Summary
In this chapter, we discovered the new era of transformer models training 100,000,000,000+ parameters on supercomputers. OpenAI's GPT models are taking NLU beyond the reach of most NLP development teams.
We first examined transformer models from a project management perspective to see if transformers can be designed to use only one GPU, for example, and remain accessible to all. We saw that by optimizing a transformer model's architecture (Reformer) and training methods such as PET, we could reduce the model's size, requiring less machine power.
We then explored the design of GPT models, which are all built on the decoder stack of the original Transformer. The masked attention sub-layer continues the philosophy of left-to-right training. However, the sheer power of the calculations and the subsequent self-attention sub-layer makes it extremely efficient.
We then implemented a 345M parameter GPT-2 model with TensorFlow. The goal was to interact...
