jingw2/demand_forecast

Deep Demand Forecast Models

Pytorch Implementation of DeepAR, MQ-RNN, Deep Factor Models, LSTNet, and TPA-LSTM. Furthermore, combine all these model to deep demand forecast model API.

Requirements

Please install Pytorch before run it, and

pip install -r requirements.txt

Run tests

# DeepAR
python deepar.py -e 100 -spe 3 -nl 1 -l g -not 168 -sp -rt -es 10 -hs 50  -sl 60 -ms

# MQ-RNN
python mq_rnn.py -e 100 -spe 3 -nl 1 -sp -sl 72 -not 168 -rt -ehs 50 -dhs 20 -ss -es 10 -ms

# Deep Factors
python deep_factors.py -e 100 -spe 3 -rt -not 168 -sp -sl 168 -ms

# TPA-LSTM
python tpa_lstm.py -e 1000 -spe 1 -nl 1 -not 168 -sl 30 -sp -rt -max

DeepAR

MQ-RNN

Deep Factors

TPA-LSTM

Arguments

Arguments	Details
-e	number of episodes
-spe	steps per episode
-sl	sequence length
-not	number of observations to train
-ms	mean scaler on y
-max	max scaler on y
-nl	number of layers
-l	likelihood to select, "g" or "nb"
-rt	run test data
-sample_size	sample size to sample after training in deep factors/deepar, default 100

TO DO

• Deep Factor Model
• TPA-LSTM pytorch
• LSTNet pytorch
• Debug Uber Extreme forcaster
• Modeling Extreme Events in TS
• Intermittent Demand Forecasting
• Model API

Demand Forecast Dataset Resources

• Solar Energy
• Traffic
• Electricity
• MuseDATA

Reference

• DeepAR
• MQ-RNN
• Deep Factor
• LSTNet
• TPA-LSTM
• LSTNet Github
• TPA-LSTM Github
• Uber Extreme Event Forecast 1
• Uber Extreme Event Forecast 2
• Modeling Extreme Events in Time Series Prediction

demand_forecast's Issues

请问下gaussian_likelihood_loss的原理是什么？

请问大佬，这个损失定义是什么意思啊：gaussian_likelihood_loss。 为什么用y值带入正态分布的密度函数，再求log的mean就是损失啊？

请问tpa-lstm中在训练的时候使用的util.generator的数据，为什么没有用X_train和Xf

num_periods，seq_len，num_obs_to_train这几个参数具体是什么意思呀
输入generator的X,y。是不是X指一个数段的数据，而y是对应的预测目标呢？

tpa_lstm 问题

在运行 python tpa_lstm.py -e 1 -spe 3 -nl 3 -not 200 -sl 100 -sp -rt -max的时候
-not 200 -sl 100 训练集+测试集大于200的时候会无法运行
报错
RuntimeError: Calculated padded input size per channel: (99 x 24). Kernel size: (199 x 1). Kernel size can't be greater than actual input size
导入的date长度为600但是无法使用更多的数据进行训练，求解，谢谢你！。

请问DeepFactor和DeepAR如何构造多条时间序列样本

DeepAR和Deep Factor可以学习多条时间序列之间的关联关系，但在样本构造上还是(n_samples, n_period, n_feature)吗

DeepAR模型训练LSTM层输入有问题？

这里输入到lstm的样本inp维度是 (1,1,12)? 那相当于喂给lstm的步长就是1？ 那这不就是一个全连接网络了吗，没有用到RNN的特性啊。是不是有哪里搞错了 。 我参数都是按照readme里面输入的

intertiment datesets for

How to guarantee the relative size between different quantiles in MQ-RNN?

There is a fixed relationship between different quantiles. For example, P90 should be bigger than P10. So how to guarantee this relationship in MQ-RNN, where different quantiles are forecasted at the same time?

DeepAR里面为什么是对y进行embedding啊？

论文里面好像是说对类别特征进行embedding，但是这里代码好像是把时序观察值进行了embedding？望大佬解答。

请问一下，为什么DeepAR中出来的sigma在softplus后要加1e-6

Consult and ask for help

Hello, jingw2：
I use matlab 2020 official deep learning example code (LSTM) to do time series data prediction, the result is that I have not learned the characteristics of the amplitude change of the original signal, but learned the quasi periodic characteristics, as shown in the figure:
overview picture

Details of the picture

Is this pure signal data time series suitable for deep learning prediction?

     I  want to improve the prediction accuracy，For example, the prediction accuracy is increased from RMSE = 289 to RMSE = 220？Please give me your advice，thank you very much！

best wishes,
Mr.Wang

question for TPA-LSTM source code

demand_forecast/tpa_lstm.py

Line 150 in 1978fef

ypred = model(ytrain)

In this line, I think that Xtrain should be used instead of ytrain. Am I wrong?
I could not have found any source code for using multi-variate features in TPA-LSTM code.

TPA-LSTM only one input feature

TPA-LSTM looks to only support one input feature. Would you like me to update it to enable multiple features? Seems important for a complete implementation unless I'm mistaken on the usage of your code.

LSTNet pytorch版本更新了嘛？

博主 LSTNet pytorch版本更新了嘛？有没有一维的代码

您好，请教下，这是时间序列预测吗？

您好，请教下，这是时间序列预测吗？做需求预测

多时间序列Deepar怎么处理呢

看了下deepar的代码，单时间序列能正常处理，但是多时间序列怎么处理呢

请问这些代码都是linux环境下的吗？

关于高斯分布的负对数似然函数

negative_likelihood = torch.log(sigma + 1) + (z - mu) ** 2 / (2 * sigma ** 2) + 6
新人求教，高斯分布的负对数似然函数的计算公式中的6是怎么来的？sigma为什么要+1？谢谢

AttributeError: module 'util' has no attribute 'MaxScaler'

Traceback (most recent call last):
File "tpa_lstm.py", line 249, in
losses, mape_list = train(X, y, args)
File "tpa_lstm.py", line 134, in train
yscaler = util.MaxScaler()
AttributeError: module 'util' has no attribute 'MaxScaler'

intertiment issunes and fixes

replace like this

@staticmethod
    def multistage(X, y, forecast_period, num_sample_points=200):
        glm = multistage_glm.MultiStageGLM()
        Xtrain = X[:-forecast_period]
        ytrain = y
        Xtest = X[-forecast_period:]
        #######      ytest = y[-forecast_period:]
        print(Xtrain.shape,ytrain.shape)
        glm.fit(Xtrain, ytrain)
        yfit = glm.predict(Xtrain)
        ypred_list = []
        for _ in range(num_sample_points):
            ypred = glm.predict(Xtest)
            ypred_list.append(ypred.reshape((-1, 1)))
        ypred = np.concatenate(ypred_list, axis=1)
        
        p50 = np.quantile(ypred, 0.5, axis=1)
        p25 = np.quantile(ypred, 0.25, axis=1)
        p90 = np.quantile(ypred, 0.90, axis=1)

        return yfit, p50

tpa-lstm怎么处理变长的batch

楼主的代码中obs_len明确规定了每批次输入数据的长度是固定的，很多任务的时序数据为变长的，这时该怎么处理呢？由于长度为变长，卷积核的大小就无法固定，也就无法在 time step上 做卷积

DeepAR model raises erros when batch size>1

DeepAR model raises erros when batch size>1, I think the code needs to be revised

quantile loss是不是错了？

针对MQRNN的问题：
相关代码：

ypred_rho - yf是不是应该改成yf - ypred_rho ？同时，按照quantile loss公式不应该是不同horizon先累加，然后平均么？我看你后面直接loss.mean了，相当于不同horizon也平均了。
还有一个问题，对于某一轮来说，如果采样一次就进行一次loss.backward()，那args.step_per_epoch 这个参数的意义何在？直接加大args.num_epoches的值不就好了？我在实现的过程中，发现如果不同args.step_per_epoch的loss做avg，然后整体进行一次loss.backward()效果也还可以。不知道你对这个怎么看？

请问pytorch版本的tpa-lstm完成了吗？

请问pytorch版本的tap-lstm完成了吗，最近在学习tpa-lstm, pytorch版本对我或许比较友好一点，请问你弄完了吗？如果上传了，感激不尽。

MQRNN为什么没有采用forking-sequences

当前测试的结果MQRNN的性能不如DeepAR，我猜测是没有采用forking-sequences的原因

DeepAR Xf在预测的时候怎么用

如果我没理解做的话，xf是test phase的X。我看你的实现里不区分原论文的zt和xt。zt是原始序列，如果把zt和xt合并起来，test phase时的zt就是target。 所以xf_test 里面有y_test， 实际上拿不到，也不能输入。请问怎么解决？或者如果我理解有误请指正。

The following error occurred when Run 'python deep_factors.py -e 100 -spe 3 -rt -not 168 -sp -sl 168 -ms',why ?

demand_forecast-master>python deep_factors.py -e 100 -spe 3 -rt -not 168 -sp -sl 168 -ms
Traceback (most recent call last):
File "C:\Anaconda\lib\site-packages\pandas\core\arrays\datetimelike.py", line 932, in _cmp_method
other = self._validate_comparison_value(other)
File "C:\Anaconda\lib\site-packages\pandas\core\arrays\datetimelike.py", line 451, in _validate_comparison_value
raise InvalidComparison(other)
pandas.core.arrays.datetimelike.InvalidComparison: 2014-01-01

During handling of the above exception, another exception occurred:

Traceback (most recent call last):
File "deep_factors.py", line 259, in
data = data.loc[(data["date"] >= date(2014, 1, 1)) & (data["date"] <= date(2014, 3, 1))]
File "C:\Anaconda\lib\site-packages\pandas\core\ops\common.py", line 65, in new_method
return method(self, other)
File "C:\Anaconda\lib\site-packages\pandas\core\arraylike.py", line 49, in ge
return self._cmp_method(other, operator.ge)
File "C:\Anaconda\lib\site-packages\pandas\core\series.py", line 4948, in _cmp_method
res_values = ops.comparison_op(lvalues, rvalues, op)
File "C:\Anaconda\lib\site-packages\pandas\core\ops\array_ops.py", line 229, in comparison_op
res_values = op(lvalues, rvalues)
File "C:\Anaconda\lib\site-packages\pandas\core\ops\common.py", line 65, in new_method
return method(self, other)
File "C:\Anaconda\lib\site-packages\pandas\core\arraylike.py", line 49, in ge
return self._cmp_method(other, operator.ge)
File "C:\Anaconda\lib\site-packages\pandas\core\arrays\datetimelike.py", line 934, in _cmp_method
return invalid_comparison(self, other, op)
File "C:\Anaconda\lib\site-packages\pandas\core\ops\invalid.py", line 34, in invalid_comparison
raise TypeError(f"Invalid comparison between dtype={left.dtype} and {typ}")
TypeError: Invalid comparison between dtype=datetime64[ns] and date

(demand) F:\MsVsPython\ImportPytorchNet\jingw2-DeepAR\11\demand_forecast-master>

Help!

@QuanhanSun，你好！请问你是在windows还是linux环境下跑通这些程序的？

生成batch是有放回抽样的

您好，您代码里batch生成的方式是直接sample了一些样本，没有记录index，这样是不是没法保证所有样本都参与到训练中？另外每个epcho跑多少个batch怎么确定呢？

Forking-Sequence in MQ-RNN

Hi,

Very nice project!

针对MQ-RNN，论文在「3.3. Training Scheme」这个章节，提到了「Forking-Sequence」的训练方法，但是代码中好像并木有相关的体现。

因此，想请问下，这部分是在哪里实现的呢？或者 我有什么遗漏的地方吗？

AttributeError: module 'util' has no attribute 'batch_generator'

Traceback (most recent call last): |
File "tpa_lstm.py", line 249, in
losses, mape_list = train(X, y, args)
File "tpa_lstm.py", line 145, in train
Xtrain, ytrain, Xf, yf = util.batch_generator(Xtr, ytr, obs_len, seq_len, args.batch_size)
AttributeError: module 'util' has no attribute 'batch_generator'