rmgogogo/nano-transformers

Nano Transformers, Transformer, GPT, GPT-2, GPT-3 (AST) etc.

nano-transformers

Nano-Transformers, a project for Transformer related education and knowledge quick reference purpose.

This project targets at

• help you easily understand Transformers from detailed simple codes
• help you easily write Keras codes, with less complains on TensorFlow
• help you easily get investment from VC if you were working on web3

2017 Transformer

Nano Transformer

2018 GPT-1

Nano GPT-1

2019 GPT-2

Nano GPT-2

2020 GPT-3

Nano GPT-3

2021 InstructGPT

InstructGPT, not fully done yet.

• Tried to get emergent ability in small model small data. Failed. It can't reasoning, but memory well.

References

• greentfrapp/attention-primer
  • It's a TensorFlow 1 implementation. That project inspired nano-transformers.
• sainathadapa/attention-primer-pytorch
  • It's a PyTorch implementation, forked from previous one.
• karpathy/nanoGPT
  • It's a PyTorch implementation of GPT2, a rewrite of karpathy/minGPT.
  • Andrej Karpathy was Tesla AI director in 2017-2022. The man behinds the magic. He is actively developing nanoGPT, we can learn a lot from his detailed optimizations.
• The 3 GPT images are captured from GPT1,2,3.
• Attention Free Transformer
• Rethinking Attention with Performers