tianzhou2011 / film Goto Github PK

License: MIT License

Python 75.37% Shell 24.63%

film's Introduction

FiLM

FiLM: Frequency improved Legendre Memory Model for Long-term Time Series Forecasting https://arxiv.org/abs/2205.08897

In long-term forecasting, FiLM achieves SOTA, with a 19% relative improvement on six benchmarks, covering five practical applications: energy, traffic, economics, weather and disease.


Figure 1. Overall structure of FiLM


Figure 2. Frequency Enhanced Layer (FEL)	Figure 3. Legendre Projection Unit (LPU)

Main Results

Get Started

Install Python 3.9, PyTorch 1.11.0.
Download data. You can obtain all the six benchmarks from Tsinghua Cloud or Google Drive. All the datasets are well pre-processed and can be used easily.
Train the model. We provide the experiment scripts of all benchmarks under the folder ./scripts. You can reproduce the Multivariate/Univariate experiment results by:

bash ./script/ETT_script/FiLM/FiLM_ETTm2.sh
bash ./script/ECL_script/FiLM/FiLM.sh
bash ./script/Exchange_script/FiLM/FiLM.sh
bash ./script/Traffic_script/FiLM/FiLM.sh
bash ./script/Weather_script/FiLM/FiLM.sh
bash ./script/ILI_script/FiLM/FiLM.sh


bash ./script/ETT_script/FiLM/FiLM_ETTm2_S.sh
bash ./script/ECL_script/FiLM/FiLM_S.sh
bash ./script/Exchange_script/FiLM/FiLM_S.sh
bash ./script/Traffic_script/FiLM/FiLM_S.sh
bash ./script/Weather_script/FiLM/FiLM_S.sh
bash ./script/ILI_script/FiLM/FiLM_S.sh

Acknowledgement

We appreciate the following github repos a lot for their valuable code base or datasets:

https://github.com/zhouhaoyi/Informer2020

https://github.com/zhouhaoyi/ETDataset

https://github.com/laiguokun/multivariate-time-series-data

https://github.com/thuml/Autoformer

film's People

Contributors

Stargazers

Watchers

Forkers

lyapunovstability dhockaday arrowooops shism2 vishalbelsare riccardogibello

film's Issues

FileNotFoundError: [WinError 3] 系统找不到指定的路径。: './checkpoints/test_FiLM_fourCroguidedm2TanhR_ab0_modes64_uwl0_ETTh1_ftM_sl2400_ll1200_pl2400_dm512_nh8_el2_dl1_df2048_fc1_ebtimeF_dtTrue_test_0'

你好，我在拜读完您的文章《FiLM: Frequency improved Legendre Memory Model for Long-term Time Series Forecasting
》之后遇到了一个Bug，run.py作为启动界面在调用exp_main.py文件中的train()方法的时候报错，报错代码在path = os.path.join(self.args.checkpoints, setting)这句话。

报错内容：FileNotFoundError: [WinError 3] 系统找不到指定的路径。: './checkpoints/test_FiLM_fourCroguidedm2TanhR_ab0_modes64_uwl0_ETTh1_ftM_sl2400_ll1200_pl2400_dm512_nh8_el2_dl1_df2048_fc1_ebtimeF_dtTrue_test_0'

其中在run.py中进行对--model的参数进行了改动，改成了FiLM这个模块，因为我想复现一下您的FiLM模型

您可以指导我一下是哪里出了问题吗，麻烦您了！

FEL Original version

In your paper, you wrote about Frequency Enhanced Layer Original version (use weights W). Where is the source of it?

FNO调用问题

老师您好，我根据论文描述，理解的是您在FiLM文件中的190行代码self.spec_conv_1 = nn.ModuleList([SpectralConv1d()])调用了FNO函数，但运行程序时，调用了第74行的class SpectralConv1d(nn.Module):。在该类中我个人没有找到调用FNO的办法，请问需要哪一步设置可以调用FNO。

脚本问题

作者您好，我想问一下script里面的脚本如何使用

FiLM consult

Hello, I am a graduate student at China University of Geosciences Wuhan, and I recently read your paper FiLM: Frequency improved Legendre Memory Model for Long-term Time Series Forecasting published in NIPS and open-sourced the code. In the paper, it is written that FLIM is a feature extraction model that can be used to extract historical information and remove historical noise. After reading your paper, I ran the open source FILM.py file, which contains examples for testing, and I encountered a problem that the output of the model is all one value.

ValueError during testing

Hello Zhou,

Thank you for releasing the code.
I am trying to recreate your results

I was able to run and recreate the results for the Film model.

Now, I am trying to run the Informer and Logformer models for comparison
I am able to train successfully

But, during testing, I am getting the following error

ValueError: operands could not be broadcast together with shapes (44,96,1) (1408,96,1)

Please let me know the fix for this, thanks.

Best regards
Niharika

Memory and Training Speed Computation

Hi, thanks for your great work! Would it be possible to share the script to calculate the memory usage and training speed presented in Figure 9? Thanks a lot!