I took the code from the original repo and simplified a bunch of stuff. The version in that repo should run correctly.
Single model accuracy on the validation split in that code gets to ~0.100. Not sure how those validation splits compare to the ones from the original competition though.
1:29:00 , I feel that by using a higher cardinality of the embedding layer, we are letting the model learn more information around what a day of week means from various perspectives in addition to it being just be one of the seven possible days (one hot encoding information). For instance, it helps the model to capture information of how Sunday is different than other days in regards to sales of some of our products or just total sales…