util.seed

seed_everything

def seed_everything(seed: int) -> None

Seed every RNG VAME draws from, so a fixed project_random_state gives stable scientific output across runs on the same machine/device.

Seeds Python random, NumPy's global RNG and Torch (CPU + CUDA). Call once at the start of each pipeline op; downstream global-RNG draws (dataset crops, KMeans/HMM/UMAP, generative and visualization sampling) then become deterministic. This reseeds the global RNGs as a side effect. It does not enable Torch deterministic algorithms (kept off to avoid the training slowdown), so cross-device bit-exactness is not guaranteed.