开源项目
说明
读完代码解析篇,我们针对开源项目中的模型预测方法做一下介绍。作者在Github上给出了模型预测方法以及Q、K图的做法,这里提供下载链接首先,在不更改任何参数的情况下跑完代码,会在项目文件夹中生成两个子文件夹checkpoints
文件夹中包含训练完成的模型,后缀名为.pth
,该模型文件包含完整的模型架构与各层权重,可以通过torch.load
函数加载模型results
文件夹中包含metrics.npy
、pred.npy
、true.npy
三个文件,pred.npy
表示模型预测值,true.npy
表示序列真实值我们可以先将pred.npy
与true.npy
文件作图进行对比,观察模型效果 setting = 'informer_power data_ftMS_sl96_ll48_pl24_dm512_nh8_el2_dl1_df2048_atprob_fc5_ebtimeF_dtTrue_mxTrue_exp_0'pred = np.load('./results/'+setting+'/pred.npy')true = np.load('./results/'+setting+'/true.npy')print(pred.shape)print(true.shape)import matplotlib.pyplot as pltplt.figure()plt.plot(true[0,:,-1], label='GroundTruth')plt.plot(pred[0,:,-1], label='Prediction')plt.legend()plt.show()
输出:
'--do_predict'
由'store_true'
变为'store_false'
,这样在代码运行完以后results
文件夹中会多出一个文件real_prediction.npy
,该文件中即是模型预测的序列值。第2种:在模型训练完以后(在jupyter notebook
)中使用exp.predict(setting, True)
得到预测值输出:
![请添加图片描述](http://zhangshiyu.com/zb_users/upload/2023/03/20230330185706168017382620474.png)
在Kaggle上使用
在Kaggle上我们可以使用免费的GPU来训练模型,达到加速效果,这里把作者的代码解析一遍。下载项目文件
!git clone https://github.com/zhouhaoyi/Informer2020.git!git clone https://github.com/zhouhaoyi/ETDataset.git!ls
导入包
from utils.tools import dotdictimport matplotlib.pyplot as pltfrom exp.exp_informer import Exp_Informerimport torch
参数传导
这个地方与在pycharm中不一样,请注意!args = dotdict()args.model = 'informer' # model of experiment, options: [informer, informerstack, informerlight(TBD)]args.data = 'ETTh1' # dataargs.root_path = './ETDataset/ETT-small/' # root path of data fileargs.data_path = 'ETTh1.csv' # data fileargs.features = 'M' # forecasting task, options:[M, S, MS]; M:multivariate predict multivariate, S:univariate predict univariate, MS:multivariate predict univariateargs.target = 'OT' # target feature in S or MS taskargs.freq = 'h' # 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 3hargs.checkpoints = './informer_checkpoints' # location of model checkpointsargs.seq_len = 96 # input sequence length of Informer encoderargs.label_len = 48 # start token length of Informer decoderargs.pred_len = 24 # prediction sequence length# Informer decoder input: concat[start token series(label_len), zero padding series(pred_len)]args.enc_in = 7 # encoder input sizeargs.dec_in = 7 # decoder input sizeargs.c_out = 7 # output sizeargs.factor = 5 # probsparse attn factorargs.d_model = 512 # dimension of modelargs.n_heads = 8 # num of headsargs.e_layers = 2 # num of encoder layersargs.d_layers = 1 # num of decoder layersargs.d_ff = 2048 # dimension of fcn in modelargs.dropout = 0.05 # dropoutargs.attn = 'prob' # attention used in encoder, options:[prob, full]args.embed = 'timeF' # time features encoding, options:[timeF, fixed, learned]args.activation = 'gelu' # activationargs.distil = True # whether to use distilling in encoderargs.output_attention = False # whether to output attention in ecoderargs.mix = Trueargs.padding = 0args.freq = 'h'args.batch_size = 32 args.learning_rate = 0.0001args.loss = 'mse'args.lradj = 'type1'args.use_amp = False # whether to use automatic mixed precision trainingargs.num_workers = 0args.itr = 1args.train_epochs = 6args.patience = 3args.des = 'exp'args.use_gpu = True if torch.cuda.is_available() else Falseargs.gpu = 0args.use_multi_gpu = Falseargs.devices = '0,1,2,3'
检查GPU
args.use_gpu = True if torch.cuda.is_available() and args.use_gpu else Falseif args.use_gpu and args.use_multi_gpu: args.devices = args.devices.replace(' ','') device_ids = args.devices.split(',') args.device_ids = [int(id_) for id_ in device_ids] args.gpu = args.device_ids[0]
定义数据加载
# Set augments by using data namedata_parser = { '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]},}if args.data in data_parser.keys(): data_info = data_parser[args.data] args.data_path = data_info['data'] args.target = data_info['T'] args.enc_in, args.dec_in, args.c_out = data_info[args.features]
训练模型
args.detail_freq = args.freqargs.freq = args.freq[-1:]Exp = Exp_Informerfor ii in range(args.itr): # setting record of experiments setting = '{}_{}_ft{}_sl{}_ll{}_pl{}_dm{}_nh{}_el{}_dl{}_df{}_at{}_fc{}_eb{}_dt{}_mx{}_{}_{}'.format(args.model, args.data, args.features, args.seq_len, args.label_len, args.pred_len, args.d_model, args.n_heads, args.e_layers, args.d_layers, args.d_ff, args.attn, args.factor, args.embed, args.distil, args.mix, args.des, ii) # set experiments exp = Exp(args) # train print('>>>>>>>start training : {}>>>>>>>>>>>>>>>>>>>>>>>>>>'.format(setting)) exp.train(setting) # test print('>>>>>>>testing : {}<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<'.format(setting)) exp.test(setting) torch.cuda.empty_cache()
到这里模型训练完毕 预测
运行完会在results
文件夹中生成real_prediction.npy
序列预测数据 import ossetting = 'informer_ETTh1_ftM_sl96_ll48_pl24_dm512_nh8_el2_dl1_df2048_atprob_fc5_ebtimeF_dtTrue_mxTrue_exp_0'exp = Exp(args)exp.predict(setting, True)prediction = np.load('./results/'+setting+'/real_prediction.npy')plt.figure()plt.plot(prediction[0,:,-1], label='Prediction')plt.legend()plt.show()
自定义项目
数据处理
首先将数据集放入项目data
文件夹中,然后修改'--data'
、'--root_path'
、'--data_path'
、'--data_path'
参数。比如在这里,我使用的是power data.csv
数据集(下载链接),那么我的参数应该修改成这样: parser.add_argument('--data', type=str, default='power data')parser.add_argument('--root_path', type=str, default='./data/')parser.add_argument('--data_path', type=str, default='power data.csv')
然后修改预测类型,这里我使用多变量预测单变量,所以--features
参数选择MS
;预测的变量名称为总有功功率
,所以修改'--target'
参数;时间采样为15分钟,所以将'--freq'
参数改为t
即: parser.add_argument('--features', type=str, default='MS')parser.add_argument('--target', type=str, default='总有功功率')parser.add_argument('--freq', type=str, default='t')
我想序列长度为192(2天),有标签预测序列长度为96(1天),无标签预测序列长度为48(半天),修改参数'--seq_len'
、'--label_len'
、'--pred_len'
为: parser.add_argument('--seq_len', type=int, default=192)parser.add_argument('--label_len', type=int, default=96)parser.add_argument('--pred_len', type=int, default=48)
接下来跳到数据处理方式data_parser
,参照项目原有处理方式进行填写: data_parser = {'power data':{'data':'power data.csv','T':'总有功功率','M':[5,5,5],'S':[1,1,1],'MS':[5,5,1]},}
找到exp_informer.py
文件,_get_data
类中data_dict
参数,使用Dataset_Custom
对数据进行处理,即: data_dict = {'power data':Dataset_Custom,}
到这里数据处理修改结束 模型参数
这里我希望模型运行结束后生成预测序列文件,所以我将'--do_predict'
参数修改为'store_false'
,即: parser.add_argument('--do_predict', action='store_false')
其他参数,比如学习率等等,大家可以根据任务要求自行修改,这里我就不再赘述了。