Set a global random seed for reproducible geodl workflows.
Details
A geodl/torch training pipeline draws from two independent random number
generator (RNG) streams: R's base RNG and torch's own RNG.
setSeed() seeds both from a single value so that dataset
shuffling, data augmentation, model weight initialization, and the
training loop all produce identical results across runs.
The following sources of randomness are covered:
- Dataset shuffling
torch::dataloader(..., shuffle = TRUE)and anydplyr::sample_n()-based shuffling (e.g.makeChipsDF(shuffle = TRUE)) read from R's base RNG.- Data augmentation
defineSegDataSet()anddefineDynamicSegDataSet()choose which augmentations to apply and their magnitudes with base R functions (runif(),sample(),dplyr::sample_n()), which read from R's base RNG.- Model weight initialization
All
defineXXXmodel constructors (e.g.defineUNet,defineDLab) rely on torch's default layer initializers, which draw from torch's RNG.- Training loop
Any stochastic operations inside
luz::fit()(e.g. dropout, if present in a given architecture) draw from torch's RNG.
Call this before creating the dataset, dataloader, and model,
and before calling luz::fit() — any random draws that occur
before seeding will not be reproducible.
When torch::dataloader() is used with num_workers > 0,
torch automatically reseeds each worker subprocess (with both
set.seed() and torch::torch_manual_seed()) from the parent
process's RNG state at the time the workers are spawned. As long as
setSeed() is called once before the dataloader is created, and
num_workers is held constant between runs, worker seeding is
deterministic and requires no further action.
Full determinism on GPU is still subject to caveats outside this function's control: some cuDNN convolution algorithms are inherently non-deterministic regardless of RNG seeding, and torch's R bindings do not currently expose a "deterministic algorithms" toggle. For strict bit-for-bit reproducibility, training on CPU remains the most reliable option.