Transformers End to End Experience training...
summary
Autodesk describes practical results using Hugging Face models for support routing and e‑commerce tasks, favoring zero-shot classification and weak supervision (Snorkel / Hazy Research) to avoid costly manual labels, spending roughly 4/5 of effort on data cleaning. Training uses SageMaker pipelines and sentence‑transformer embeddings rather than full end‑to‑end fine‑tuning; deployment explores distillation, quantization and AWS Inferentia achieving 4.9x throughput on typical 20–30 GB models.
AI-written summary · about
chapters (10)
- 0:00Introduction and Autodesk overview
- 2:41What are Transformers
- 8:16Business use cases for Transformers
- 12:06Data cleaning and preparation
- 15:33Zero-shot classification
- 20:31Weak supervision techniques
- 22:17Training models with Transformers
- 28:12Testing and model discipline
- 30:18Inference and deployment
- 32:05Conclusion and key takeaways
related talks
01h 09mReinventing Machine Learning with Transformers and Hugging Face
25mTransformers.js: Building Next-Generation WebAI Applications
58mAn Introduction to Transfer Learning and Hugging Face
33mMachine Learning 2.0 with Hugging Face
01h 16mHow Do Transformers Work?
23mUnveiling the Journey of NLP
32mAn Introduction to Transfer Learning in NLP and HuggingFace
54mDeep learning on AWS with NVIDIA: From training to deployment