Best practices

This section collects habits that keep PhenoMe analyses reproducible, fast enough for the hardware you have, and easy to share. Each topic is a short focused page:

Topic	What you’ll find
Reproducibility	Seeds, deterministic mode, saving configurations, DR backend choice.
Performance	GPU sizing, batch size, memory management, troubleshooting OOM and slow runs.
Data organization	Directory layout, naming conventions, inspection, filtering.
Batch correction	Correcting for technical variation (plate effects) using controls.
Collaboration	Sharing experiments, DVC, environment specs.

Quick reference

Topic	Key tip
Reproducibility	PhenoMe(seed=42); use export_experiment_config() after save_results().
DR backends	use_gpu_for_dr=False for strict reproducibility (sklearn / umap-learn).
DR GPU OOM	Install pykeops for memory-efficient t-SNE / UMAP; PCA uses batched ExactIncrementalPCA.
GPU OOM	Reduce batch_size, pick a smaller model, process in chunks.
Portability	HDF5 stores relative paths - call find_files first, then load_results.
Batch correction	Use pheno.correct_batches(batch_metadata_key="Plate", control_filters={"condition": "Control"}).
Data inspection	Run pheno.inspect_data() before processing (after find_files).