Can't reproduce accuracy numbers in "Structured data learning with TabTransformer"

[rama100]

In Structured data learning with TabTransformer, we have the following:

• The baseline linear model achieves ~81% validation accuracy.
• The TabTransformer model achieves ~85% validation accuracy.

I ran the colab as-is with no modifications and the accuracy numbers don't match the above - I get 83.78% and 82.8% for the baseline and the TabTransformer model respectively i.e. the TabTransformer underperforms the baseline :-(.

(If random seeds had been set in the notebook, perhaps this could be avoided)