Closed whalegeek closed 2 years ago
Hi @whalegeek, I believe the issue is that your time series is too short (just 4 steps as defined by the window_length). MiniRocket cannot handle sequences shorter than 9 steps. You can either increase the length (if that makes sense) or use any other model that doesn't have this constraint, like the ones you mentioned. I'll make a note to update the documentation clarify this.
Hi @oguiza,
You are absolutely right. I conducted additional tests, using different numbers for time steps, and it confirmed that MiniRocket could only accept time step values of at least 9.
Appreciate your valuable support very much.
There is an issue during the training of a MiniRocket (Pytorch implementation) model when SlidingWindow is used to split data and add time step. However, the training works successful when the df2xy is used to split the data. Nevertheless, it is possible to train model based on architecture like XCM and TST using the same data processed by the SlidingWindow. Here is the code used (I can provide additional information if required):