torchFastText: Modernizing Text Classification at Insee with PyTorch-based models
summary
Open-sourced torchFastText replaces archived fastText for INSEE by providing a PyTorch training stack that trains on GPU but runs CPU-only inference, targets 200 ms inference per sentence, and handles 700+ classes of NACE labels. The pipeline uses MLflow for tracking and model packaging, FastAPI + ArgoCD for serving, and Captum for explainability. The claim is concrete: the PyTorch model beats the previous fastText on INSEE data, is packaged for secure CPU production, and is available for others to inspect and extend.
AI-written summary · about
chapters (14)
- 0:00Modernizing text classification
- 3:22Current MLOps pipeline
- 6:11Data cleaning and rules
- 7:58MLflow for training
- 9:45Human-readable predictions
- 10:49Custom metric for trade-offs
- 11:30FastAPI and Argo CD
- 13:14Model monitoring
- 14:38Why torchFastText
- 17:52torchFastText features
- 20:02Explainability with Captum
- 21:07Future MLOps automation
- 21:50torchTextClassifiers roadmap
- 25:36Q&A
related talks
01h 32mNeural Networks for Natural Language Processing
32mDeep learning and spark
18mPutting the Fast in fast.ai - Machine Learning Models in 15 mins
06mTorchtext 0.4 with Supervised Learning Datasets: A Quick Introduction
08mTorchText
18mDeploying ML solutions with low latency in Python
26mMachine Learning in Fastly's Compute@Edge
17mF8 Refresh | Deep Learning with PyTorch