Open TanCari opened 6 years ago
For time series forecasting, how you define your problem and preprocess you data makes a big difference. For example how many timesteps/lag variables do you want to look back for prediction? How far ahead do you want to make a prediction? Are you going to deseasonalize your data or not? Will take the moving average when making a prediction multiple days ahead?
I don't think Tensor2Tensor is equipped to preprocess time series data like this. Using Keras may be a lot simpler, than using one of T2T's models. From my understanding, their default hyperparameters are for the datasets they supply, not ones you upload yourself.
Take a look at this article, which gives an example of making a time series problem with keras, and preprocessing data https://machinelearningmastery.com/time-series-prediction-lstm-recurrent-neural-networks-python-keras/
Hi TanCari, T2T team just added a time series data generator two days ago and there is a transformer parameter set for time series. Maybe it support time series prediction now.
@theJiangYu Hello. I see your comment about the time series data generator but i am unable to find it despite extensively search it. Can you add more help for the same.
Hi guys, congrats and thanks for development and help.
I would like to apply the most modern deep-learning techniques for analysis of sequential data to TIMESERIES coming from physical and chemical phenomena. So, no image processing and no automatic translation. Input and output are just a bunch of real numbers at every time step (the same number of them at every step). Ouputs depend only on the previous story of inputs.
My question: am I right at tensor2tensor or is this mainly thought for image and language processing? In case: is any simple example of subclassing "problem" for this simpler case available?
Thanx a lot for your precious help.