yfzhang114 / onenet

This is an official PyTorch implementation of the NeurIPS 2023 paper 《OneNet: Enhancing Time Series Forecasting Models under Concept Drift by Online Ensembling》

Hi there, nice to meet you! 👋

My GitHub Stats

onenet's Issues

Some basic questions to ask

Hello, I am a freshman. I have read your literature and main code and have some basic questions :

1. May I ask if in your experiments, you include a comparison between the results of applying online learning and not applying online learning?
2. for mitigating conceptual drift, did you try in your experiments: the comparison of periodically updating the model after training then predicting and direct prediction without updating the model?
   I hope you could discuss this problem with me.
   Looking forward to your reply. Thank you.

upload files

Your work is excellent and very helpful to me. Can you upload the MOE and Average files in the ensembling method?

Regarding the results of the paper

Thank you for your contribution. But when I tried to reproduce your code multiple times, I got results different from the article. When the prediction step size was 1 in the ETTh2 dataset, I got a result of 0.660 0.435. When the prediction step size was 24 in the WTH dataset, I got a result of 0.192 0.272. When I ran it on the ECL dataset, I got a result of 2.394 0.273, 2.324 0.360, 3.089 0.382. What is the reason?

Do you notice the phenomenon of memory leak in the code of FSNet

I've fully experimented with their code taking up CPU memory issues. It is evident that the CPU memory decreases gradually when the program is running. Fortunately, their process will not be killed because the datasets in their experiment are too small.
However, I tested their code on a dataset with 100,000 entries. And the process was killed when we got to one tenth of the way through the test because of memory leak.
Long data series are not uncommon in the real world of online learning, and online learning is geared towards applications in the display world. Therefore, we are supposed to take this issue seriously.
I have asked the author of the FSNet but got no reply. Considering the in-depth research you have conducted on time series online learning， I hope you could discuss this problem with me.
Looking forward to your reply.

The dataset division of ETTm1

I notice that ETTm1 is divided in the same way as ETTh2 in your code and FSNet. However, the observations of them are recorded hourly and in 15-minutes intervals respectively, which means you only used a quarter of the data from ETTm1 for your experiment. I think that's unreasonable.

A hyperparameter setting question about d3a

I noticed that the hyperparameter var_weight is not set when using the d3a script, which results in the augmented data being the same as the unaugmented data when doing data augmentation?Looking forward to it!Looking forward to your answer

这个地方Pandas和np改了一点，float和 dt.isocalendar().week.to_numpy()

if timeenc == 2:
    dt = pd.to_datetime(dates.date.values)
    return np.stack([
        dt.minute.to_numpy(),
        dt.hour.to_numpy(),
        dt.dayofweek.to_numpy(),
        dt.day.to_numpy(),
        dt.dayofyear.to_numpy(),
        dt.month.to_numpy(),
        dt.isocalendar().week.to_numpy()
    ], axis=1).astype(float)