The seminar "Building Machine Learning Apps with Hugging Face: LLMs to Diffusion Modeling" from a few weeks ago from Hugging Face provides some interesting insights on LLMs to Diffusion Modeling.
History of MLOps
- Software is eating the world
- Deep learning is eating the world
- Transformers are eating the Deep learning
Underthehood, transfer learning is the key player for transformers eating the deep learning.
Transfer learning
There are many pre-trained models which have been trained on large-scale data that we cannot do by ourselves. But thankfully, we can use the output (the pre-trained models) for our tasks.
So nowadays, we can easily use those models, e.g., via Hugging Face, by following steps:
1. Identify the task matching the problem
- Text, Images, Speech
2. Pick a pre-trained SOTA
- no need to build, label, and clean a large dataset
- 2 lines of code to download and test a model
3. If needed, fine-tune it on your dataset
- build customized layers on top of pre-trained models for our needs
- only train the parameters of the customized layers
Mission of Hugging Face: Democratize Good ML
- open source, community, ethics-first
Some facts about Hugging Face
- Diffusers is growing faster than Transformers
- 150k free public models
- Over 50% of Hub models are private
- 1M model downloads/day
- How we make money? Sell compute services, expert support
- Get started with /tasks
- Bloomz from Bloom: an instruction fine-tuned open source LLM
- Fast ControlNet: Guide image generation
- Mac app for that (Diffusers)
- Free diffusion course from Hugging Face (git)
- Javascript client - huggingface.js
Building ML with Hugging Face
- Explore models
- Manage your models (every model is on github)
- Collaborate
- Use
from transformers import pipeline
- pipeline - encapsulize all - text in text out
AutoTrain
- allows to train based on your data (pricing)
Spaces
- allows you to create your app
Deploy models at scale
- headache about GPUs, CPUs, Docker containers etc.
- Inference Endpoints (we made models plug and play)
LLMs 3types
- Bloom, GPT
- Bloomz, GPT-3: intruction guided
- ChatGPT: Human feedback (RL-based)
Pinecone: vector database company
ModelScope from text to generate video