Finding BAD IMAGE DATA using UMAP and Prodigy
Prodigy is a modern annotation tool for collecting training data for machine learning models, developed by the makers of spaCy. In this video, we'll show you how to use Prodigy to find bad examples in the Google QuickDraw dataset. We will be leveraging a technique that involves UMAP to find strange images semi-automatically.
[00:00] Introduction
[04:04] Using Quick!Draw!
[07:27] Exploration in Jupyter
[11:05] UMAP Clusters
[14:48] UMAP in Jupyter
[18:27] Introducing Prodigy
[19:10] Project Setup
[23:30] Labelling with Prodigy
[27:25] Prodigy Output
[30:54] Manually Hashing in Prodigy
[33:13] More Labelling
[35:18] Lessons Learned
PRODIGY
● Website & docs: https://prodi.gy
● Live demo: https://prodi.gy/demo
● Forum: https://support.prodi.gy
THIS TUTORIAL
● Code & data: https://github.com/explosion/prodigy-...
● Jupyter Notebook: https://github.com/explosion/prodigy-...
● UMAP Docs: https://umap-learn.readthedocs.io/en/...
● Google Quick! Draw!: https://quickdraw.withgoogle.com/
● Custom Prodigy Recipe Docs: https://prodi.gy/docs/custom-recipes
FOLLOW US
● Explosion: / explosion_ai
● Subscribe to our newsletter: http://eepurl.com/hZyFbf
● We offer new services! spaCy Custom Solutions✨ https://explosion.ai/custom-solutions