- Download the ENSO data from tianchi, put it in
tcdata - Train and inference
# a single 42.61 model
cd code
sh run.sh- Af first, I try to practice LightGBM in time series, but its performance is not satisfied here.
- I optimized a PyTorch baseline. Then I decide to use TensorFlow, but my TF model only get 19 score with the same configuration.
- Then I turned back to time series model, I use a GRU seq2seq model modified from my time series repository, it helps a lot.
- https://github.com/jerrywn121/TianChi_AIEarth
- https://github.com/icodeworld/AI-competition
- https://github.com/Wangjw6/Tianchi_Prediction
- https://tianchi.aliyun.com/notebook-ai/detail?spm=5176.12586969.1002.21.561d5330ApD1af&postId=196536
- I should win the 3rd place. The current 3rd team, from the host university, use lots of accounts to overfitting. It's funny that they admitted it in comments here.