Seed torch's global RNG, returning its previous state to restore later

torch::torch_manual_seed() reseeds torch's one global generator, and there is no way to ask it for an independent stream instead. A caller that reseeds it and never restores the previous state leaves every later unseeded torch call in the same session – another fit, another de_sample() call, a c2st() call with no seed of its own – drawing from wherever the seeded call left the generator, rather than from a fresh one (GitHub #275, the torch analogue of #272).

Usage

set_torch_seed(seed)

Arguments

seed

Integer seed to set torch's global RNG to.

Value

The RNG state captured before reseeding, suitable for torch::torch_set_rng_state().

Details

Pair this with on.exit(torch::torch_set_rng_state(old), add = TRUE) right after calling it. That is the same save-then-restore shape with_fixed_seed() uses for R's own RNG, but split into a value-returning helper plus the caller's own on.exit() rather than an expression-wrapping one: both call sites need the seed live for the rest of a long-running function – a network's random initialization and the epochs of shuffled batches that follow it, not just one expression's value – so the restore has to fire when that function returns, not when this helper does.

Requires torch (call require_torch() first; both call sites are already past that check when they reach this).