const_lr=args.no_schedule,
double_bias_lr=args.double_bias_lr,
model_variance=args.model_variance,
input_dim=dataset.D,
output_dim=output_dim,
apply_var=args.noise_var,
**model_cfg.kwargs,
**extra_args)
mname = args.model
if args.swag:
mname = mname + args.subspace + args.inference
bb_args = argparse.Namespace(model=mname,
dataset=args.dataset,
split=args.split,
seed=args.seed,
database_path=args.database_path)
bb_result = run(bb_args,
data=dataset,
model=regression_model,
is_test=args.database_path == '')
print([(k, bb_result[k]) for k in sorted(bb_result)])
utils.save_checkpoint(args.dir,
args.epochs,
model_state_dict=regression_model.model.state_dict(),
optimizer=regression_model.optimizer.state_dict(),
result=bb_result)
tablefmt='simple',
floatfmt='8.4f'))
else:
print(
tabulate.tabulate([values],
columns,
tablefmt='plain',
floatfmt='8.4f').split('\n')[1])
print("sigma:",
torch.nn.functional.softplus(vi_model.inv_softplus_sigma.detach().cpu()))
if not args.no_mu:
print("mu:", vi_model.mu.detach().cpu().data)
utils.save_checkpoint(args.dir,
epoch,
name='vi',
state_dict=vi_model.state_dict())
eval_model = model_cfg.base(num_classes=num_classes,
*model_cfg.args,
**model_cfg.kwargs)
eval_model.to(args.device)
num_samples = args.num_samples
query_length = len(meta_testloader.dataset) * len(
meta_testloader.dataset[0][-1])
ens_predictions = np.zeros((query_length, args.n_ways))
targets = np.zeros((query_length, args.n_ways))
#printf, logfile = utils.get_logging_print(os.path.join(args.dir, args.log_fname + '-%s.txt'))
model.load_state_dict(checkpoint['state_dict'])
optimizer.load_state_dict(checkpoint['optimizer'])
if args.swag and args.swag_resume is not None:
checkpoint = torch.load(args.swag_resume)
swag_model.subspace.rank = torch.tensor(0)
swag_model.load_state_dict(checkpoint['state_dict'])
columns = ['ep', 'lr', 'tr_loss', 'tr_acc', 'te_loss', 'te_acc', 'time', 'mem_usage']
if args.swag:
columns = columns[:-2] + ['swa_te_loss', 'swa_te_acc'] + columns[-2:]
swag_res = {'loss': None, 'accuracy': None}
utils.save_checkpoint(
args.dir,
start_epoch,
state_dict=model.state_dict(),
optimizer=optimizer.state_dict()
)
sgd_ens_preds = None
sgd_targets = None
n_ensembled = 0.
for epoch in range(start_epoch, args.epochs):
time_ep = time.time()
if not args.no_schedule:
lr = schedule(epoch, args)
utils.adjust_learning_rate(optimizer, lr)
else:
lr = args.lr_init
