Creating embeddings – tokenizers and other key steps for smart features

Now that you have your data loader tested, built, and possibly scaled, you’re thinking to yourself, what do I do with all of these raw images and/or natural language strings? Do I throw them straight into my neural network? Actually, the last five years of learning representations have proven this definitively: no, you should not put raw images or text into your neural network right off the bat. You should convert your raw inputs to embeddings by using another model.

The intuition for this is simple: before you teach your model how to recognize relationships in your dataset, you first have to introduce it to the concept of a dataset. Creating embeddings is basically a way of doing this; you use a data structure that has been trained from another process to create vector representations of your data. That is to say, you provide your raw text and images as input, and you get high-dimensional vectors...