s
	iters: 200, epoch: 41 | loss: 0.3641948
	speed: 0.0906s/iter; left time: 1422.0832s
Epoch: 41 cost time: 24.570287466049194
--------start to validate-----------
normed mse:0.5108, mae:0.4747, rmse:0.7147, mape:5.9908, mspe:25702.7811, corr:0.7920
denormed mse:7.2514, mae:1.5723, rmse:2.6928, mape:inf, mspe:inf, corr:0.7920
--------start to test-----------
normed mse:0.3664, mae:0.4001, rmse:0.6053, mape:7.6782, mspe:30989.9618, corr:0.7178
denormed mse:8.2571, mae:1.5634, rmse:2.8735, mape:inf, mspe:inf, corr:0.7178
Epoch: 41, Steps: 265 | Train Loss: 0.3702444 valid Loss: 0.4746509 Test Loss: 0.4000920
	iters: 100, epoch: 42 | loss: 0.3643743
	speed: 0.2015s/iter; left time: 3130.5999s
	iters: 200, epoch: 42 | loss: 0.3464577
	speed: 0.1015s/iter; left time: 1566.1000s
Epoch: 42 cost time: 25.76799440383911
--------start to validate-----------
normed mse:0.5101, mae:0.4743, rmse:0.7142, mape:5.9707, mspe:25459.9669, corr:0.7923
denormed mse:7.2425, mae:1.5713, rmse:2.6912, mape:inf, mspe:inf, corr:0.7923
--------start to test-----------
normed mse:0.3670, mae:0.4010, rmse:0.6058, mape:7.6564, mspe:30790.0708, corr:0.7179
denormed mse:8.2969, mae:1.5701, rmse:2.8804, mape:inf, mspe:inf, corr:0.7179
Epoch: 42, Steps: 265 | Train Loss: 0.3700826 valid Loss: 0.4743312 Test Loss: 0.4009686
	iters: 100, epoch: 43 | loss: 0.3849421
	speed: 0.2019s/iter; left time: 3083.0659s
	iters: 200, epoch: 43 | loss: 0.3757646
	speed: 0.0981s/iter; left time: 1487.8231s
Epoch: 43 cost time: 25.635279893875122
--------start to validate-----------
normed mse:0.5105, mae:0.4744, rmse:0.7145, mape:5.9568, mspe:25381.2960, corr:0.7922
denormed mse:7.2566, mae:1.5721, rmse:2.6938, mape:inf, mspe:inf, corr:0.7922
--------start to test-----------
normed mse:0.3674, mae:0.4014, rmse:0.6061, mape:7.6480, mspe:30700.9283, corr:0.7180
denormed mse:8.3153, mae:1.5732, rmse:2.8836, mape:inf, mspe:inf, corr:0.7180
Epoch: 43, Steps: 265 | Train Loss: 0.3698175 valid Loss: 0.4744163 Test Loss: 0.4013726
Early stopping
>>>>>>>testing : SCINet_ETTh1_ftM_sl96_ll48_pl48_lr0.0001_bs32_hid1_s1_l3_dp0.5_invFalse_itr0<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<
test 2833
normed mse:0.3660, mae:0.3998, rmse:0.6050, mape:7.7062, mspe:31254.7139, corr:0.7174
TTTT denormed mse:8.2374, mae:1.5608, rmse:2.8701, mape:inf, mspe:inf, corr:0.7174
Final mean normed mse:0.3660,mae:0.3998,denormed mse:8.2374,mae:1.5608

• 跑完以后项目文件中会生成两个文件夹，exp文件夹中存放模型文件，后缀名为.pht；event文件夹中有tensorboard记录的loss文件，这里展示一下
  

后记

• 如果大家有自定义项目（跑自己数据）的需求，可以在文章下留言，后期有时间我会专门写一篇SCINet模型如何自定义项目，以及哪些参数需要着重调整，该怎么调等等。