Build NLP Models with Amazon SageMaker
summary
Demo shows end-to-end BERT fine-tuning on Amazon SageMaker using TensorFlow (script mode), with HuggingFace DistilBERT tokeniser and TFRecord features derived from Amazon’s 130+ million review dataset; training runs on two p3.2xlarge GPUs (sharded input) and completes in ~15 minutes for the sample, producing confident predictions (e.g., 5‑star at 92%). Setup uses SageMaker Experiments and SageMaker Debugger for metrics and rules, and highlights cost savings with Managed Spot Training plus checkpointing.
AI-written summary · about
chapters (15)
- 0:00Introduction to session
- 0:48What is NLP
- 4:54NLP history and algorithms
- 7:20BERT and attention mechanism
- 9:32BERT family models
- 16:03NLP with Amazon SageMaker
- 18:30Demo setup and dataset
- 19:53BERT fine-tuning with TensorFlow
- 29:46Training job execution
- 31:57Model evaluation results
- 33:24Experiment tracking and SageMaker features
- 35:12PyTorch on AWS and TorchServe
- 38:33Deepset and BERT optimization
- 42:14Apache MXNet and GluonNLP
- 44:23Resources and closing
related talks
30mBuild consistent and portable ML environments with containers
25mAccelerate NLP training with Amazon SageMaker
45mAutomate your ML workflows with pipelines
31mTrain and tune ML models to the highest accuracy using Amazon SageMaker
39mTrain ML models at scale with Amazon SageMaker Training
33mBuild, train, and deploy ML models using Amazon SageMaker
36mSmokey and the Multi-Armed Bandit featuring BERT Reynolds
30mTrain ML models faster with better insights using Amazon SageMaker Debugger