Modern AI Stack

tags : Programming Languages, Machine Learning, Computer Vision

This is to understand what happened in AI post 2020. Also to keep track of things.

NLP §

See NLP (Natural Language Processing)

Transformers §

Transformers FAQ §

zero-shot/one-zero/few-shot learning? §

• Full training: This is not even in the shot spectrum, we train the whole data to get predictions. We can do “shot” kind of things on pre-trained models.
• few-shot: few more example/runs, we give it few examples etc.
• one-shot: give one example, it’ll be able to do it.
• zero-shot
  • We do not want to give any concrete examples at all
  • But instruct the model in a different way. (Eg. Prompting)

What attention? §

• In each layer of the network
  • Encoders taking words from an input sentence, converting them into a representation
  • Each decoder takes all the encoders’ representation of words and transforms them into words in another language.
• Decoders getting all the encoders’ output provides a wider context and enables Attention.

Usage (Modality) §

• Text
  • Text classification
  • Test generation
  • Summarization
• Audio
  • Audio classification
  • Automatic speech recognition
• Vision
  • Object detection
  • Image classification
  • Image segmentation
• Multi Modal
  • Visual QA
  • Document QA
  • Image captioning

Diffusion models §

• These are different from transformers in arch, training process, how they infer, usecase etc.
• See
  • ⭐ Diffusion Models
  • What are Diffusion Models? | Lil’Log
  • Diffusion models are autoencoders – Sander Dieleman

Fineturning diffusion models (Stable Difussion) §

Dreambooth §

Textual Inversion §

LoRA §

Depth-2-Img §

ControlNet §

• It’s a training strategy, a way of doing fine tuning
• It’s different from Dreambooth and LoRA in ways that they don’t freeze the original model.

• The complimentary external model can be distributed independently or can be baked into one model.
• The complimentary model is specific to freezed main model so it’ll only work with that version so we need to care about compatibility

TTS §

Bark §

• https://github.com/suno-ai/bark
• https://github.com/coqui-ai/TTS
• https://github.com/serp-ai/bark-with-voice-clone

tortoise §

• https://github.com/neonbjb/tortoise-tts
• https://git.ecker.tech/mrq/ai-voice-cloning
• https://github.com/facebookresearch/fairseq/tree/main/examples/mms

Piper §

• https://github.com/rhasspy/piper

StyleTTS §

• StyleTTS2 – open-source Eleven-Labs-quality Text To Speech | Hacker News

STT §

Whisper §

• https://github.com/guillaumekln/faster-whisper