A training checkpoint is a saved state from which a run can continue as if it had not stopped: the model’s parameters, the optimiser’s state (moment estimates, step counts), the data sampler’s position and random state, the random generators’ states and the step count.
Quantitative Finance · Begrippenlijst