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IDparser.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IDparser.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_sizeparser.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 sizeparser.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 blockparser.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 = INNself.kernel_size = kernelself.dilation = 1self.dropout = dropoutself.hidden_size = hidden_sizeself.groups = groups# 如果通道数为偶数if self.kernel_size % 2 == 0:# 1 * (kernel -2) // 2 + 1pad_l = self.dilation * (self.kernel_size - 2) // 2 + 1 #by default: stride==1# 1 * kernel // 2 + 1pad_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 = 2self.splitting = splittingself.split = Splitting()modules_P = []modules_U = []modules_psi = []modules_phi = []prev_size = 1size_hidden = self.hidden_sizemodules_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0if 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_levelself.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0if 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, Ddef 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_levelsself.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 xSCINet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_dimself.input_len = input_lenself.output_len = output_lenself.hidden_size = hid_sizeself.num_levels = num_levelsself.groups = groupsself.modified = modifiedself.kernel_size = kernelself.dropout = dropoutself.single_step_output_One = single_step_output_Oneself.concat_len = concat_lenself.pe = positionalEself.RIN=RINself.num_decoder_layer = num_decoder_layerself.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_stacksfor 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_channelsm.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//4self.div_len = self.input_len//6# 若解码层大于1if 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 --> 16lens = 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 encodingself.pe_hidden_size = input_dimif self.pe_hidden_size % 2 == 1:self.pe_hidden_size += 1num_timescales = self.pe_hidden_size // 2max_timescale = 10000.0min_timescale = 1.0log_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 1temp2 = 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 signaldef 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()#meanx = x - means#varstdev = 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stackres1 = x# 进入编码器x = self.blocks1(x)# 相加操作x += res1# 如果解码层为1if 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1if 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 + meansreturn x# 若stacks为2elif self.stacks == 2:# 赋值中间层输出MidOutPut = x# 若concat_len不为0if 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stacksres2 = 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_biasMidOutPut = MidOutPut / (self.affine_weight + 1e-10)MidOutPut = MidOutPut * stdevMidOutPut = MidOutPut + means# 反转RIN激活if self.RIN:x = x - self.affine_biasx = x / (self.affine_weight + 1e-10)x = x * stdevx = x + means# 输出结果以及中间层特征输出return x, MidOutPutdef 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 = 0for epoch in range(epoch_start, self.args.train_epochs):iter_count = 0train_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 += 1model_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1if self.args.stacks == 1:# loss损失为mae(真实值+预测值)loss = criterion(pred, true)# stacks为2elif 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_countleft_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 = 0time_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 8497val 2833test 2833iters: 100, epoch: 41 | loss: 0.3506456speed: 0.2028s/iter; left time: 3204.9921siters: 200, epoch: 41 | loss: 0.3641948speed: 0.0906s/iter; left time: 1422.0832sEpoch: 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:0.7920denormed 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:0.7178denormed mse:8.2571, mae:1.5634, rmse:2.8735, mape:inf, mspe:inf, corr:0.7178Epoch: 41, Steps: 265 | Train Loss: 0.3702444 valid Loss: 0.4746509 Test Loss: 0.4000920iters: 100, epoch: 42 | loss: 0.3643743speed: 0.2015s/iter; left time: 3130.5999siters: 200, epoch: 42 | loss: 0.3464577speed: 0.1015s/iter; left time: 1566.1000sEpoch: 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:0.7923denormed 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:0.7179denormed mse:8.2969, mae:1.5701, rmse:2.8804, mape:inf, mspe:inf, corr:0.7179Epoch: 42, Steps: 265 | Train Loss: 0.3700826 valid Loss: 0.4743312 Test Loss: 0.4009686iters: 100, epoch: 43 | loss: 0.3849421speed: 0.2019s/iter; left time: 3083.0659siters: 200, epoch: 43 | loss: 0.3757646speed: 0.0981s/iter; left time: 1487.8231sEpoch: 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:0.7922denormed 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:0.7180denormed mse:8.3153, mae:1.5732, rmse:2.8836, mape:inf, mspe:inf, corr:0.7180Epoch: 43, Steps: 265 | Train Loss: 0.3698175 valid Loss: 0.4744163 Test Loss: 0.4013726Early stopping>>>>>>>testing : SCINet_ETTh1_ftM_sl96_ll48_pl48_lr0.0001_bs32_hid1_s1_l3_dp0.5_invFalse_itr0<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<test 2833normed mse:0.3660, mae:0.3998, rmse:0.6050, mape:7.7062, mspe:31254.7139, corr:0.7174TTTT denormed mse:8.2374, mae:1.5608, rmse:2.8701, mape:inf, mspe:inf, corr:0.7174Final 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。