def do_validate(conf, model, optimizer, criterion, scheduler, metrics,
data_loader):
"""Evaluate the model on the test dataset and save to the checkpoint."""
# wait until the whole group enters this function, and then evaluate.
print("Enter validation phase.")
performance = validate(conf, model, optimizer, criterion, scheduler,
metrics, data_loader)
# remember best performance and display the val info.
scheduler.best_tracker.update(performance[0], scheduler.epoch_)
dispaly_best_test_stat(conf, scheduler)
# save to the checkpoint.
if not conf.train_fast:
save_to_checkpoint(
conf,
{
"arch": conf.arch,
"current_epoch": scheduler.epoch,
"local_index": scheduler.local_index,
"best_perf": scheduler.best_tracker.best_perf,
"optimizer": optimizer.state_dict(),
"state_dict": model.state_dict(),
},
scheduler.best_tracker.is_best,
dirname=conf.checkpoint_dir,
filename="checkpoint.pth.tar",
save_all=conf.save_all_models,
)
print("Finished validation.")
def do_validation(
conf,
coordinator,
model,
criterion,
metrics,
data_loaders,
performance=None,
label=None,
):
"""Evaluate the model on the test dataset and save to the checkpoint."""
# wait until the whole group enters this function, and then evaluate.
conf.logger.log(f"Master enters the validation phase.")
if performance is None:
performance = get_avg_perf_on_dataloaders(
conf, coordinator, model, criterion, metrics, data_loaders, label
)
# remember best performance and display the val info.
coordinator.update_perf(performance)
dispaly_best_test_stat(conf, coordinator)
# save to the checkpoint.
conf.logger.log(f"Master finished the validation.")
if not conf.train_fast:
checkpoint.save_to_checkpoint(
conf,
{
"arch": conf.arch,
"current_comm_round": conf.graph.comm_round,
"best_perf": coordinator.best_trackers["top1"].best_perf,
"state_dict": model.state_dict(),
},
coordinator.best_trackers["top1"].is_best,
dirname=conf.checkpoint_root,
filename="checkpoint.pth.tar",
save_all=conf.save_all_models,
)
conf.logger.log(f"Master saved to checkpoint.")
