transformers python

A GPT-2 implementation using as vanilla python as possible, for future conversion to C++.

steps

• implement the decoder stack (masked self-attention, feed forward) using numpy
• implement transformer things (positional encoding)
• standardize weight-loading schema, test to ensure that it's actually working
• tally used numpy methods and implement a naive matrix class that supports thoes operations

carried by

• Jay Alammar: The Illustrated GPT-2
• GPT-2 paper section 2.3
• GPT paper section 4.1 Model specifications
• Attention Is All You Need section 3.1