model-optimization_skill

Overview

This skill provides practical tools and recipes for improving model performance, reducing model size, and accelerating inference. It covers hyperparameter tuning, AutoML, quantization, pruning, knowledge distillation, feature selection, and inference deployment optimizations. Use it to balance accuracy, latency, and resource cost across training and production.

How this skill works

The skill exposes common optimization patterns and code examples for Python-based ML workflows: grid and Bayesian search, Optuna, and AutoML frameworks for automated model selection. It includes model compression techniques (post-training quantization, quantization-aware training, pruning, distillation) plus conversion and runtime tips (ONNX, TensorRT) for faster inference. It also covers feature selection, learning rate scheduling, and early stopping to streamline training and avoid overfitting.

When to use it

• When models are too large for deployment on edge or limited hardware
• When inference latency must be reduced for real-time applications
• To systematically search and tune hyperparameters for better performance
• When you need automated pipeline discovery or a quick baseline with AutoML
• To reduce training time or prevent overfitting via scheduling and early stopping

Best practices

• Start with a simple baseline and profile to identify bottlenecks before optimizing
• Measure improvements using the same validation and production-like data; use A/B tests for changes in production
• Balance trade-offs explicitly: accuracy vs latency vs model size
• Prefer automated search (Optuna/Bayes) for complex spaces and manual grid search for small discrete sets
• Apply compression progressively: prune/quantize after validating accuracy, consider distillation for large teacher->small student transitions
• Export to standard runtimes (ONNX) and benchmark on target hardware before finalizing deployment

Example use cases

• Quantize a PyTorch image model to reduce memory 4x and speed up CPU inference for a mobile app
• Use Optuna or BayesSearchCV to tune XGBoost hyperparameters for higher F1 on imbalanced data
• Run Auto-sklearn or H2O AutoML to discover strong baselines and candidate pipelines quickly
• Prune 20% of network weights and fine-tune to cut model size with minimal accuracy loss
• Convert a trained model to ONNX and optimize with TensorRT for 10x faster GPU inference