In order to train this model on Kaggle Kernels with 16GB RAM instead of 32GB RAM machine, I performed some optimizations to work with these limited resources. This is part 2 of the series.