thuml / iTransformer

Official implementation for "iTransformer: Inverted Transformers Are Effective for Time Series Forecasting" (ICLR 2024 Spotlight), https://openreview.net/forum?id=JePfAI8fah
https://arxiv.org/abs/2310.06625
MIT License
1.25k stars 218 forks source link

How iTransformer could training blazingly fast? #136

Open ikhsansdqq opened 2 weeks ago

ikhsansdqq commented 2 weeks ago

Hi, thanks for sharing the iTransformer code earlier. I’ve made several adjustments to make it more flexible for my research. I’m curious, how is the iTransformer able to train so quickly with univariate time-series data? On average, it only takes around 20 epochs and less than 10 seconds of training time. Could you explain how that's possible?

I used iTransformer external libraries from @lucidrains

Please take a look at my epoch training size, I've set the maximum epochs of 500.

image

WenWeiTHU commented 1 week ago

For univariate time series, the token number of iTransformer is only one. It's just like an MLP composed of FFNs.