def main():
args = parser.parse_args()
args, logging, writer = utils.parse_args(args)
logging.info('# Start Re-training #')
criterion = LOSS_FACTORY[args.task](args, args.loss_scaling)
if args.model_type == "stochastic":
model_temp = STOCHASTIC_FACTORY[args.model]
else:
raise NotImplementedError("Other models have not been implemented!")
model = model_temp(args.input_size, args.output_size, args.layers,
args.activation, args, True)
logging.info('## Model created: ##')
logging.info(model.__repr__())
logging.info("### Param size = %f MB, Total number of params = %d ###" %
utils.count_parameters_in_MB(model, args))
logging.info('### Loading model to parallel GPUs ###')
utils.profile(model, args, logging)
model = utils.model_to_gpus(model, args)
logging.info('### Preparing schedulers and optimizers ###')
optimizer = SGLD(model.parameters(),
args.learning_rate,
norm_sigma=args.weight_decay)
scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(
optimizer, args.epochs)
logging.info('## Downloading and preparing data ##')
train_loader, valid_loader = get_train_loaders(args)
logging.info('## Beginning Training ##')
train = Trainer(model, criterion, optimizer, scheduler, args)
best_error, train_time, val_time = train.train_loop(
train_loader, valid_loader, logging, writer)
logging.info(
'## Finished training, the best observed validation error: {}, total training time: {}, total validation time: {} ##'
.format(best_error, timedelta(seconds=train_time),
timedelta(seconds=val_time)))
logging.info('## Beginning Plotting ##')
del model
with torch.no_grad():
args.samples = 100
args.model_path = args.save
model = model_temp(args.input_size, args.output_size, args.layers,
args.activation, args, False)
model = utils.model_to_gpus(model, args)
model.eval()
plot_regression_uncertainty(model, PLT, train_loader, args)
logging.info('# Finished #')
def main(wrapper):
args = wrapper.args
model = wrapper.model
save_dir = args.save
try:
wrapper.load()
logging.info('loaded previously saved weights')
except Exception as e:
print(e)
logging.info("param size = %fMB", utils.count_parameters_in_MB(model))
logging.info('Args: {}'.format(args))
if args.eval_only:
assert save_dir is not None
data_size = 25000
time_steps = 1
B = int(args.epochs * data_size / args.batch_size / time_steps)
searcher = Random_NAS(B, wrapper, args.seed, save_dir)
logging.info('budget: %d' % (searcher.B))
if not args.eval_only:
searcher.run()
archs = searcher.get_eval_arch(args.randomnas_rounds, args.n_samples)
else:
np.random.seed(args.seed + 1)
archs = searcher.get_eval_arch(2)
logging.info(archs)
#arch = ' '.join([str(a) for a in archs[0][0]])
arch = str(archs[0][0])
arch = parse_arch_to_darts('cnn', ast.literal_eval(arch), args.space)
with open(os.path.join(args.save, 'arch_{}'.format(args.task_id)),
'w') as f:
f.write(str(arch))
logging.info(str(arch))
utils.write_yaml_results(args, args.results_file_arch, str(arch))
return arch