Train ML models faster with better insights using Amazon SageMaker Debugger
summary
Amazon SageMaker Debugger captures full tensors and system metrics during training, runs built‑in rules to flag overfitting, vanishing gradients, CPU/IO bottlenecks, and can auto‑terminate stalled jobs. Demo fine‑tunes a ResNet on traffic signs, fixes class imbalance with weighted resampling, enables pin_memory and more data loader subprocesses, raises batch size 64→264, and cuts training time from ~2200s to ~800s while raising test accuracy 92%→94%. A Mobileye run shows profiler rules firing 16 times when tensor capture every 50 steps causes GPU utilization drops.
AI-written summary · about
chapters (9)
- 0:00Introduction
- 0:46Understanding debugger
- 4:12Live demo: Bug detection
- 10:15Analyzing tensor data
- 13:23Performance profiling
- 17:17Optimizing and results
- 19:56Mobileye use case
- 22:53Solving development challenges
- 26:13Profiling and conclusion
related talks
28mGain better insights while training ML models with SageMaker Debugger
37mStreamlining Machine Learning Operations
27mWaking the Data Scientist at 2am: Detect...
39mTrain ML models at scale with Amazon SageMaker Training
37mStreamlining Machine Learning Operations
33mBuild, train, and deploy ML models using Amazon SageMaker
19mDeveloping large-scale machine learning algorithms on Amazon SageMaker
25mAccelerate NLP training with Amazon SageMaker