Peicewise linear encoding location?

[pqrz]

Great work!
I am unable to locate piecewise linear encoding in this code - rtdl.
If possible, could you point me to it?

[Yura52]

Great work!

Thanks!

I am unable to locate piecewise linear encoding in this code

Thanks for the question! We need to add more instructions to README in that regard.

• paper implementation: https://github.com/Yura52/tabular-dl-num-embeddings
• all neural networks are implemented in bin/train4.py. For PLE-based models, you should use the bins option
• then, check the configuration files in the exp folder. The inner folder names follow the names from the paper. For example, here is the directory for tuning and evaluation of MLP-T-LR on the California Housing dataset: link.
• note that in this specific case, the older version of train4.py, namely train1.py, was used, which can be found by looking at the tuning config
• so, to run the already tuned MLP-T-LR on the California Housing dataset, you can copy any config from the evaluation directory and run train1.py with it
  1. cp exp/mlp-t-lr/california/0_evaluation/0.toml exp/mlp-t-lr/california/0_evaluation/0_reproduce.toml
  2. python bin/train1.py exp/mlp-t-lr/california/0_evaluation/0_reproduce.toml
• however, for futher usage, we recommend adapting the configs for bin/train4.py

Does this help?

[Yura52]

P.S. as of now, in the rtdl library, we do not provide the embedding modules from the paper

[Yura52]

Hi @pqrz ! Just in case, sharing that I created an easy-to-use package for this paper "On Embeddings for Numerical Features in Tabular Deep Learning":

• The package
• The announcement with more details

P.S. More projects on tabular deep learning are available here.