Cleanup and Model Improvements

[sshleifer]

Changes that improve performance to 3.00 - 3.02 test MAE:

• Learning Rate Decay (default =.97)
• skip connection if cat_feat_gc=True
• default nhid=40 (instead of 32)
• default clip=3 (instead of 5)
• replace 0s with mean in StandardScaler

Changes that don't affect performance:

• only save one checkpoint
• add early stopping
• simplification of do_graph_conv command line args
• share parser code between train.py and test.py
• misc renamings
• store metrics in pandas and write them to disk nicely
• no expid --save specifies a directory.
• save args to --save for reproducibility
• train.py skip batches where all targets are 0: they will certainly have 0 loss.
• misc renamings
• check in best_model.pth. Dont need to retrain to inspect.
• calculate all three metrics in one function instead of 3.

Metrics: baseline: 3.00 - 3.02 finetuning: 2.99-3.00

Paper has more details on experiments that didn't work: http://arxiv.org/abs/1912.07390

[nnzhan]

Dear all

Thanks for your great contribution. I am impressed by your wonderful efforts to improve Graph Wavenet. I will review the codes as soon as possible.

Bests Zonghan

[nnzhan]

Dear all,

I think you have done a very good job. To respect your contribution mostly, I would highly recommend to keep your own repository and not merge into the project. I will instead add a link referring to your repo.

Best regards, Zonghan