This is a totally unofficial fork of Google Research's TabNet repo at https://github.com/google-research/google-research/tree/master/tabnet.

The original code accompanies a manuscript, TabNet: Attentive Interpretable Tabular Learning, which was written by Sercan O. Arik and Tomas Pfister.

Please refer to the original repository for the Forest Cover experiments described in the paper.

After having installed the dependencies, perhaps into a virtual environment, with

pip install -r requirements.txt

Run a command like

python train_tabnet.py --csv-path data/adult.csv \
                       --target-name "<=50K" \
                       --categorical-features workclass,education,marital.status,occupation,relationship,race,sex,native.country
                       --task classification 

There are also many other parameters that can be specified, e.g.

python train_tabnet.py --csv-path data/adult.csv \
                       --target-name "<=50K" \
                       --categorical-features workclass,education,marital.status,occupation,relationship,race,sex,native.country \
                       --task classification
                       --val-frac 0.2 \
                       --test-frac 0.1 \
                       --model-name tabnet_adult_census \
                       --tb-log-location adult_census_logs \
                       --emb-size 1 \ 
                       --feature_dim 16 \
                       --output_dim 16 \
                       --n_steps 5 \
                       --lambda-sparsity 0.0001 \
                       --gamma 1.5 \
                       --lr 0.02 \
                       --batch-momentum 0.98 \
                       --batch-size 4096 \
                       --virtual-batch-size 128 \
                       --decay-every 500 \
                       --max-steps 3000

Because there are so many possible parameters, some "ready-made" configurations are available as shell scripts

Run a command like

python opt_tabnet.py --csv-path data/poker_train.csv --target-name CLASS --categorical-features S1,S2,S3,S4,S5

Note that you should probably change the parameters max_steps and early_stop_steps manually depending on the dataset.

I wanted to be able to run TabNet in a more flexible way, where I would not have to worry about manually changing the code as in the original implementation. I also wanted to be able to train TabNet models directly on CSV files. This means that you should only run this code if your dataset fits in RAM. Otherwise, it is better to use the author's original CSV data loader (which is still retained in this version).

The code here has functionality to train TabNet models from the command line, and includes some toy datasets for reference: Adult Census, Poker Hands, and AutoMPG (the latter one is for regression rather than classification.)

Caveats:

• I am not sure I have implemented the regression loss correctly, as there was no example of that in the original code.
• The default size of the embedding layers is one (!). This was the case in the original code and one of the authors explained to me that it is done for the sake of interpretability.