How to Implement a Production Ready Deep Learning Model in 12 Weeks
summary
Claims a production-ready image aesthetic classifier in 12 weeks by combining Lean Startup cycles with ML engineering. Uses a pretrained CNN (initially VGG16 then a lighter model), follows the NIMA paper with earth mover distance, deploys on cloud GPU instances and OpenShift, and ships a one-day labeling app to collect domain labels. Handles ~300 million images by sampling a million, retraining on in-house labels, scaling to dozens of replicas, and delivers usable ordering and high-score detection after fine-tuning.
AI-written summary · about
chapters (27)
- 0:00Introduction
- 4:10Why this talk
- 8:53Why is this important
- 12:58Example
- 13:25Most important criteria
- 14:05The issue
- 14:36Diallo example
- 15:24Gallery example
- 15:54Perspective
- 16:29Reception
- 17:04Aesthetics
- 18:12Hotel
- 18:40Ideal Case
- 18:53The Problem
- 20:36What You Should Not Do
- 21:07Machine Learning Team
- 21:50Machine Learning
- 24:05Pretrained Neural Networks
- 24:31Deployment
- 25:24First and Second iteration
- 26:31Results of First iteration
- 30:50What we learned
- 34:20Deployed
- 36:24Results
- 36:49What would work
- 37:04Conclusion
- 37:39Summary
related talks
33mRunning Machine Learning in Production - a Journey to Success
23mBest Practices for Building Production-Grade Deep Learning Systems
01h 01mA practical guide to deep learning
35mGetting Machine Learning Models Ready For Production Using Python
18mPutting the Fast in fast.ai - Machine Learning Models in 15 mins
18mDeep learning for Space imagery applications
35mLevel Up Your Machine Learning Lifecycle
44mDeep Learning in Production