elapsed / pos – the lifetime average since the bar started – is what
the old ETA extrapolated from, and it is a poor estimate of the current
speed: it is dragged down for the whole run by the first step's one-time
setup cost (device/net initialization, first CUDA/MPS kernel compilation,
R JIT warmup), and across training restarts it blends restarts that may
cost very different amounts per epoch. times/positions are a short
rolling window of recent calls (oldest first), so the rate reflects what
the last few steps actually cost, not the whole run's history. Falls back
to the lifetime average when the window is too short to trust (fewer than
two samples, or no time/progress elapsed within it yet).
Arguments
- times, positions
Parallel vectors of recent
Sys.time()timestamps (as numeric seconds) and theposreported at each.- elapsed, pos
Lifetime elapsed seconds and current position, used only as the fallback.