search

DAMO-DI-ML / NeurIPS2023-One-Fits-All

The official code for "One Fits All: Power General Time Series Analysis by Pretrained LM (NeurIPS 2023 Spotlight)"
459 stars 64 forks source link

Clarification on the Scope of RevIN in Time-Series Applications #5

Open blacksnail789521 opened 11 months ago

blacksnail789521 commented 11 months ago

Dear Authors,

I recently read your paper where you discuss the implementation of RevIN in your model. My understanding is that RevIN is primarily designed for time-series forecasting tasks. In reviewing your source code, I noticed that RevIN is not utilized in classification tasks. Could you please clarify if RevIN is intended exclusively for time-series forecasting, or if it can be applied to other time-series tasks as well? This distinction wasn't explicitly addressed in your paper, and I believe a clarification would be beneficial.

Thank you for your time and for the insights your paper offers.

tianzhou2011 commented 11 months ago

RevIN can be applied to a wide range of regression-based tasks. It involves computing the mean and variance of the input data, applying normalization, and subsequently denormalizing the output. The underlying assumption is that the input distribution is either identical to or possesses similar mean and variance as the output distribution. Consequently, when it comes to classification tasks, utilizing RevIN does not yield meaningful results. This is due to the fact that the output, being a label, does not originate from the same distribution as the input time series sequence.

On Mon, Nov 6, 2023 at 10:34 PM blacksnail789521 @.***> wrote:

Dear Authors,

I recently read your paper where you discuss the implementation of RevIN in your model. My understanding is that RevIN is primarily designed for time-series forecasting tasks. In reviewing your source code, I noticed that RevIN is not utilized in classification tasks. Could you please clarify if RevIN is intended exclusively for time-series forecasting, or if it can be applied to other time-series tasks as well? This distinction wasn't explicitly addressed in your paper, and I believe a clarification would be beneficial.

Thank you for your time and for the insights your paper offers.

— Reply to this email directly, view it on GitHub https://github.com/DAMO-DI-ML/NeurIPS2023-One-Fits-All/issues/5, or unsubscribe https://github.com/notifications/unsubscribe-auth/AB3JGO5S32PSWK4Q6A6BPBTYDDYQPAVCNFSM6AAAAAA67TTHI6VHI2DSMVQWIX3LMV43ASLTON2WKOZRHE3TSMZRHE3DANA . You are receiving this because you are subscribed to this thread.Message ID: @.***>

blacksnail789521 commented 11 months ago

Thank you for your detailed response. From your explanation, it appears that the key to effectively applying RevIN is the similarity in distribution between the input and output. This similarity enables the "reversal" of instance normalization. Given this, would it be correct to say that RevIN is only suitable for tasks where the input and output share similar distributions, primarily seen in time-series forecasting? I'm trying to understand if this characteristic limits its applicability primarily to forecasting scenarios. Please don't hesitate to correct me if my understanding is not accurate.