Fastai Tabular Shared Embedding

Part 1 (2019)
1

I’m basically using 2 inputs with multiple (categorical) features that needs to share the embeddings between them. So, input_1<cat_1> and input_2<cat_1> should pass through the same embedding layer.

This is how I do it in Keras -

image
image.png986×337 18.4 KB

But, I’m now trying to port this to fast.ai to use the excellent fit_one_cycle among other things!

But, I’m not entirely sure if just building a TabularModel class as per my specification is enough, or do I need to modify the DataBunch definition as well?
Anybody worked on something similar?

2

Hi, have you had a look into: https://docs.fast.ai/tutorial.itemlist.html ?

3

Nevermind! The data can be as it is.

We can just build a new Tabular model to accommodate the changes