jerrywn121 / tianchi_aiearth Goto Github PK

TianChi AIEarth Contest Solution

Python 100.00%

convlstm spatial-temporal-forecasting attention-mechanism transformer timesformer pytorch

tianchi_aiearth's Introduction

TianChi AIEarth

This is the top 1 (B榜) solution for the TianChi AIEarth Contest (https://tianchi.aliyun.com/competition/entrance/531871/introduction?lang=zh-cn)

The three models are implemented in Pytorch for Nino prediction and currently support single-GPU training only. Note that the model architecture of the SAConvLSTM and Conv-TT-LSTM are from the original papers, and we novelly modify the TimeSformer as STTransformer for spatio-temporal prediction.

Dependencies

In general, several major packages are needed
xarray==0.17.0
netcdf4==1.5.6
torch==1.7.1
numpy==1.19.1

References

SAConvLSTM
Self-Attention ConvLSTM for Spatiotemporal Prediction (Lin et al., 2020) https://ojs.aaai.org/index.php/AAAI/article/view/6819
Conv-TT-LSTM
Convolutional Tensor-Train LSTM for Spatio-temporal Learning (Su et al., 2020) https://arxiv.org/abs/2002.09131
TimeSformer
Is Space-Time Attention All You Need for Video Understanding? (Bertasius et al., 2021) https://arxiv.org/abs/2102.05095

tianchi_aiearth's People

Contributors

Stargazers

Watchers

tianchi_aiearth's Issues

One question for SA-ConvLSTM

Hi.

This model can do spatio-temporal prediction like MovingMNIST dataset?

Is there a tip for applying MovingMNIST dataset to your code?

thank you.

Doubts about the time sliding window of the sample set

Hello, I'm a graduate student of geography science and I do some research on the prediction of ENSO recently. I want to ask how to set the sliding window of this training sample set? After looking at the code for a long time, I didn't understand this question. Does this competition use the data from January to December to predict the Nino3.4 index from January to December in the next two years? Why is the netCDF data of the Tianchi data set to month=36? If the sliding window does not start in January, how will the data conversion be handled? I noticed that this code set input_length=12 and pred_length=26, but I don't quite understand the purpose of this.

bug

SAConvLSTM/trainer.py#L101
missing a return value

The way CMIP data is processed（cmip_dataset）

Hello, thank you for the code open source. I would like to ask, what is the idea of the class cmip_dataset to handle CMIP schema data? I would like to learn it. thanks！

想请教一下test.py

我在学习时间序列预测的时候，想自己试一试这个比赛，之后参考了您的模型，但是我写的inference总是报错。能不能参考一下您的inference.py或是test.py。十分感谢

关于SAConvLSTM的一些疑问

大佬好，请问用SAConvLSTM训练TaxiBj数据集时用了early stop吗？还是像paper里训练了8w epoch？
1、如果用了early stop大概在100epoch时模型就停止了，MAPE=1.26 RMSE=110。
2、用2060本地跑一轮epoch 35s，如果训练8w epoch需要32天。还是说大概在几百几千epoch就early stop了呢？

我没有改模型架构，只是改了下Dataloader和input_length、output_length这些参数

How to run the test set?

Hello, I am a new learner. Could you please upload the .py file to run the test set? That will be very helpful, thanks a lot.

Question for inference code

Dear Jerry,

I hope this message finds you well. I wanted to reach out to you regarding the inference code that I built based on your training code. After testing it on the contest's B data set, I noticed that my results were significantly different from those displayed on the competition board.

I was wondering if you would be willing to share your inference code with me so that I can compare it with mine and see if there are any discrepancies. Your assistance would be greatly appreciated.

Thank you for your time and consideration.

Best regards

Question about SSTtransformer

Thank you for your released codes. I wonder which predictive model achieve the best performance？

When I run SSTtransformer, I find that there is a big difference between the predictions of model(Train=False) and model(Train=True). May I ask how you solved this problem? Also how did you adjust the ratio for scheduled sampling ?