nishide-dev/claude-code-ml-research Skills

Building professional CLIs with Typer and Rich - type-safe argument parsing, progress bars, model visualization, Hydra integration, RichHandler logging, and multi-process handling for ML workflows

Generate and manage Hydra configuration files for machine learning experiments. Use when creating new configs (model, data, trainer, logger, experiment, sweep), organizing config hierarchies, or setting up hyperparameter sweeps with Optuna.

Create and manage data loading, preprocessing, and augmentation pipelines (DataModule, transforms, data loaders). Use when implementing DataModules, setting up data loaders, or optimizing data pipelines for computer vision, NLP, or graph ML tasks.

Debug common ML training issues (NaN loss, OOM, slow training, convergence problems) and provide solutions. Use when training fails, metrics don't improve, or encountering errors like NaN loss, CUDA OOM, or slow convergence.

Manage ML experiments, track results, and compare performance across different configurations. Use when setting up experiment tracking, creating experiment configs, comparing runs, or analyzing experiment results with W&B, TensorBoard, or MLflow.

Format Python code with ruff formatter and optionally fix auto-fixable linting issues. Use when formatting code, preparing code for commit, or ensuring consistent code style across the project.

Comprehensive guide for Hydra configuration management, hierarchical configs, experiment management, Optuna integration, and Lightning integration patterns

Comprehensive guide for PyTorch Lightning - LightningModule, Trainer, distributed training, PyTorch 2.0 torch.compile integration, Lightning Fabric, and production best practices

Run comprehensive code quality checks with ruff (format, lint) and ty (type checking). Use when checking code quality, fixing linting errors, or ensuring code follows best practices before commits or PRs.

Export trained PyTorch models to various formats (ONNX, TorchScript, TensorRT) and upload to model registries (Hugging Face Hub, MLflow). Use when deploying models, sharing trained weights, or preparing for production inference.

Profile ML training performance to identify bottlenecks (data loading, compute, memory usage) and optimize GPU utilization. Use when training is slow, GPU utilization is low, or experiencing memory issues.

Initialize a new ML research project using the ML Research template with PyTorch Lightning, Hydra, and modern Python tooling. Use when starting a new ML project from scratch.

Complete guide for PyTorch Geometric (PyG) - graph neural networks, message passing, large-scale distributed graph learning, Lightning integration, and heterogeneous graphs

Setup development environment with modern Python tooling (uv/pixi), install dependencies, and configure development tools (ruff, ty, pytest). Use when setting up new ML projects, configuring environments, or installing dependencies.

Execute training runs with proper monitoring, checkpointing, and experiment tracking. Use when starting training, resuming training, debugging training issues, or setting up multi-GPU/distributed training with PyTorch Lightning and Hydra.

Hugging Face Transformers with PyTorch Lightning - LightningModule integration, distributed training (FSDP/DeepSpeed), PEFT (LoRA/QLoRA), data pipelines with HF Datasets, evaluation metrics, and common NLP tasks

Comprehensive validation of ML project structure, configurations, code quality, and training readiness. Use when setting up a new project, before training runs, or debugging configuration issues. Validates config loading, data pipeline, model architecture, and dependencies.

Complete guide for Weights & Biases (W&B) - experiment tracking, hyperparameter sweeps, artifact management, model registry, and PyTorch Lightning integration