performance-optimization

Optimize training performance for Kaggle competitions. Use PROACTIVELY when training is slow, GPU memory issues occur, or user asks about performance tuning. Keywords: 遅い, 最適化, OOM, CUDA, batch_size, num_workers, 高速化, パフォーマンス

⭐0stars🔀0forks📁View on GitHub🕐Updated Dec 17, 2025⬇️1downloads

When & Why to Use This Skill

This Claude skill is designed to maximize machine learning training efficiency, specifically tailored for Kaggle competitions and high-performance GPU computing. It proactively identifies and resolves performance bottlenecks such as slow training speeds and CUDA Out of Memory (OOM) errors by implementing advanced optimization strategies like Mixed Precision Training and Gradient Accumulation.

Use Cases

  • Resolving 'CUDA Out of Memory' (OOM) errors by applying memory profiling and hardware-specific optimization techniques.
  • Accelerating deep learning model training through the implementation of Mixed Precision (FP16) and optimized Gradient Accumulation patterns.
  • Fine-tuning data loading performance by recommending optimal batch_size and num_workers configurations based on specific GPU hardware specifications.
  • Optimizing resource utilization in Kaggle environments to ensure faster iteration cycles and better competition results.
nameperformance-optimization
descriptionOptimize training performance for Kaggle competitions. Use PROACTIVELY when training is slow, GPU memory issues occur, or user asks about performance tuning. Keywords: 遅い, 最適化, OOM, CUDA, batch_size, num_workers, 高速化, パフォーマンス

When to Use (PROACTIVE)

This skill should be activated automatically when:

  • User reports slow training
  • CUDA Out of Memory (OOM) errors occur
  • User asks about batch_size or num_workers tuning
  • User mentions: "遅い", "最適化", "OOM", "高速化"
  • Training speed is suboptimal

What This Skill Does

Provides GPU-specific optimization strategies:

  • Batch size and num_workers recommendations by GPU
  • Mixed Precision Training implementation
  • Gradient Accumulation patterns
  • Memory optimization techniques
  • Parallel processing patterns

How to Use

Refer to gpu-tuning-guide.md for:

  • GPU-specific hyperparameter recommendations
  • Mixed precision training code
  • Gradient accumulation implementation
  • Memory profiling techniques
  • Common performance bottlenecks and solutions