Split out so train_conditional_de() can wrap the whole loop – all
restarts, one progress bar – in a progress-reporting context.
Usage
train_restarts(
build_net,
log_prob_fn,
theta,
x,
max_epochs,
batch_size,
lr,
validation_fraction,
patience,
n_restarts,
clip_grad_norm,
lr_patience,
lr_factor,
min_lr,
seed,
verbose,
device = "cpu"
)Details
device here is still the raw, unresolved keyword train_conditional_de()
was given ("cpu", "cuda", "mps", "gpu" or "auto"); resolving it to
an actual, available device needs torch loaded (see resolve_device()),
and this is the first point that is guaranteed true – require_torch() is
the line above. Doing it here rather than earlier in
train_conditional_de() keeps check_train_controls() (which needs no
torch at all) running first, so a bad batch_size is still reported before
an unavailable device is.