This is the official PyTorch implementation of Reversible Instance Normalization for Accurate Time-Series Forecasting against Distribution Shift.
Statistical properties such as mean and variance often change over time in time series, i.e., time-series data suffer from a distribution shift problem. This change in temporal distribution is one of the main challenges that prevent accurate time-series forecasting. To address this issue, we propose a simple yet effective normalization method called reversible instance normalization (RevIN), a generally-applicable normalization-and-denormalization method with learnable affine transformation. The proposed method is symmetrically structured to remove and restore the statistical information of a time-series instance, leading to significant performance improvements in time-series forecasting, as shown in Fig. 1. We demonstrate the effectiveness of RevIN via extensive quantitative and qualitative analyses on various real-world datasets, addressing the distribution shift problem.
The code was developed using python 3.8 on Ubuntu 18.04.
The experiments were performed on a single NVIDIA TITAN RTX or NVIDIA TITAN Xp.
- Install PyTorch >= v.1.8.0 following the official instruction.
- tested with PyTorch v.1.8.0 and PyTorch v.1.11.0.
- Clone this repository:
git clone https://github.com/ts-kim/RevIN.git
RevIN calculates the mean and standard deviation of each feature separately for each sequence in a mini-batch.
To be reversible, the input and output tensors should have the same number of features. The input tensors should be provided as (..., feature). For example,
- x_in: (batch_size, sequence_length, num_features)
- x_out: (batch_size, prediction_length, num_features)
RevIN can be added in any arbitrarily chosen layers of a model as follows:
>>> from RevIN import RevIN
>>> revin_layer = RevIN(num_features)
>>> x_in = revin_layer(x_in, 'norm')
>>> x_out = blocks(x_in) # your model or subnetwork within the model
>>> x_out = revin_layer(x_out, 'denorm')
We updated the training and evaluation codes for the baselines, Informer and SCINet.
Please go to the baselines forder or the link below.
If you find this work or code is helpful in your research, please cite:
@inproceedings{kim2021reversible,
title = {Reversible Instance Normalization for Accurate Time-Series Forecasting against Distribution Shift},
author = {Kim, Taesung and
Kim, Jinhee and
Tae, Yunwon and
Park, Cheonbok and
Choi, Jang-Ho and
Choo, Jaegul},
booktitle = {International Conference on Learning Representations},
year = {2021},
url = {https://openreview.net/forum?id=cGDAkQo1C0p}
}