def main():
torch.set_num_threads(3)
if not torch.cuda.is_available():
logging.info('no gpu device available')
sys.exit(1)
np.random.seed(args.seed)
gpu = ig_utils.pick_gpu_lowest_memory() if args.gpu == 'auto' else int(
args.gpu)
torch.cuda.set_device(gpu)
cudnn.benchmark = True
torch.manual_seed(args.seed)
cudnn.enabled = True
torch.cuda.manual_seed(args.seed)
logging.info("args = %s", args)
logging.info('gpu device = %d' % gpu)
if not args.fast:
api = API('../data/NAS-Bench-201-v1_0-e61699.pth')
#### model
criterion = nn.CrossEntropyLoss()
search_space = SearchSpaceNames[args.search_space]
if args.method in ['darts', 'blank']:
model = TinyNetworkDarts(C=args.init_channels,
N=5,
max_nodes=4,
num_classes=n_classes,
criterion=criterion,
search_space=search_space,
args=args)
elif args.method in ['darts-proj', 'blank-proj']:
model = TinyNetworkDartsProj(C=args.init_channels,
N=5,
max_nodes=4,
num_classes=n_classes,
criterion=criterion,
search_space=search_space,
args=args)
model = model.cuda()
logging.info("param size = %fMB", ig_utils.count_parameters_in_MB(model))
architect = Architect(model, args)
#### data
if args.dataset == 'cifar10':
train_transform, valid_transform = ig_utils._data_transforms_cifar10(
args)
train_data = dset.CIFAR10(root=args.data,
train=True,
download=True,
transform=train_transform)
valid_data = dset.CIFAR10(root=args.data,
train=False,
download=True,
transform=valid_transform)
elif args.dataset == 'cifar100':
train_transform, valid_transform = ig_utils._data_transforms_cifar100(
args)
train_data = dset.CIFAR100(root=args.data,
train=True,
download=True,
transform=train_transform)
valid_data = dset.CIFAR100(root=args.data,
train=False,
download=True,
transform=valid_transform)
elif args.dataset == 'imagenet16-120':
import torchvision.transforms as transforms
from nasbench201.DownsampledImageNet import ImageNet16
mean = [x / 255 for x in [122.68, 116.66, 104.01]]
std = [x / 255 for x in [63.22, 61.26, 65.09]]
lists = [
transforms.RandomHorizontalFlip(),
transforms.RandomCrop(16, padding=2),
transforms.ToTensor(),
transforms.Normalize(mean, std)
]
train_transform = transforms.Compose(lists)
train_data = ImageNet16(root=os.path.join(args.data, 'imagenet16'),
train=True,
transform=train_transform,
use_num_of_class_only=120)
valid_data = ImageNet16(root=os.path.join(args.data, 'imagenet16'),
train=False,
transform=train_transform,
use_num_of_class_only=120)
assert len(train_data) == 151700
num_train = len(train_data)
indices = list(range(num_train))
split = int(np.floor(args.train_portion * num_train))
train_queue = torch.utils.data.DataLoader(
train_data,
batch_size=args.batch_size,
sampler=torch.utils.data.sampler.SubsetRandomSampler(indices[:split]),
pin_memory=True)
valid_queue = torch.utils.data.DataLoader(
train_data,
batch_size=args.batch_size,
sampler=torch.utils.data.sampler.SubsetRandomSampler(
indices[split:num_train]),
pin_memory=True)
#### scheduler
scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(
model.optimizer, float(args.epochs), eta_min=args.learning_rate_min)
#### resume
start_epoch = 0
if args.resume_epoch != 0:
logging.info('loading checkpoint from {}'.format(expid))
file = 'checkpoint.pth.tar' if args.resume_epoch == -1 else 'checkpoint_{}.pth.tar'.format(
args.resume_epoch)
filename = os.path.join(args.save, file)
if os.path.isfile(filename):
logging.info("=> loading checkpoint '{}'".format(filename))
checkpoint = torch.load(filename, map_location='cpu')
start_epoch = checkpoint['epoch'] # epoch
model_state_dict = checkpoint['state_dict']
if '_arch_parameters' in model_state_dict:
del model_state_dict['_arch_parameters']
model.load_state_dict(model_state_dict) # model
saved_arch_parameters = checkpoint['alpha'] # arch
model.set_arch_parameters(saved_arch_parameters)
scheduler.load_state_dict(checkpoint['scheduler'])
model.optimizer.load_state_dict(
checkpoint['optimizer']) # optimizer
architect.optimizer.load_state_dict(checkpoint['arch_optimizer'])
logging.info("=> loaded checkpoint '{}' (epoch {})".format(
filename, start_epoch - 1))
else:
print("=> no checkpoint found at '{}'".format(filename))
#### training
for epoch in range(start_epoch, args.epochs):
lr = scheduler.get_lr()[0]
## data aug
if args.cutout:
train_transform.transforms[
-1].cutout_prob = args.cutout_prob * epoch / (args.epochs - 1)
logging.info('epoch %d lr %e cutout_prob %e', epoch, lr,
train_transform.transforms[-1].cutout_prob)
else:
logging.info('epoch %d lr %e', epoch, lr)
## pre logging
genotype = model.genotype()
logging.info('genotype = %s', genotype)
model.printing(logging)
## training
train_acc, train_obj = train(train_queue, valid_queue, model,
architect, model.optimizer, lr, epoch)
logging.info('train_acc %f', train_acc)
logging.info('train_loss %f', train_obj)
## eval
valid_acc, valid_obj = infer(valid_queue, model, criterion, log=False)
logging.info('valid_acc %f', valid_acc)
logging.info('valid_loss %f', valid_obj)
## logging
if not args.fast:
# nasbench201
cifar10_train, cifar10_test, cifar100_train, cifar100_valid, \
cifar100_test, imagenet16_train, imagenet16_valid, imagenet16_test = query(api, model.genotype(), logging)
# tensorboard
writer.add_scalars('accuracy', {
'train': train_acc,
'valid': valid_acc
}, epoch)
writer.add_scalars('loss', {
'train': train_obj,
'valid': valid_obj
}, epoch)
writer.add_scalars('nasbench201/cifar10', {
'train': cifar10_train,
'test': cifar10_test
}, epoch)
writer.add_scalars(
'nasbench201/cifar100', {
'train': cifar100_train,
'valid': cifar100_valid,
'test': cifar100_test
}, epoch)
writer.add_scalars(
'nasbench201/imagenet16', {
'train': imagenet16_train,
'valid': imagenet16_valid,
'test': imagenet16_test
}, epoch)
#### scheduling
scheduler.step()
#### saving
save_state = {
'epoch': epoch + 1,
'state_dict': model.state_dict(),
'alpha': model.arch_parameters(),
'optimizer': model.optimizer.state_dict(),
'arch_optimizer': architect.optimizer.state_dict(),
'scheduler': scheduler.state_dict()
}
if save_state['epoch'] % args.ckpt_interval == 0:
ig_utils.save_checkpoint(save_state,
False,
args.save,
per_epoch=True)
#### architecture selection / projection
if args.dev == 'proj':
pt_project(train_queue, valid_queue, model, architect, criterion,
model.optimizer, start_epoch, args, infer, query)
writer.close()
'avoid loading NASBench2 api an instead load a pickle file with tuple (index, arch_str)'
)
args = parser.parse_args()
args.device = torch.device(
"cuda:" + str(args.gpu) if torch.cuda.is_available() else "cpu")
return args
if __name__ == '__main__':
args = parse_arguments()
if args.noacc:
api = pickle.load(open(args.api_loc, 'rb'))
else:
from nas_201_api import NASBench201API as API
api = API(args.api_loc)
torch.manual_seed(args.seed)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
train_loader, val_loader = get_cifar_dataloaders(args.batch_size,
args.batch_size,
args.dataset,
args.num_data_workers)
cached_res = []
pre = 'cf' if 'cifar' in args.dataset else 'im'
pfn = f'nb2_{pre}{get_num_classes(args)}_seed{args.seed}_dl{args.dataload}_dlinfo{args.dataload_info}_initw{args.init_w_type}_initb{args.init_b_type}.p'
op = os.path.join(args.outdir, pfn)
type=str,
help="Folder to save checkpoints and log.")
parser.add_argument(
"--arch_nas_dataset",
type=str,
help="The path to load the architecture dataset (tiny-nas-benchmark).",
)
parser.add_argument("--print_freq",
type=int,
help="print frequency (default: 200)")
parser.add_argument("--rand_seed", type=int, help="manual seed")
args = parser.parse_args()
# if args.rand_seed is None or args.rand_seed < 0: args.rand_seed = random.randint(1, 100000)
if args.arch_nas_dataset is None or not os.path.isfile(
args.arch_nas_dataset):
nas_bench = None
else:
print("{:} build NAS-Benchmark-API from {:}".format(
time_string(), args.arch_nas_dataset))
nas_bench = API(args.arch_nas_dataset)
if args.rand_seed < 0:
save_dir, all_indexes, num = None, [], 500
for i in range(num):
print("{:} : {:03d}/{:03d}".format(time_string(), i, num))
args.rand_seed = random.randint(1, 100000)
save_dir, index = main(args, nas_bench)
all_indexes.append(index)
torch.save(all_indexes, save_dir / "results.pth")
else:
main(args, nas_bench)
def main(xargs):
assert torch.cuda.is_available(), 'CUDA is not available.'
torch.backends.cudnn.enabled = True
torch.backends.cudnn.benchmark = False
torch.backends.cudnn.deterministic = True
torch.set_num_threads( xargs.workers )
prepare_seed(xargs.rand_seed)
logger = prepare_logger(args)
train_data, valid_data, xshape, class_num = get_datasets(xargs.dataset, xargs.data_path, -1)
#config_path = 'configs/nas-benchmark/algos/DARTS.config'
config = load_config(xargs.config_path, {'class_num': class_num, 'xshape': xshape}, logger)
search_loader, _, valid_loader = get_nas_search_loaders(train_data, valid_data, xargs.dataset, 'configs/nas-benchmark/', config.batch_size, xargs.workers)
logger.log('||||||| {:10s} ||||||| Search-Loader-Num={:}, Valid-Loader-Num={:}, batch size={:}'.format(xargs.dataset, len(search_loader), len(valid_loader), config.batch_size))
logger.log('||||||| {:10s} ||||||| Config={:}'.format(xargs.dataset, config))
search_space = get_search_spaces('cell', xargs.search_space_name)
if xargs.model_config is None:
model_config = dict2config({'name': 'DARTS-V1', 'C': xargs.channel, 'N': xargs.num_cells,
'max_nodes': xargs.max_nodes, 'num_classes': class_num,
'space' : search_space,
'affine' : False, 'track_running_stats': bool(xargs.track_running_stats)}, None)
else:
model_config = load_config(xargs.model_config, {'num_classes': class_num, 'space' : search_space,
'affine' : False, 'track_running_stats': bool(xargs.track_running_stats)}, None)
search_model = get_cell_based_tiny_net(model_config)
logger.log('search-model :\n{:}'.format(search_model))
w_optimizer, w_scheduler, criterion = get_optim_scheduler(search_model.get_weights(), config)
a_optimizer = torch.optim.Adam(search_model.get_alphas(), lr=xargs.arch_learning_rate, betas=(0.5, 0.999), weight_decay=xargs.arch_weight_decay)
logger.log('w-optimizer : {:}'.format(w_optimizer))
logger.log('a-optimizer : {:}'.format(a_optimizer))
logger.log('w-scheduler : {:}'.format(w_scheduler))
logger.log('criterion : {:}'.format(criterion))
flop, param = get_model_infos(search_model, xshape)
#logger.log('{:}'.format(search_model))
logger.log('FLOP = {:.2f} M, Params = {:.2f} MB'.format(flop, param))
if xargs.arch_nas_dataset is None:
api = None
else:
api = API(xargs.arch_nas_dataset)
logger.log('{:} create API = {:} done'.format(time_string(), api))
last_info, model_base_path, model_best_path = logger.path('info'), logger.path('model'), logger.path('best')
network, criterion = torch.nn.DataParallel(search_model).cuda(), criterion.cuda()
if last_info.exists(): # automatically resume from previous checkpoint
logger.log("=> loading checkpoint of the last-info '{:}' start".format(last_info))
last_info = torch.load(last_info)
start_epoch = last_info['epoch']
checkpoint = torch.load(last_info['last_checkpoint'])
genotypes = checkpoint['genotypes']
valid_accuracies = checkpoint['valid_accuracies']
search_model.load_state_dict( checkpoint['search_model'] )
w_scheduler.load_state_dict ( checkpoint['w_scheduler'] )
w_optimizer.load_state_dict ( checkpoint['w_optimizer'] )
a_optimizer.load_state_dict ( checkpoint['a_optimizer'] )
logger.log("=> loading checkpoint of the last-info '{:}' start with {:}-th epoch.".format(last_info, start_epoch))
else:
logger.log("=> do not find the last-info file : {:}".format(last_info))
start_epoch, valid_accuracies, genotypes = 0, {'best': -1}, {-1: search_model.genotype()}
# start training
start_time, search_time, epoch_time, total_epoch = time.time(), AverageMeter(), AverageMeter(), config.epochs + config.warmup
for epoch in range(start_epoch, total_epoch):
w_scheduler.update(epoch, 0.0)
need_time = 'Time Left: {:}'.format( convert_secs2time(epoch_time.val * (total_epoch-epoch), True) )
epoch_str = '{:03d}-{:03d}'.format(epoch, total_epoch)
logger.log('\n[Search the {:}-th epoch] {:}, LR={:}'.format(epoch_str, need_time, min(w_scheduler.get_lr())))
search_w_loss, search_w_top1, search_w_top5 = search_func(search_loader, network, criterion, w_scheduler, w_optimizer, a_optimizer, epoch_str, xargs.print_freq, logger, xargs.gradient_clip)
search_time.update(time.time() - start_time)
logger.log('[{:}] searching : loss={:.2f}, accuracy@1={:.2f}%, accuracy@5={:.2f}%, time-cost={:.1f} s'.format(epoch_str, search_w_loss, search_w_top1, search_w_top5, search_time.sum))
valid_a_loss , valid_a_top1 , valid_a_top5 = valid_func(valid_loader, network, criterion)
logger.log('[{:}] evaluate : loss={:.2f}, accuracy@1={:.2f}%, accuracy@5={:.2f}%'.format(epoch_str, valid_a_loss, valid_a_top1, valid_a_top5))
# check the best accuracy
valid_accuracies[epoch] = valid_a_top1
if valid_a_top1 > valid_accuracies['best']:
valid_accuracies['best'] = valid_a_top1
genotypes['best'] = search_model.genotype()
find_best = True
else: find_best = False
genotypes[epoch] = search_model.genotype()
logger.log('<<<--->>> The {:}-th epoch : {:}'.format(epoch_str, genotypes[epoch]))
# save checkpoint
save_path = save_checkpoint({'epoch' : epoch + 1,
'args' : deepcopy(xargs),
'search_model': search_model.state_dict(),
'w_optimizer' : w_optimizer.state_dict(),
'a_optimizer' : a_optimizer.state_dict(),
'w_scheduler' : w_scheduler.state_dict(),
'genotypes' : genotypes,
'valid_accuracies' : valid_accuracies},
model_base_path, logger)
last_info = save_checkpoint({
'epoch': epoch + 1,
'args' : deepcopy(args),
'last_checkpoint': save_path,
}, logger.path('info'), logger)
if find_best:
logger.log('<<<--->>> The {:}-th epoch : find the highest validation accuracy : {:.2f}%.'.format(epoch_str, valid_a_top1))
copy_checkpoint(model_base_path, model_best_path, logger)
with torch.no_grad():
#logger.log('arch-parameters :\n{:}'.format( nn.functional.softmax(search_model.arch_parameters, dim=-1).cpu() ))
logger.log('{:}'.format(search_model.show_alphas()))
if api is not None: logger.log('{:}'.format(api.query_by_arch( genotypes[epoch] )))
# measure elapsed time
epoch_time.update(time.time() - start_time)
start_time = time.time()
logger.log('\n' + '-'*100)
logger.log('DARTS-V1 : run {:} epochs, cost {:.1f} s, last-geno is {:}.'.format(total_epoch, search_time.sum, genotypes[total_epoch-1]))
if api is not None: logger.log('{:}'.format( api.query_by_arch(genotypes[total_epoch-1]) ))
logger.close()
def check_cor_for_bandit(meta_file,
test_epoch,
use_less_or_not,
is_rand=True,
need_print=False):
if isinstance(meta_file, API):
api = meta_file
else:
api = API(str(meta_file))
cifar10_currs = []
cifar10_valid = []
cifar10_test = []
cifar100_valid = []
cifar100_test = []
imagenet_test = []
imagenet_valid = []
for idx, arch in enumerate(api):
results = api.get_more_info(idx, "cifar10-valid", test_epoch - 1,
use_less_or_not, is_rand)
cifar10_currs.append(results["valid-accuracy"])
# --->>>>>
results = api.get_more_info(idx, "cifar10-valid", None, False, is_rand)
cifar10_valid.append(results["valid-accuracy"])
results = api.get_more_info(idx, "cifar10", None, False, is_rand)
cifar10_test.append(results["test-accuracy"])
results = api.get_more_info(idx, "cifar100", None, False, is_rand)
cifar100_test.append(results["test-accuracy"])
cifar100_valid.append(results["valid-accuracy"])
results = api.get_more_info(idx, "ImageNet16-120", None, False,
is_rand)
imagenet_test.append(results["test-accuracy"])
imagenet_valid.append(results["valid-accuracy"])
def get_cor(A, B):
return float(np.corrcoef(A, B)[0, 1])
cors = []
for basestr, xlist in zip(
["C-010-V", "C-010-T", "C-100-V", "C-100-T", "I16-V", "I16-T"],
[
cifar10_valid,
cifar10_test,
cifar100_valid,
cifar100_test,
imagenet_valid,
imagenet_test,
],
):
correlation = get_cor(cifar10_currs, xlist)
if need_print:
print(
"With {:3d}/{:}-epochs-training, the correlation between cifar10-valid and {:} is : {:}"
.format(
test_epoch,
"012" if use_less_or_not else "200",
basestr,
correlation,
))
cors.append(correlation)
# print ('With {:3d}/200-epochs-training, the correlation between cifar10-valid and {:} is : {:}'.format(test_epoch, basestr, get_cor(cifar10_valid_200, xlist)))
# print('-'*200)
# print('*'*230)
return cors
def main():
api = API(None)
info = api.get_more_info(100, 'cifar100', 199, False, True)
def main(xargs):
PID = os.getpid()
assert torch.cuda.is_available(), 'CUDA is not available.'
torch.backends.cudnn.enabled = True
torch.backends.cudnn.benchmark = False
torch.backends.cudnn.deterministic = True
prepare_seed(xargs.rand_seed)
if xargs.timestamp == 'none':
xargs.timestamp = "{:}".format(time.strftime('%h-%d-%C_%H-%M-%s', time.gmtime(time.time())))
train_data, valid_data, xshape, class_num = get_datasets(xargs.dataset, xargs.data_path, -1)
##### config & logging #####
config = edict()
config.class_num = class_num
config.xshape = xshape
config.batch_size = xargs.batch_size
xargs.save_dir = xargs.save_dir + \
"/repeat%d-prunNum%d-prec%d-%s-batch%d"%(
xargs.repeat, xargs.prune_number, xargs.precision, xargs.init, config["batch_size"]) + \
"/{:}/seed{:}".format(xargs.timestamp, xargs.rand_seed)
config.save_dir = xargs.save_dir
logger = prepare_logger(xargs)
###############
if xargs.dataset != 'imagenet-1k':
search_loader, train_loader, valid_loader = get_nas_search_loaders(train_data, valid_data, xargs.dataset, 'configs/', config.batch_size, xargs.workers)
else:
train_loader = torch.utils.data.DataLoader(train_data, batch_size=xargs.batch_size, shuffle=True, num_workers=args.workers, pin_memory=True)
logger.log('||||||| {:10s} ||||||| Train-Loader-Num={:}, batch size={:}'.format(xargs.dataset, len(train_loader), config.batch_size))
logger.log('||||||| {:10s} ||||||| Config={:}'.format(xargs.dataset, config))
search_space = get_search_spaces('cell', xargs.search_space_name)
if xargs.search_space_name == 'nas-bench-201':
model_config = edict({'name': 'DARTS-V1',
'C': 3, 'N': 1, 'depth': -1, 'use_stem': True,
'max_nodes': xargs.max_nodes, 'num_classes': class_num,
'space': search_space,
'affine': True, 'track_running_stats': bool(xargs.track_running_stats),
})
model_config_thin = edict({'name': 'DARTS-V1',
'C': 1, 'N': 1, 'depth': 1, 'use_stem': False,
'max_nodes': xargs.max_nodes, 'num_classes': class_num,
'space': search_space,
'affine': True, 'track_running_stats': bool(xargs.track_running_stats),
})
elif xargs.search_space_name == 'darts':
model_config = edict({'name': 'DARTS-V1',
'C': 1, 'N': 1, 'depth': 2, 'use_stem': True, 'stem_multiplier': 1,
'num_classes': class_num,
'space': search_space,
'affine': True, 'track_running_stats': bool(xargs.track_running_stats),
'super_type': xargs.super_type,
'steps': 4,
'multiplier': 4,
})
model_config_thin = edict({'name': 'DARTS-V1',
'C': 1, 'N': 1, 'depth': 2, 'use_stem': False, 'stem_multiplier': 1,
'max_nodes': xargs.max_nodes, 'num_classes': class_num,
'space': search_space,
'affine': True, 'track_running_stats': bool(xargs.track_running_stats),
'super_type': xargs.super_type,
'steps': 4,
'multiplier': 4,
})
network = get_cell_based_tiny_net(model_config)
logger.log('model-config : {:}'.format(model_config))
arch_parameters = [alpha.detach().clone() for alpha in network.get_alphas()]
for alpha in arch_parameters:
alpha[:, :] = 0
# TODO Linear_Region_Collector
lrc_model = Linear_Region_Collector(input_size=(1000, 1, 3, 3), sample_batch=3, dataset=xargs.dataset, data_path=xargs.data_path, seed=xargs.rand_seed)
# ### all params trainable (except train_bn) #########################
flop, param = get_model_infos(network, xshape)
logger.log('FLOP = {:.2f} M, Params = {:.2f} MB'.format(flop, param))
logger.log('search-space [{:} ops] : {:}'.format(len(search_space), search_space))
if xargs.arch_nas_dataset is None or xargs.search_space_name == 'darts':
api = None
else:
api = API(xargs.arch_nas_dataset)
logger.log('{:} create API = {:} done'.format(time_string(), api))
network = network.cuda()
genotypes = {}; genotypes['arch'] = {-1: network.genotype()}
arch_parameters_history = []
arch_parameters_history_npy = []
start_time = time.time()
epoch = -1
for alpha in arch_parameters:
alpha[:, 0] = -INF
arch_parameters_history.append([alpha.detach().clone() for alpha in arch_parameters])
arch_parameters_history_npy.append([alpha.detach().clone().cpu().numpy() for alpha in arch_parameters])
np.save(os.path.join(xargs.save_dir, "arch_parameters_history.npy"), arch_parameters_history_npy)
while not is_single_path(network):
epoch += 1
torch.cuda.empty_cache()
print("<< ============== JOB (PID = %d) %s ============== >>"%(PID, '/'.join(xargs.save_dir.split("/")[-6:])))
arch_parameters, op_pruned = prune_func_rank(xargs, arch_parameters, model_config, model_config_thin, train_loader, lrc_model, search_space,
precision=xargs.precision,
prune_number=xargs.prune_number
)
# rebuild supernet
network = get_cell_based_tiny_net(model_config)
network = network.cuda()
network.set_alphas(arch_parameters)
arch_parameters_history.append([alpha.detach().clone() for alpha in arch_parameters])
arch_parameters_history_npy.append([alpha.detach().clone().cpu().numpy() for alpha in arch_parameters])
np.save(os.path.join(xargs.save_dir, "arch_parameters_history.npy"), arch_parameters_history_npy)
genotypes['arch'][epoch] = network.genotype()
logger.log('operators remaining (1s) and prunned (0s)\n{:}'.format('\n'.join([str((alpha > -INF).int()) for alpha in network.get_alphas()])))
if xargs.search_space_name == 'darts':
print("===>>> Prune Edge Groups...")
arch_parameters = prune_func_rank_group(xargs, arch_parameters, model_config, model_config_thin, train_loader, lrc_model, search_space,
edge_groups=[(0, 2), (2, 5), (5, 9), (9, 14)], num_per_group=2,
precision=xargs.precision,
)
network = get_cell_based_tiny_net(model_config)
network = network.cuda()
network.set_alphas(arch_parameters)
arch_parameters_history.append([alpha.detach().clone() for alpha in arch_parameters])
arch_parameters_history_npy.append([alpha.detach().clone().cpu().numpy() for alpha in arch_parameters])
np.save(os.path.join(xargs.save_dir, "arch_parameters_history.npy"), arch_parameters_history_npy)
logger.log('<<<--->>> End: {:}'.format(network.genotype()))
logger.log('operators remaining (1s) and prunned (0s)\n{:}'.format('\n'.join([str((alpha > -INF).int()) for alpha in network.get_alphas()])))
end_time = time.time()
logger.log('\n' + '-'*100)
logger.log("Time spent: %d s"%(end_time - start_time))
# check the performance from the architecture dataset
if api is not None:
logger.log('{:}'.format(api.query_by_arch(genotypes['arch'][epoch])))
logger.close()
def __init__(self, dataset, apiloc):
self.dataset = dataset
self.api = API(apiloc, verbose=False)
self.epochs = '12'
def make(ww="without"):
from nas_201_api import NASBench201API as API
api = API('/Users/madoibito80/NAS-Bench-201-v1_0-e61699.pth')
macc = {}
nparams = {}
for mode in modes:
for way in ways:
if mode == "P" and way == "best":
continue
for trial in trials:
fname = str(trial) + "_" + mode
arcs = opener("./snapshot/" + ww + "/" + way + "/" + fname +
".txt",
freeze=50,
mode=mode)
accs = []
for i in range(len(arcs)):
if i % 100 == 0:
print(mode, way, trial, i)
index = api.query_index_by_arch(arcs[i])
flg = True
try:
info = api.query_meta_info_by_index(index)
except:
print("error: ", mode, way, trial, i)
flg = False
if flg:
res = info.get_metrics('cifar10', 'ori-test', None,
False)
# cifar10 : training the model on the CIFAR-10 training + validation set.
# ,criteria , , False=3average of NAS-Bench)
acc = res['accuracy']
accs.append(float(acc))
else:
accs.append(accs[-1])
if trial == 0:
macc[mode + way] = np.zeros((len(trials), len(accs)))
nparams[mode + way] = np.zeros((len(trials), 1))
macc[mode + way][trial] = np.array(accs)
# get final performance for table
try:
print(dir(info))
except:
print("no dir")
metric = info.get_comput_costs('cifar10')
flop, param, latency = metric['flops'], metric[
'params'], metric['latency']
nparams[mode + way][trial] = float(param)
f = open("./" + ww + ".pickle", "wb")
pickle.dump(macc, f)
pickle.dump(nparams, f)
f.close()
print(macc)
print(nparams)
return macc
def __init__(self, data_dir, task='cifar10-valid', log_scale=True, negative=True,
use_12_epochs_result=False,
seed=None):
"""
data_dir: data directory that contains NAS-Bench-201-v1_0-e61699.pth file
task: the target image tasks. Options: cifar10-valid, cifar100, ImageNet16-120
log_scale: whether output the objective in log scale
negative: whether output the objective in negative form
use_12_epochs_result: whether use the statistics at the end of training of the 12th epoch instead of all the
way till the end.
seed: set the random seed to access trained model performance: Options: 0, 1, 2
seed=None will select the seed randomly
"""
self.api = API(os.path.join(data_dir, 'NAS-Bench-201-v1_1-096897.pth'))
if isinstance(task, list):
task = task[0]
self.task = task
self.use_12_epochs_result = use_12_epochs_result
if task == 'cifar10-valid':
best_val_arch_index = 6111
best_val_acc = 91.60666665039064 / 100
best_test_arch_index = 1459
best_test_acc = 91.52333333333333 / 100
elif task == 'cifar100':
best_val_arch_index = 9930
best_val_acc = 73.49333323567708 / 100
best_test_arch_index = 9930
best_test_acc = 73.51333326009114 / 100
elif task == 'ImageNet16-120':
best_val_arch_index = 10676
best_val_acc = 46.766666727701825 / 100
best_test_arch_index = 857
best_test_acc = 47.311111097547744 / 100
else:
raise NotImplementedError("task" + str(task) + " is not implemented in the dataset.")
if log_scale:
best_val_acc = np.log(best_val_acc)
best_val_err = 1. - best_val_acc
best_test_err = 1. - best_test_acc
if log_scale:
best_val_err = np.log(best_val_err)
best_test_err = np.log(best_val_err)
if negative:
best_val_err = -best_val_err
best_test_err = -best_test_err
self.best_val_err = best_val_err
self.best_test_err = best_test_err
self.best_val_acc = best_val_acc
self.best_test_acc = best_test_acc
super(NAS201, self).__init__(dim=None, optimum_location=best_test_arch_index, optimal_val=best_test_err,
bounds=None)
self.log_scale = log_scale
self.seed = seed
self.X = []
self.y_valid_acc = []
self.y_test_acc = []
self.costs = []
self.negative = negative
def main(xargs):
assert torch.cuda.is_available(), 'CUDA is not available.'
torch.backends.cudnn.enabled = True
torch.backends.cudnn.benchmark = False
torch.backends.cudnn.deterministic = True
torch.set_num_threads(xargs.workers)
prepare_seed(xargs.rand_seed)
logger = prepare_logger(args)
train_data, valid_data, xshape, class_num = get_datasets(
xargs.dataset, xargs.data_path, -1)
#config_path = 'configs/nas-benchmark/algos/GDAS.config'
config = load_config(xargs.config_path, {
'class_num': class_num,
'xshape': xshape
}, logger)
search_loader, _, valid_loader = get_nas_search_loaders(
train_data, valid_data, xargs.dataset, 'configs/nas-benchmark/',
config.batch_size, xargs.workers)
logger.log(
'||||||| {:10s} ||||||| Search-Loader-Num={:}, batch size={:}'.format(
xargs.dataset, len(search_loader), config.batch_size))
logger.log('||||||| {:10s} ||||||| Config={:}'.format(
xargs.dataset, config))
search_space = get_search_spaces('cell', xargs.search_space_name)
if xargs.model_config is None and not args.constrain:
model_config = dict2config(
{
'name': 'SNAS',
'C': xargs.channel,
'N': xargs.num_cells,
'max_nodes': xargs.max_nodes,
'num_classes': class_num,
'space': search_space,
'inp_size': 0,
'affine': False,
'track_running_stats': bool(xargs.track_running_stats)
}, None)
elif xargs.model_config is None:
model_config = dict2config(
{
'name': 'SNAS',
'C': xargs.channel,
'N': xargs.num_cells,
'max_nodes': xargs.max_nodes,
'num_classes': class_num,
'space': search_space,
'inp_size': 32,
'affine': False,
'track_running_stats': bool(xargs.track_running_stats)
}, None)
else:
model_config = load_config(
xargs.model_config, {
'num_classes': class_num,
'space': search_space,
'affine': False,
'track_running_stats': bool(xargs.track_running_stats)
}, None)
search_model = get_cell_based_tiny_net(model_config)
#logger.log('search-model :\n{:}'.format(search_model))
logger.log('model-config : {:}'.format(model_config))
w_optimizer, w_scheduler, criterion = get_optim_scheduler(
search_model.get_weights(), config)
a_optimizer = torch.optim.Adam(search_model.get_alphas(),
lr=xargs.arch_learning_rate,
betas=(0.5, 0.999),
weight_decay=xargs.arch_weight_decay)
logger.log('w-optimizer : {:}'.format(w_optimizer))
logger.log('a-optimizer : {:}'.format(a_optimizer))
logger.log('w-scheduler : {:}'.format(w_scheduler))
logger.log('criterion : {:}'.format(criterion))
flop, param = get_model_infos(search_model, xshape)
#logger.log('{:}'.format(search_model))
logger.log('FLOP = {:.2f} M, Params = {:.2f} MB'.format(flop, param))
logger.log('search-space [{:} ops] : {:}'.format(len(search_space),
search_space))
if xargs.arch_nas_dataset is None:
api = None
else:
api = API(xargs.arch_nas_dataset)
logger.log('{:} create API = {:} done'.format(time_string(), api))
last_info, model_base_path, model_best_path = logger.path(
'info'), logger.path('model'), logger.path('best')
network, criterion = torch.nn.DataParallel(
search_model).cuda(), criterion.cuda()
#network, criterion = search_model.cuda(), criterion.cuda()
if last_info.exists(): # automatically resume from previous checkpoint
logger.log("=> loading checkpoint of the last-info '{:}' start".format(
last_info))
last_info = torch.load(last_info)
start_epoch = last_info['epoch']
checkpoint = torch.load(last_info['last_checkpoint'])
genotypes = checkpoint['genotypes']
valid_accuracies = checkpoint['valid_accuracies']
search_model.load_state_dict(checkpoint['search_model'])
w_scheduler.load_state_dict(checkpoint['w_scheduler'])
w_optimizer.load_state_dict(checkpoint['w_optimizer'])
a_optimizer.load_state_dict(checkpoint['a_optimizer'])
logger.log(
"=> loading checkpoint of the last-info '{:}' start with {:}-th epoch."
.format(last_info, start_epoch))
else:
logger.log("=> do not find the last-info file : {:}".format(last_info))
start_epoch, valid_accuracies, genotypes = 0, {
'best': -1
}, {
-1: search_model.genotype()
}
# start training
start_time, search_time, epoch_time, total_epoch = time.time(
), AverageMeter(), AverageMeter(), config.epochs + config.warmup
sampled_weights = []
for epoch in range(start_epoch, total_epoch + config.t_epochs):
w_scheduler.update(epoch, 0.0)
need_time = 'Time Left: {:}'.format(
convert_secs2time(
epoch_time.val * (total_epoch - epoch + config.t_epochs),
True))
epoch_str = '{:03d}-{:03d}'.format(epoch, total_epoch)
search_model.set_tau(xargs.tau_max -
(xargs.tau_max - xargs.tau_min) * epoch /
(total_epoch - 1))
logger.log('\n[Search the {:}-th epoch] {:}, tau={:}, LR={:}'.format(
epoch_str, need_time, search_model.get_tau(),
min(w_scheduler.get_lr())))
if epoch < total_epoch:
search_w_loss, search_w_top1, search_w_top5, valid_a_loss , valid_a_top1 , valid_a_top5 \
= search_func(search_loader, network, criterion, w_scheduler, w_optimizer, a_optimizer, epoch_str, xargs.print_freq, logger, xargs.bilevel)
else:
try:
search_w_loss, search_w_top1, search_w_top5, valid_a_loss , valid_a_top1 , valid_a_top5, arch_iter \
= train_func(search_loader, network, criterion, w_scheduler, w_optimizer, epoch_str, xargs.print_freq, sampled_weights[0], arch_iter, logger)
except IndexError:
weights = search_model.sample_weights(100)
sampled_weights.append(weights)
arch_iter = iter(weights)
search_w_loss, search_w_top1, search_w_top5, valid_a_loss , valid_a_top1 , valid_a_top5, arch_iter \
= train_func(search_loader, network, criterion, w_scheduler, w_optimizer, epoch_str, xargs.print_freq, sampled_weights[0], arch_iter, logger)
search_time.update(time.time() - start_time)
logger.log(
'[{:}] searching : loss={:.2f}, accuracy@1={:.2f}%, accuracy@5={:.2f}%, time-cost={:.1f} s'
.format(epoch_str, search_w_loss, search_w_top1, search_w_top5,
search_time.sum))
logger.log(
'[{:}] evaluate : loss={:.2f}, accuracy@1={:.2f}%, accuracy@5={:.2f}%'
.format(epoch_str, valid_a_loss, valid_a_top1, valid_a_top5))
if (epoch + 1) % 50 == 0 and not config.t_epochs:
weights = search_model.sample_weights(100)
sampled_weights.append(weights)
elif (epoch + 1) == total_epoch and config.t_epochs:
weights = search_model.sample_weights(100)
sampled_weights.append(weights)
arch_iter = iter(weights)
# validate with single arch
single_weight = search_model.sample_weights(1)[0]
single_valid_acc = AverageMeter()
network.eval()
for i in range(10):
try:
val_input, val_target = next(valid_iter)
except Exception as e:
valid_iter = iter(valid_loader)
val_input, val_target = next(valid_iter)
n_val = val_input.size(0)
with torch.no_grad():
val_target = val_target.cuda(non_blocking=True)
_, logits, _ = network(val_input, weights=single_weight)
val_acc1, val_acc5 = obtain_accuracy(logits.data,
val_target.data,
topk=(1, 5))
single_valid_acc.update(val_acc1.item(), n_val)
logger.log('[{:}] valid : accuracy = {:.2f}'.format(
epoch_str, single_valid_acc.avg))
# check the best accuracy
valid_accuracies[epoch] = valid_a_top1
if valid_a_top1 > valid_accuracies['best']:
valid_accuracies['best'] = valid_a_top1
genotypes['best'] = search_model.genotype()
find_best = True
else:
find_best = False
if epoch < total_epoch:
genotypes[epoch] = search_model.genotype()
logger.log('<<<--->>> The {:}-th epoch : {:}'.format(
epoch_str, genotypes[epoch]))
# save checkpoint
save_path = save_checkpoint(
{
'epoch': epoch + 1,
'args': deepcopy(xargs),
'search_model': search_model.state_dict(),
'w_optimizer': w_optimizer.state_dict(),
'a_optimizer': a_optimizer.state_dict(),
'w_scheduler': w_scheduler.state_dict(),
'genotypes': genotypes,
'valid_accuracies': valid_accuracies
}, model_base_path, logger)
last_info = save_checkpoint(
{
'epoch': epoch + 1,
'args': deepcopy(args),
'last_checkpoint': save_path,
}, logger.path('info'), logger)
if find_best:
logger.log(
'<<<--->>> The {:}-th epoch : find the highest validation accuracy : {:.2f}%.'
.format(epoch_str, valid_a_top1))
copy_checkpoint(model_base_path, model_best_path, logger)
with torch.no_grad():
logger.log('{:}'.format(search_model.show_alphas()))
if api is not None and epoch < total_epoch:
logger.log('{:}'.format(api.query_by_arch(genotypes[epoch])))
# measure elapsed time
epoch_time.update(time.time() - start_time)
start_time = time.time()
network.eval()
# Evaluate the architectures sampled throughout the search
for i in range(len(sampled_weights) - 1):
logger.log('Sample eval : epoch {}'.format((i + 1) * 50 - 1))
for w in sampled_weights[i]:
sample_valid_acc = AverageMeter()
for i in range(10):
try:
val_input, val_target = next(valid_iter)
except Exception as e:
valid_iter = iter(valid_loader)
val_input, val_target = next(valid_iter)
n_val = val_input.size(0)
with torch.no_grad():
val_target = val_target.cuda(non_blocking=True)
_, logits, _ = network(val_input, weights=w)
val_acc1, val_acc5 = obtain_accuracy(logits.data,
val_target.data,
topk=(1, 5))
sample_valid_acc.update(val_acc1.item(), n_val)
w_gene = search_model.genotype(w)
if api is not None:
ind = api.query_index_by_arch(w_gene)
info = api.query_meta_info_by_index(ind)
metrics = info.get_metrics('cifar10', 'ori-test')
acc = metrics['accuracy']
else:
acc = 0.0
logger.log(
'sample valid : val_acc = {:.2f} test_acc = {:.2f}'.format(
sample_valid_acc.avg, acc))
# Evaluate the final sampling separately to find the top 10 architectures
logger.log('Final sample eval')
final_archs = []
for w in sampled_weights[-1]:
sample_valid_acc = AverageMeter()
for i in range(10):
try:
val_input, val_target = next(valid_iter)
except Exception as e:
valid_iter = iter(valid_loader)
val_input, val_target = next(valid_iter)
n_val = val_input.size(0)
with torch.no_grad():
val_target = val_target.cuda(non_blocking=True)
_, logits, _ = network(val_input, weights=w)
val_acc1, val_acc5 = obtain_accuracy(logits.data,
val_target.data,
topk=(1, 5))
sample_valid_acc.update(val_acc1.item(), n_val)
w_gene = search_model.genotype(w)
if api is not None:
ind = api.query_index_by_arch(w_gene)
info = api.query_meta_info_by_index(ind)
metrics = info.get_metrics('cifar10', 'ori-test')
acc = metrics['accuracy']
else:
acc = 0.0
logger.log('sample valid : val_acc = {:.2f} test_acc = {:.2f}'.format(
sample_valid_acc.avg, acc))
final_archs.append((w, sample_valid_acc.avg))
top_10 = sorted(final_archs, key=lambda x: x[1], reverse=True)[:10]
# Evaluate the top 10 architectures on the entire validation set
logger.log('Evaluating top archs')
for w, prev_acc in top_10:
full_valid_acc = AverageMeter()
for val_input, val_target in valid_loader:
n_val = val_input.size(0)
with torch.no_grad():
val_target = val_target.cuda(non_blocking=True)
_, logits, _ = network(val_input, weights=w)
val_acc1, val_acc5 = obtain_accuracy(logits.data,
val_target.data,
topk=(1, 5))
full_valid_acc.update(val_acc1.item(), n_val)
w_gene = search_model.genotype(w)
logger.log('genotype {}'.format(w_gene))
if api is not None:
ind = api.query_index_by_arch(w_gene)
info = api.query_meta_info_by_index(ind)
metrics = info.get_metrics('cifar10', 'ori-test')
acc = metrics['accuracy']
else:
acc = 0.0
logger.log(
'full valid : val_acc = {:.2f} test_acc = {:.2f} pval_acc = {:.2f}'
.format(full_valid_acc.avg, acc, prev_acc))
logger.log('\n' + '-' * 100)
# check the performance from the architecture dataset
logger.log(
'SNAS : run {:} epochs, cost {:.1f} s, last-geno is {:}.'.format(
total_epoch, search_time.sum, genotypes[total_epoch - 1]))
if api is not None:
logger.log('{:}'.format(api.query_by_arch(genotypes[total_epoch - 1])))
logger.close()
def main():
torch.set_num_threads(3)
if not torch.cuda.is_available():
logging.info('no gpu device available')
sys.exit(1)
np.random.seed(args.seed)
torch.cuda.set_device(args.gpu)
cudnn.benchmark = True
torch.manual_seed(args.seed)
cudnn.enabled = True
torch.cuda.manual_seed(args.seed)
logging.info('gpu device = %d' % args.gpu)
logging.info("args = %s", args)
if not 'debug' in args.save:
api = API('pth file path')
criterion = nn.CrossEntropyLoss()
criterion = criterion.cuda()
if args.method == 'gdas' or args.method == 'snas':
# Create the decrease step for the gumbel softmax temperature
tau_step = (args.tau_min - args.tau_max) / args.epochs
tau_epoch = args.tau_max
if args.method == 'gdas':
model = TinyNetworkGDAS(C=args.init_channels,
N=5,
max_nodes=4,
num_classes=n_classes,
criterion=criterion,
search_space=NAS_BENCH_201)
else:
model = TinyNetwork(C=args.init_channels,
N=5,
max_nodes=4,
num_classes=n_classes,
criterion=criterion,
search_space=NAS_BENCH_201,
k=args.k,
species='gumbel')
elif args.method == 'dirichlet':
model = TinyNetwork(C=args.init_channels,
N=5,
max_nodes=4,
num_classes=n_classes,
criterion=criterion,
search_space=NAS_BENCH_201,
k=args.k,
species='dirichlet')
elif args.method == 'darts':
model = TinyNetwork(C=args.init_channels,
N=5,
max_nodes=4,
num_classes=n_classes,
criterion=criterion,
search_space=NAS_BENCH_201,
k=args.k,
species='softmax')
model = model.cuda()
logging.info("param size = %fMB", utils.count_parameters_in_MB(model))
optimizer = torch.optim.SGD(model.get_weights(),
args.learning_rate,
momentum=args.momentum,
weight_decay=args.weight_decay)
if args.dataset == 'cifar10':
train_transform, valid_transform = utils._data_transforms_cifar10(args)
train_data = dset.CIFAR10(root=args.data,
train=True,
download=True,
transform=train_transform)
elif args.dataset == 'cifar100':
train_transform, valid_transform = utils._data_transforms_cifar100(
args)
train_data = dset.CIFAR100(root=args.data,
train=True,
download=True,
transform=train_transform)
elif args.dataset == 'svhn':
train_transform, valid_transform = utils._data_transforms_svhn(args)
train_data = dset.SVHN(root=args.data,
split='train',
download=True,
transform=train_transform)
elif args.dataset == 'imagenet16-120':
import torchvision.transforms as transforms
from nasbench201.DownsampledImageNet import ImageNet16
mean = [x / 255 for x in [122.68, 116.66, 104.01]]
std = [x / 255 for x in [63.22, 61.26, 65.09]]
lists = [
transforms.RandomHorizontalFlip(),
transforms.RandomCrop(16, padding=2),
transforms.ToTensor(),
transforms.Normalize(mean, std)
]
train_transform = transforms.Compose(lists)
train_data = ImageNet16(root=os.path.join(args.data, 'imagenet16'),
train=True,
transform=train_transform,
use_num_of_class_only=120)
assert len(train_data) == 151700
num_train = len(train_data)
indices = list(range(num_train))
split = int(np.floor(args.train_portion * num_train))
train_queue = torch.utils.data.DataLoader(
train_data,
batch_size=args.batch_size,
sampler=torch.utils.data.sampler.SubsetRandomSampler(indices[:split]),
pin_memory=True)
valid_queue = torch.utils.data.DataLoader(
train_data,
batch_size=args.batch_size,
sampler=torch.utils.data.sampler.SubsetRandomSampler(
indices[split:num_train]),
pin_memory=True)
scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(
optimizer, float(args.epochs), eta_min=args.learning_rate_min)
architect = Architect(model, args)
for epoch in range(args.epochs):
lr = scheduler.get_lr()[0]
logging.info('epoch %d lr %e', epoch, lr)
genotype = model.genotype()
logging.info('genotype = %s', genotype)
model.show_arch_parameters()
# training
train_acc, train_obj = train(train_queue, valid_queue, model,
architect, criterion, optimizer, lr,
epoch)
logging.info('train_acc %f', train_acc)
# validation
valid_acc, valid_obj = infer(valid_queue, model, criterion)
logging.info('valid_acc %f', valid_acc)
if not 'debug' in args.save:
# nasbench201
result = api.query_by_arch(model.genotype())
logging.info('{:}'.format(result))
cifar10_train, cifar10_test, cifar100_train, cifar100_valid, \
cifar100_test, imagenet16_train, imagenet16_valid, imagenet16_test = distill(result)
logging.info('cifar10 train %f test %f', cifar10_train,
cifar10_test)
logging.info('cifar100 train %f valid %f test %f', cifar100_train,
cifar100_valid, cifar100_test)
logging.info('imagenet16 train %f valid %f test %f',
imagenet16_train, imagenet16_valid, imagenet16_test)
# tensorboard
writer.add_scalars('accuracy', {
'train': train_acc,
'valid': valid_acc
}, epoch)
writer.add_scalars('loss', {
'train': train_obj,
'valid': valid_obj
}, epoch)
writer.add_scalars('nasbench201/cifar10', {
'train': cifar10_train,
'test': cifar10_test
}, epoch)
writer.add_scalars(
'nasbench201/cifar100', {
'train': cifar100_train,
'valid': cifar100_valid,
'test': cifar100_test
}, epoch)
writer.add_scalars(
'nasbench201/imagenet16', {
'train': imagenet16_train,
'valid': imagenet16_valid,
'test': imagenet16_test
}, epoch)
utils.save_checkpoint(
{
'epoch': epoch + 1,
'state_dict': model.state_dict(),
'optimizer': optimizer.state_dict(),
'alpha': model.arch_parameters()
}, False, args.save)
scheduler.step()
if args.method == 'gdas' or args.method == 'snas':
# Decrease the temperature for the gumbel softmax linearly
tau_epoch += tau_step
logging.info('tau %f', tau_epoch)
model.set_tau(tau_epoch)
writer.close()
def __init__(self):
"""Construct the Nasbench201 class."""
super(Nasbench201, self).__init__()
self.args.data_path = FileOps.download_dataset(self.args.data_path)
self.nasbench201_api = API('self.args.data_path')
def main(xargs):
assert torch.cuda.is_available(), 'CUDA is not available.'
torch.backends.cudnn.enabled = True
torch.backends.cudnn.benchmark = False
torch.backends.cudnn.deterministic = True
torch.set_num_threads(xargs.workers)
prepare_seed(xargs.rand_seed)
logger = prepare_logger(args)
train_data, valid_data, xshape, class_num = get_datasets(
xargs.dataset, xargs.data_path, -1)
if xargs.overwite_epochs is None:
extra_info = {'class_num': class_num, 'xshape': xshape}
else:
extra_info = {
'class_num': class_num,
'xshape': xshape,
'epochs': xargs.overwite_epochs
}
config = load_config(xargs.config_path, extra_info, logger)
search_loader, train_loader, valid_loader = get_nas_search_loaders(
train_data, valid_data, xargs.dataset, 'configs/nas-benchmark/',
(config.batch_size, config.test_batch_size), xargs.workers)
logger.log(
'||||||| {:10s} ||||||| Search-Loader-Num={:}, Valid-Loader-Num={:}, batch size={:}'
.format(xargs.dataset, len(search_loader), len(valid_loader),
config.batch_size))
logger.log('||||||| {:10s} ||||||| Config={:}'.format(
xargs.dataset, config))
search_space = get_search_spaces(xargs.search_space, 'nas-bench-301')
model_config = dict2config(
dict(name='generic',
super_type='search-shape',
candidate_Cs=search_space['candidates'],
max_num_Cs=search_space['numbers'],
num_classes=class_num,
genotype=args.genotype,
affine=bool(xargs.affine),
track_running_stats=bool(xargs.track_running_stats)), None)
logger.log('search space : {:}'.format(search_space))
logger.log('model config : {:}'.format(model_config))
search_model = get_cell_based_tiny_net(model_config)
search_model.set_algo(xargs.algo)
logger.log('{:}'.format(search_model))
w_optimizer, w_scheduler, criterion = get_optim_scheduler(
search_model.weights, config)
a_optimizer = torch.optim.Adam(search_model.alphas,
lr=xargs.arch_learning_rate,
betas=(0.5, 0.999),
weight_decay=xargs.arch_weight_decay,
eps=xargs.arch_eps)
logger.log('w-optimizer : {:}'.format(w_optimizer))
logger.log('a-optimizer : {:}'.format(a_optimizer))
logger.log('w-scheduler : {:}'.format(w_scheduler))
logger.log('criterion : {:}'.format(criterion))
params = count_parameters_in_MB(search_model)
logger.log('The parameters of the search model = {:.2f} MB'.format(params))
logger.log('search-space : {:}'.format(search_space))
if bool(xargs.use_api):
api = API(verbose=False)
else:
api = None
logger.log('{:} create API = {:} done'.format(time_string(), api))
last_info, model_base_path, model_best_path = logger.path(
'info'), logger.path('model'), logger.path('best')
network, criterion = search_model.cuda(), criterion.cuda(
) # use a single GPU
last_info, model_base_path, model_best_path = logger.path(
'info'), logger.path('model'), logger.path('best')
if last_info.exists(): # automatically resume from previous checkpoint
logger.log("=> loading checkpoint of the last-info '{:}' start".format(
last_info))
last_info = torch.load(last_info)
start_epoch = last_info['epoch']
checkpoint = torch.load(last_info['last_checkpoint'])
genotypes = checkpoint['genotypes']
valid_accuracies = checkpoint['valid_accuracies']
search_model.load_state_dict(checkpoint['search_model'])
w_scheduler.load_state_dict(checkpoint['w_scheduler'])
w_optimizer.load_state_dict(checkpoint['w_optimizer'])
a_optimizer.load_state_dict(checkpoint['a_optimizer'])
logger.log(
"=> loading checkpoint of the last-info '{:}' start with {:}-th epoch."
.format(last_info, start_epoch))
else:
logger.log("=> do not find the last-info file : {:}".format(last_info))
start_epoch, valid_accuracies, genotypes = 0, {
'best': -1
}, {
-1: network.random
}
# start training
start_time, search_time, epoch_time, total_epoch = time.time(
), AverageMeter(), AverageMeter(), config.epochs + config.warmup
for epoch in range(start_epoch, total_epoch):
w_scheduler.update(epoch, 0.0)
need_time = 'Time Left: {:}'.format(
convert_secs2time(epoch_time.val * (total_epoch - epoch), True))
epoch_str = '{:03d}-{:03d}'.format(epoch, total_epoch)
logger.log('\n[Search the {:}-th epoch] {:}, LR={:}'.format(
epoch_str, need_time, min(w_scheduler.get_lr())))
if xargs.algo == 'fbv2' or xargs.algo == 'tas':
network.set_tau(xargs.tau_max -
(xargs.tau_max - xargs.tau_min) * epoch /
(total_epoch - 1))
logger.log('[RESET tau as : {:}]'.format(network.tau))
search_w_loss, search_w_top1, search_w_top5, search_a_loss, search_a_top1, search_a_top5 \
= search_func(search_loader, network, criterion, w_scheduler, w_optimizer, a_optimizer, xargs.algo, epoch_str, xargs.print_freq, logger)
search_time.update(time.time() - start_time)
logger.log(
'[{:}] search [base] : loss={:.2f}, accuracy@1={:.2f}%, accuracy@5={:.2f}%, time-cost={:.1f} s'
.format(epoch_str, search_w_loss, search_w_top1, search_w_top5,
search_time.sum))
logger.log(
'[{:}] search [arch] : loss={:.2f}, accuracy@1={:.2f}%, accuracy@5={:.2f}%'
.format(epoch_str, search_a_loss, search_a_top1, search_a_top5))
genotype = network.genotype
logger.log('[{:}] - [get_best_arch] : {:}'.format(epoch_str, genotype))
valid_a_loss, valid_a_top1, valid_a_top5 = valid_func(
valid_loader, network, criterion, logger)
logger.log(
'[{:}] evaluate : loss={:.2f}, accuracy@1={:.2f}%, accuracy@5={:.2f}% | {:}'
.format(epoch_str, valid_a_loss, valid_a_top1, valid_a_top5,
genotype))
valid_accuracies[epoch] = valid_a_top1
genotypes[epoch] = genotype
logger.log('<<<--->>> The {:}-th epoch : {:}'.format(
epoch_str, genotypes[epoch]))
# save checkpoint
save_path = save_checkpoint(
{
'epoch': epoch + 1,
'args': deepcopy(xargs),
'search_model': search_model.state_dict(),
'w_optimizer': w_optimizer.state_dict(),
'a_optimizer': a_optimizer.state_dict(),
'w_scheduler': w_scheduler.state_dict(),
'genotypes': genotypes,
'valid_accuracies': valid_accuracies
}, model_base_path, logger)
last_info = save_checkpoint(
{
'epoch': epoch + 1,
'args': deepcopy(args),
'last_checkpoint': save_path,
}, logger.path('info'), logger)
with torch.no_grad():
logger.log('{:}'.format(search_model.show_alphas()))
if api is not None:
logger.log('{:}'.format(api.query_by_arch(genotypes[epoch], '90')))
# measure elapsed time
epoch_time.update(time.time() - start_time)
start_time = time.time()
# the final post procedure : count the time
start_time = time.time()
genotype = network.genotype
search_time.update(time.time() - start_time)
valid_a_loss, valid_a_top1, valid_a_top5 = valid_func(
valid_loader, network, criterion, logger)
logger.log(
'Last : the gentotype is : {:}, with the validation accuracy of {:.3f}%.'
.format(genotype, valid_a_top1))
logger.log('\n' + '-' * 100)
# check the performance from the architecture dataset
logger.log('[{:}] run {:} epochs, cost {:.1f} s, last-geno is {:}.'.format(
xargs.algo, total_epoch, search_time.sum, genotype))
if api is not None:
logger.log('{:}'.format(api.query_by_arch(genotype, '90')))
logger.close()
def main(xargs):
assert torch.cuda.is_available(), 'CUDA is not available.'
torch.backends.cudnn.enabled = True
torch.backends.cudnn.benchmark = False
torch.backends.cudnn.deterministic = True
torch.set_num_threads(xargs.workers)
prepare_seed(xargs.rand_seed)
logger = prepare_logger(args)
train_data, valid_data, xshape, class_num = get_datasets(
xargs.dataset, xargs.data_path, -1)
#config_path = 'configs/nas-benchmark/algos/GDAS.config'
config = load_config(xargs.config_path, {
'class_num': class_num,
'xshape': xshape
}, logger)
search_loader, _, valid_loader = get_nas_search_loaders(
train_data, valid_data, xargs.dataset, 'configs/nas-benchmark/',
config.batch_size, xargs.workers)
if xargs.ood_inner or xargs.ood_outer:
mean = [x / 255 for x in [125.3, 123.0, 113.9]]
std = [x / 255 for x in [63.0, 62.1, 66.7]]
# lists = [transforms.RandomHorizontalFlip(), transforms.RandomCrop(32, padding=4), transforms.ToTensor(), transforms.Normalize(mean, std)]
lists = [transforms.ToTensor(), transforms.Normalize(mean, std)]
# lists += [CUTOUT(-1)]
ood_transform = transforms.Compose(lists)
ood_data = dset.SVHN(root=args.data_path,
split='train',
download=True,
transform=ood_transform)
ood_loader = torch.utils.data.DataLoader(
ood_data,
batch_size=config.batch_size,
sampler=torch.utils.data.sampler.SubsetRandomSampler(
list(range(len(ood_data)))[:len(train_data)]),
pin_memory=True,
num_workers=xargs.workers)
else:
ood_loader = None
logger.log(
'||||||| {:10s} ||||||| Search-Loader-Num={:}, batch size={:}'.format(
xargs.dataset, len(search_loader), config.batch_size))
logger.log('||||||| {:10s} ||||||| Config={:}'.format(
xargs.dataset, config))
global search_space
search_space = get_search_spaces('cell', xargs.search_space_name)
if xargs.model_config is None:
model_config = dict2config(
{
'name': 'GDAS',
'C': xargs.channel,
'N': xargs.num_cells,
'max_nodes': xargs.max_nodes,
'num_classes': class_num,
'space': search_space,
'affine': False,
'track_running_stats': bool(xargs.track_running_stats),
}, None)
else:
model_config = load_config(
xargs.model_config, {
'num_classes': class_num,
'space': search_space,
'affine': False,
'track_running_stats': bool(xargs.track_running_stats)
}, None)
search_model = get_cell_based_tiny_net(model_config)
# logger.log('search-model :\n{:}'.format(search_model))
logger.log('model-config : {:}'.format(model_config))
w_optimizer, w_scheduler, criterion = get_optim_scheduler(
search_model.get_weights(), config)
a_optimizer = torch.optim.Adam(search_model.get_alphas(),
lr=xargs.arch_learning_rate,
betas=(0.5, 0.999),
weight_decay=xargs.arch_weight_decay)
logger.log('w-optimizer : {:}'.format(w_optimizer))
logger.log('a-optimizer : {:}'.format(a_optimizer))
logger.log('w-scheduler : {:}'.format(w_scheduler))
logger.log('criterion : {:}'.format(criterion))
flop, param = get_model_infos(search_model, xshape)
logger.log('FLOP = {:.2f} M, Params = {:.2f} MB'.format(flop, param))
logger.log('search-space [{:} ops] : {:}'.format(len(search_space),
search_space))
if xargs.arch_nas_dataset is None:
api = None
else:
api = API(xargs.arch_nas_dataset)
logger.log('{:} create API = {:} done'.format(time_string(), api))
last_info, model_base_path, model_best_path = logger.path(
'info'), logger.path('model'), logger.path('best')
# network, criterion = torch.nn.DataParallel(search_model).cuda(), criterion.cuda()
network, criterion = search_model.cuda(), criterion.cuda()
if last_info.exists(): # automatically resume from previous checkpoint
logger.log("=> loading checkpoint of the last-info '{:}' start".format(
last_info))
last_info = torch.load(last_info)
start_epoch = last_info['epoch']
checkpoint = torch.load(last_info['last_checkpoint'])
genotypes = checkpoint['genotypes']
valid_accuracies = checkpoint['valid_accuracies']
search_model.load_state_dict(checkpoint['search_model'])
w_scheduler.load_state_dict(checkpoint['w_scheduler'])
w_optimizer.load_state_dict(checkpoint['w_optimizer'])
a_optimizer.load_state_dict(checkpoint['a_optimizer'])
logger.log(
"=> loading checkpoint of the last-info '{:}' start with {:}-th epoch."
.format(last_info, start_epoch))
else:
logger.log("=> do not find the last-info file : {:}".format(last_info))
start_epoch, valid_accuracies, genotypes = 0, {
'best': -1
}, {
-1: search_model.genotype()
}
# start training
start_time, search_time, epoch_time, total_epoch = time.time(
), AverageMeter(), AverageMeter(), config.epochs + config.warmup
for epoch in range(start_epoch, total_epoch):
w_scheduler.update(epoch, 0.0)
need_time = 'Time Left: {:}'.format(
convert_secs2time(epoch_time.val * (total_epoch - epoch), True))
epoch_str = '{:03d}-{:03d}'.format(epoch, total_epoch)
search_model.set_tau(xargs.tau_max -
(xargs.tau_max - xargs.tau_min) * epoch /
(total_epoch - 1))
logger.log('\n[Search the {:}-th epoch] {:}, tau={:}, LR={:}'.format(
epoch_str, need_time, search_model.get_tau(),
min(w_scheduler.get_lr())))
search_w_loss, search_w_top1, search_w_top5, valid_a_loss , valid_a_top1 , valid_a_top5 \
= search_func(search_loader, network, criterion, w_scheduler, w_optimizer, a_optimizer, epoch_str, xargs, logger, ood_loader)
search_time.update(time.time() - start_time)
logger.log(
'[{:}] searching : loss={:.2f}, accuracy@1={:.2f}%, accuracy@5={:.2f}%, time-cost={:.1f} s'
.format(epoch_str, search_w_loss, search_w_top1, search_w_top5,
search_time.sum))
logger.log(
'[{:}] evaluate : loss={:.2f}, accuracy@1={:.2f}%, accuracy@5={:.2f}%'
.format(epoch_str, valid_a_loss, valid_a_top1, valid_a_top5))
# check the best accuracy
valid_accuracies[epoch] = valid_a_top1
if valid_a_top1 > valid_accuracies['best']:
valid_accuracies['best'] = valid_a_top1
genotypes['best'] = search_model.genotype()
find_best = True
else:
find_best = False
genotypes[epoch] = search_model.genotype()
logger.log('<<<--->>> The {:}-th epoch : {:}'.format(
epoch_str, genotypes[epoch]))
# save checkpoint
save_path = save_checkpoint(
{
'epoch': epoch + 1,
'args': deepcopy(xargs),
'search_model': search_model.state_dict(),
'w_optimizer': w_optimizer.state_dict(),
'a_optimizer': a_optimizer.state_dict(),
'w_scheduler': w_scheduler.state_dict(),
'genotypes': genotypes,
'valid_accuracies': valid_accuracies
}, model_base_path, logger)
last_info = save_checkpoint(
{
'epoch': epoch + 1,
'args': deepcopy(args),
'last_checkpoint': save_path,
}, logger.path('info'), logger)
if find_best:
logger.log(
'<<<--->>> The {:}-th epoch : find the highest validation accuracy : {:.2f}%.'
.format(epoch_str, valid_a_top1))
copy_checkpoint(model_base_path, model_best_path, logger)
with torch.no_grad():
logger.log('{:}'.format(search_model.show_alphas()))
if api is not None:
logger.log('{:}'.format(api.query_by_arch(genotypes[epoch],
'200')))
# measure elapsed time
epoch_time.update(time.time() - start_time)
start_time = time.time()
logger.log('\n' + '-' * 100)
# check the performance from the architecture dataset
logger.log(
'GDAS : run {:} epochs, cost {:.1f} s, last-geno is {:}.'.format(
total_epoch, search_time.sum, genotypes[total_epoch - 1]))
if api is not None:
logger.log('{:}'.format(
api.query_by_arch(genotypes[total_epoch - 1], '200')))
logger.close()
import xnas.core.optimizer as optim
import xnas.datasets.loader as loader
from xnas.core.config import cfg
from xnas.search_space.cell_based_nasben1shot import INPUT, OUTPUT, CONV1X1, CONV3X3, MAXPOOL3X3, OUTPUT_NODE
import sys
from nas_201_api import NASBench201API as API
from nasbench import api
import ConfigSpace
logger = logging.get_logger(__name__)
nasbench1shot1_path = 'benchmark/nasbench_full.tfrecord'
nasbench201_path = 'benchmark/NAS-Bench-102-v1_0-e61699.pth'
api_nasben201 = API(nasbench201_path, verbose=False)
nasbench = api.NASBench(nasbench1shot1_path)
def setup_env():
"""Sets up environment for training or testing."""
if dist.is_master_proc():
# Ensure that the output dir exists
os.makedirs(cfg.OUT_DIR, exist_ok=True)
# Save the config
config.dump_cfg()
# Setup logging
logging.setup_logging()
# Log the config as both human readable and as a json
logger.info("Config:\n{}".format(cfg))
logger.info(logging.dump_log_data(cfg, "cfg"))
def main():
torch.set_num_threads(3)
if not torch.cuda.is_available():
logging.info('no gpu device available')
sys.exit(1)
np.random.seed(args.seed)
torch.cuda.set_device(args.gpu)
cudnn.benchmark = True
torch.manual_seed(args.seed)
cudnn.enabled = True
torch.cuda.manual_seed(args.seed)
logging.info('gpu device = %d' % args.gpu)
logging.info("args = %s", args)
if args.perturb_alpha == 'none':
perturb_alpha = None
elif args.perturb_alpha == 'pgd_linf':
perturb_alpha = Linf_PGD_alpha
elif args.perturb_alpha == 'random':
perturb_alpha = Random_alpha
api = API('/nfs/data/xiangning/data/NAS-Bench-201-v1_0-e61699.pth')
criterion = nn.CrossEntropyLoss()
criterion = criterion.cuda()
model = Network(C=args.init_channels, N=5, max_nodes=4, num_classes=n_classes, criterion=criterion)
model = model.cuda()
logging.info("param size = %fMB", utils.count_parameters_in_MB(model))
optimizer = torch.optim.SGD(
model.parameters(),
args.learning_rate,
momentum=args.momentum,
weight_decay=args.weight_decay)
if args.dataset == 'cifar10':
train_transform, valid_transform = utils._data_transforms_cifar10(args)
train_data = dset.CIFAR10(root=args.data, train=True, download=True, transform=train_transform)
elif args.dataset == 'cifar100':
train_transform, valid_transform = utils._data_transforms_cifar100(args)
train_data = dset.CIFAR100(root=args.data, train=True, download=True, transform=train_transform)
elif args.dataset == 'svhn':
train_transform, valid_transform = utils._data_transforms_svhn(args)
train_data = dset.SVHN(root=args.data, split='train', download=True, transform=train_transform)
num_train = len(train_data)
indices = list(range(num_train))
split = int(np.floor(args.train_portion * num_train))
if 'debug' in args.save:
split = args.batch_size
num_train = 2 * args.batch_size
train_queue = torch.utils.data.DataLoader(
train_data, batch_size=args.batch_size,
sampler=torch.utils.data.sampler.SubsetRandomSampler(indices[:split]),
pin_memory=True)
valid_queue = torch.utils.data.DataLoader(
train_data, batch_size=args.batch_size,
sampler=torch.utils.data.sampler.SubsetRandomSampler(indices[split:num_train]),
pin_memory=True)
scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(
optimizer, float(args.epochs), eta_min=args.learning_rate_min)
architect = Architect(model, args)
for epoch in range(args.epochs):
scheduler.step()
lr = scheduler.get_lr()[0]
if args.cutout:
# increase the cutout probability linearly throughout search
train_transform.transforms[-1].cutout_prob = args.cutout_prob * epoch / (args.epochs - 1)
logging.info('epoch %d lr %e cutout_prob %e', epoch, lr,
train_transform.transforms[-1].cutout_prob)
else:
logging.info('epoch %d lr %e', epoch, lr)
if args.perturb_alpha:
epsilon_alpha = 0.03 + (args.epsilon_alpha - 0.03) * epoch / args.epochs
logging.info('epoch %d epsilon_alpha %e', epoch, epsilon_alpha)
genotype = model.genotype()
logging.info('genotype = %s', genotype)
model.show_alphas()
# training
train_acc, train_obj = train(train_queue, valid_queue, model, architect, criterion, optimizer, lr,
perturb_alpha, epsilon_alpha)
logging.info('train_acc %f', train_acc)
writer.add_scalar('Acc/train', train_acc, epoch)
writer.add_scalar('Obj/train', train_obj, epoch)
# validation
valid_acc, valid_obj = infer(valid_queue, model, criterion)
logging.info('valid_acc %f', valid_acc)
writer.add_scalar('Acc/valid', valid_acc, epoch)
writer.add_scalar('Obj/valid', valid_obj, epoch)
# nasbench201
result = api.query_by_arch(model.genotype())
logging.info('{:}'.format(result))
cifar10_train, cifar10_test, cifar100_train, cifar100_valid, \
cifar100_test, imagenet16_train, imagenet16_valid, imagenet16_test = distill(result)
writer.add_scalars('nasbench201/cifar10', {'train':cifar10_train,'test':cifar10_test}, epoch)
writer.add_scalars('nasbench201/cifar100', {'train':cifar100_train,'valid':cifar100_valid, 'test':cifar100_test}, epoch)
writer.add_scalars('nasbench201/imagenet16', {'train':imagenet16_train,'valid':imagenet16_valid, 'test':imagenet16_test}, epoch)
utils.save_checkpoint({
'epoch': epoch + 1,
'state_dict': model.state_dict(),
'optimizer': optimizer.state_dict(),
'alpha': model.arch_parameters()
}, False, args.save)
writer.close()
default=None,
help='The path to the NAS-Bench-201 benchmark file.')
args = parser.parse_args()
vis_save_dir = Path(args.save_dir)
vis_save_dir.mkdir(parents=True, exist_ok=True)
meta_file = Path(args.api_path)
assert meta_file.exists(), 'invalid path for api : {:}'.format(meta_file)
#visualize_rank_over_time(str(meta_file), vis_save_dir / 'over-time')
#write_video(vis_save_dir / 'over-time')
#visualize_info(str(meta_file), 'cifar10' , vis_save_dir)
#visualize_info(str(meta_file), 'cifar100', vis_save_dir)
#visualize_info(str(meta_file), 'ImageNet16-120', vis_save_dir)
#visualize_relative_ranking(vis_save_dir)
api = API(args.api_path)
#show_reinforce(api, vis_save_dir, 'cifar10-valid' , 'x-valid', 'REINFORCE-CIFAR-10', (85, 92, 2))
#show_rea (api, vis_save_dir, 'cifar10-valid' , 'x-valid', 'REA-CIFAR-10', (88, 92, 1))
#plot_results_nas_v2(api, ('cifar10-valid' , 'x-valid'), ('cifar10' , 'ori-test'), vis_save_dir, 'nas-com-v2-cifar010.pdf', (85,95, 1))
#plot_results_nas_v2(api, ('cifar100' , 'x-valid'), ('cifar100' , 'x-test' ), vis_save_dir, 'nas-com-v2-cifar100.pdf', (60,75, 3))
#plot_results_nas_v2(api, ('ImageNet16-120', 'x-valid'), ('ImageNet16-120', 'x-test' ), vis_save_dir, 'nas-com-v2-imagenet.pdf', (35,48, 2))
show_nas_sharing_w_v2(api, ('cifar10-valid', 'x-valid'),
('cifar10', 'ori-test'), vis_save_dir, 'BN0',
'BN0-DARTS-CIFAR010.pdf', (0, 100, 10), 50)
show_nas_sharing_w_v2(api, ('cifar100', 'x-valid'), ('cifar100', 'x-test'),
vis_save_dir, 'BN0', 'BN0-DARTS-CIFAR100.pdf',
(0, 100, 10), 50)
show_nas_sharing_w_v2(api, ('ImageNet16-120', 'x-valid'),
('ImageNet16-120', 'x-test'), vis_save_dir, 'BN0',
def main(xargs):
assert torch.cuda.is_available(), 'CUDA is not available.'
torch.backends.cudnn.enabled = True
torch.backends.cudnn.benchmark = False
torch.backends.cudnn.deterministic = True
torch.set_num_threads(xargs.workers)
prepare_seed(xargs.rand_seed)
logger = prepare_logger(args)
train_data, valid_data, xshape, class_num = get_datasets(
xargs.dataset, xargs.data_path, -1)
if xargs.dataset == 'cifar10' or xargs.dataset == 'cifar100':
split_Fpath = 'configs/nas-benchmark/cifar-split.txt'
cifar_split = load_config(split_Fpath, None, None)
train_split, valid_split = cifar_split.train, cifar_split.valid
logger.log('Load split file from {:}'.format(split_Fpath))
elif xargs.dataset.startswith('ImageNet16'):
split_Fpath = 'configs/nas-benchmark/{:}-split.txt'.format(
xargs.dataset)
imagenet16_split = load_config(split_Fpath, None, None)
train_split, valid_split = imagenet16_split.train, imagenet16_split.valid
logger.log('Load split file from {:}'.format(split_Fpath))
else:
raise ValueError('invalid dataset : {:}'.format(xargs.dataset))
config_path = 'configs/nas-benchmark/algos/DARTS.config'
config = load_config(config_path, {
'class_num': class_num,
'xshape': xshape
}, logger)
# To split data
train_data_v2 = deepcopy(train_data)
train_data_v2.transform = valid_data.transform
valid_data = train_data_v2
search_data = SearchDataset(xargs.dataset, train_data, train_split,
valid_split)
# data loader
search_loader = torch.utils.data.DataLoader(search_data,
batch_size=config.batch_size,
shuffle=True,
num_workers=xargs.workers,
pin_memory=True)
valid_loader = torch.utils.data.DataLoader(
valid_data,
batch_size=config.batch_size,
sampler=torch.utils.data.sampler.SubsetRandomSampler(valid_split),
num_workers=xargs.workers,
pin_memory=True)
logger.log(
'||||||| {:10s} ||||||| Search-Loader-Num={:}, Valid-Loader-Num={:}, batch size={:}'
.format(xargs.dataset, len(search_loader), len(valid_loader),
config.batch_size))
logger.log('||||||| {:10s} ||||||| Config={:}'.format(
xargs.dataset, config))
search_space = get_search_spaces('cell', xargs.search_space_name)
model_config = dict2config(
{
'name': 'DARTS-V2',
'C': xargs.channel,
'N': xargs.num_cells,
'max_nodes': xargs.max_nodes,
'num_classes': class_num,
'space': search_space
}, None)
search_model = get_cell_based_tiny_net(model_config)
logger.log('search-model :\n{:}'.format(search_model))
w_optimizer, w_scheduler, criterion = get_optim_scheduler(
search_model.get_weights(), config)
a_optimizer = torch.optim.Adam(search_model.get_alphas(),
lr=xargs.arch_learning_rate,
betas=(0.5, 0.999),
weight_decay=xargs.arch_weight_decay)
logger.log('w-optimizer : {:}'.format(w_optimizer))
logger.log('a-optimizer : {:}'.format(a_optimizer))
logger.log('w-scheduler : {:}'.format(w_scheduler))
logger.log('criterion : {:}'.format(criterion))
flop, param = get_model_infos(search_model, xshape)
#logger.log('{:}'.format(search_model))
logger.log('FLOP = {:.2f} M, Params = {:.2f} MB'.format(flop, param))
if xargs.arch_nas_dataset is None:
api = None
else:
api = API(xargs.arch_nas_dataset)
logger.log('{:} create API = {:} done'.format(time_string(), api))
last_info, model_base_path, model_best_path = logger.path(
'info'), logger.path('model'), logger.path('best')
network, criterion = torch.nn.DataParallel(
search_model).cuda(), criterion.cuda()
logger.close()
def visualize_info(meta_file, dataset, vis_save_dir):
print('{:} start to visualize {:} information'.format(
time_string(), dataset))
cache_file_path = vis_save_dir / '{:}-cache-info.pth'.format(dataset)
if not cache_file_path.exists():
print('Do not find cache file : {:}'.format(cache_file_path))
nas_bench = API(str(meta_file))
params, flops, train_accs, valid_accs, test_accs, otest_accs = [], [], [], [], [], []
for index in range(len(nas_bench)):
info = nas_bench.query_by_index(index, use_12epochs_result=False)
resx = info.get_comput_costs(dataset)
flop, param = resx['flops'], resx['params']
if dataset == 'cifar10':
res = info.get_metrics('cifar10', 'train')
train_acc = res['accuracy']
res = info.get_metrics('cifar10-valid', 'x-valid')
valid_acc = res['accuracy']
res = info.get_metrics('cifar10', 'ori-test')
test_acc = res['accuracy']
res = info.get_metrics('cifar10', 'ori-test')
otest_acc = res['accuracy']
else:
res = info.get_metrics(dataset, 'train')
train_acc = res['accuracy']
res = info.get_metrics(dataset, 'x-valid')
valid_acc = res['accuracy']
res = info.get_metrics(dataset, 'x-test')
test_acc = res['accuracy']
res = info.get_metrics(dataset, 'ori-test')
otest_acc = res['accuracy']
if index == 11472: # resnet
resnet = {
'params': param,
'flops': flop,
'index': 11472,
'train_acc': train_acc,
'valid_acc': valid_acc,
'test_acc': test_acc,
'otest_acc': otest_acc
}
flops.append(flop)
params.append(param)
train_accs.append(train_acc)
valid_accs.append(valid_acc)
test_accs.append(test_acc)
otest_accs.append(otest_acc)
#resnet = {'params': 0.559, 'flops': 78.56, 'index': 11472, 'train_acc': 99.99, 'valid_acc': 90.84, 'test_acc': 93.97}
info = {
'params': params,
'flops': flops,
'train_accs': train_accs,
'valid_accs': valid_accs,
'test_accs': test_accs,
'otest_accs': otest_accs
}
info['resnet'] = resnet
torch.save(info, cache_file_path)
else:
print('Find cache file : {:}'.format(cache_file_path))
info = torch.load(cache_file_path)
params, flops, train_accs, valid_accs, test_accs, otest_accs = info[
'params'], info['flops'], info['train_accs'], info[
'valid_accs'], info['test_accs'], info['otest_accs']
resnet = info['resnet']
print('{:} collect data done.'.format(time_string()))
indexes = list(range(len(params)))
dpi, width, height = 300, 2600, 2600
figsize = width / float(dpi), height / float(dpi)
LabelSize, LegendFontsize = 22, 22
resnet_scale, resnet_alpha = 120, 0.5
fig = plt.figure(figsize=figsize)
ax = fig.add_subplot(111)
plt.xticks(np.arange(0, 1.6, 0.3), fontsize=LegendFontsize)
if dataset == 'cifar10':
plt.ylim(50, 100)
plt.yticks(np.arange(50, 101, 10), fontsize=LegendFontsize)
elif dataset == 'cifar100':
plt.ylim(25, 75)
plt.yticks(np.arange(25, 76, 10), fontsize=LegendFontsize)
else:
plt.ylim(0, 50)
plt.yticks(np.arange(0, 51, 10), fontsize=LegendFontsize)
ax.scatter(params, valid_accs, marker='o', s=0.5, c='tab:blue')
ax.scatter([resnet['params']], [resnet['valid_acc']],
marker='*',
s=resnet_scale,
c='tab:orange',
label='resnet',
alpha=0.4)
plt.grid(zorder=0)
ax.set_axisbelow(True)
plt.legend(loc=4, fontsize=LegendFontsize)
ax.set_xlabel('#parameters (MB)', fontsize=LabelSize)
ax.set_ylabel('the validation accuracy (%)', fontsize=LabelSize)
save_path = (vis_save_dir /
'{:}-param-vs-valid.pdf'.format(dataset)).resolve()
fig.savefig(save_path, dpi=dpi, bbox_inches='tight', format='pdf')
save_path = (vis_save_dir /
'{:}-param-vs-valid.png'.format(dataset)).resolve()
fig.savefig(save_path, dpi=dpi, bbox_inches='tight', format='png')
print('{:} save into {:}'.format(time_string(), save_path))
fig = plt.figure(figsize=figsize)
ax = fig.add_subplot(111)
plt.xticks(np.arange(0, 1.6, 0.3), fontsize=LegendFontsize)
if dataset == 'cifar10':
plt.ylim(50, 100)
plt.yticks(np.arange(50, 101, 10), fontsize=LegendFontsize)
elif dataset == 'cifar100':
plt.ylim(25, 75)
plt.yticks(np.arange(25, 76, 10), fontsize=LegendFontsize)
else:
plt.ylim(0, 50)
plt.yticks(np.arange(0, 51, 10), fontsize=LegendFontsize)
ax.scatter(params, test_accs, marker='o', s=0.5, c='tab:blue')
ax.scatter([resnet['params']], [resnet['test_acc']],
marker='*',
s=resnet_scale,
c='tab:orange',
label='resnet',
alpha=resnet_alpha)
plt.grid()
ax.set_axisbelow(True)
plt.legend(loc=4, fontsize=LegendFontsize)
ax.set_xlabel('#parameters (MB)', fontsize=LabelSize)
ax.set_ylabel('the test accuracy (%)', fontsize=LabelSize)
save_path = (vis_save_dir /
'{:}-param-vs-test.pdf'.format(dataset)).resolve()
fig.savefig(save_path, dpi=dpi, bbox_inches='tight', format='pdf')
save_path = (vis_save_dir /
'{:}-param-vs-test.png'.format(dataset)).resolve()
fig.savefig(save_path, dpi=dpi, bbox_inches='tight', format='png')
print('{:} save into {:}'.format(time_string(), save_path))
fig = plt.figure(figsize=figsize)
ax = fig.add_subplot(111)
plt.xticks(np.arange(0, 1.6, 0.3), fontsize=LegendFontsize)
if dataset == 'cifar10':
plt.ylim(50, 100)
plt.yticks(np.arange(50, 101, 10), fontsize=LegendFontsize)
elif dataset == 'cifar100':
plt.ylim(20, 100)
plt.yticks(np.arange(20, 101, 10), fontsize=LegendFontsize)
else:
plt.ylim(25, 76)
plt.yticks(np.arange(25, 76, 10), fontsize=LegendFontsize)
ax.scatter(params, train_accs, marker='o', s=0.5, c='tab:blue')
ax.scatter([resnet['params']], [resnet['train_acc']],
marker='*',
s=resnet_scale,
c='tab:orange',
label='resnet',
alpha=resnet_alpha)
plt.grid()
ax.set_axisbelow(True)
plt.legend(loc=4, fontsize=LegendFontsize)
ax.set_xlabel('#parameters (MB)', fontsize=LabelSize)
ax.set_ylabel('the trarining accuracy (%)', fontsize=LabelSize)
save_path = (vis_save_dir /
'{:}-param-vs-train.pdf'.format(dataset)).resolve()
fig.savefig(save_path, dpi=dpi, bbox_inches='tight', format='pdf')
save_path = (vis_save_dir /
'{:}-param-vs-train.png'.format(dataset)).resolve()
fig.savefig(save_path, dpi=dpi, bbox_inches='tight', format='png')
print('{:} save into {:}'.format(time_string(), save_path))
fig = plt.figure(figsize=figsize)
ax = fig.add_subplot(111)
plt.xlim(0, max(indexes))
plt.xticks(np.arange(min(indexes), max(indexes),
max(indexes) // 5),
fontsize=LegendFontsize)
if dataset == 'cifar10':
plt.ylim(50, 100)
plt.yticks(np.arange(50, 101, 10), fontsize=LegendFontsize)
elif dataset == 'cifar100':
plt.ylim(25, 75)
plt.yticks(np.arange(25, 76, 10), fontsize=LegendFontsize)
else:
plt.ylim(0, 50)
plt.yticks(np.arange(0, 51, 10), fontsize=LegendFontsize)
ax.scatter(indexes, test_accs, marker='o', s=0.5, c='tab:blue')
ax.scatter([resnet['index']], [resnet['test_acc']],
marker='*',
s=resnet_scale,
c='tab:orange',
label='resnet',
alpha=resnet_alpha)
plt.grid()
ax.set_axisbelow(True)
plt.legend(loc=4, fontsize=LegendFontsize)
ax.set_xlabel('architecture ID', fontsize=LabelSize)
ax.set_ylabel('the test accuracy (%)', fontsize=LabelSize)
save_path = (vis_save_dir /
'{:}-test-over-ID.pdf'.format(dataset)).resolve()
fig.savefig(save_path, dpi=dpi, bbox_inches='tight', format='pdf')
save_path = (vis_save_dir /
'{:}-test-over-ID.png'.format(dataset)).resolve()
fig.savefig(save_path, dpi=dpi, bbox_inches='tight', format='png')
print('{:} save into {:}'.format(time_string(), save_path))
plt.close('all')
def main():
api = API(RAWPATH)
r, e = create_record(outpath=OUTPATH, api=api)
shuffle_data_n_times_and_store(OUTPATH, len(api))
def visualize_rank_over_time(meta_file, vis_save_dir):
print('\n' + '-' * 150)
vis_save_dir.mkdir(parents=True, exist_ok=True)
print('{:} start to visualize rank-over-time into {:}'.format(
time_string(), vis_save_dir))
cache_file_path = vis_save_dir / 'rank-over-time-cache-info.pth'
if not cache_file_path.exists():
print('Do not find cache file : {:}'.format(cache_file_path))
nas_bench = API(str(meta_file))
print('{:} load nas_bench done'.format(time_string()))
params, flops, train_accs, valid_accs, test_accs, otest_accs = [], [], defaultdict(
list), defaultdict(list), defaultdict(list), defaultdict(list)
#for iepoch in range(200): for index in range( len(nas_bench) ):
for index in tqdm(range(len(nas_bench))):
info = nas_bench.query_by_index(index, use_12epochs_result=False)
for iepoch in range(200):
res = info.get_metrics('cifar10', 'train', iepoch)
train_acc = res['accuracy']
res = info.get_metrics('cifar10-valid', 'x-valid', iepoch)
valid_acc = res['accuracy']
res = info.get_metrics('cifar10', 'ori-test', iepoch)
test_acc = res['accuracy']
res = info.get_metrics('cifar10', 'ori-test', iepoch)
otest_acc = res['accuracy']
train_accs[iepoch].append(train_acc)
valid_accs[iepoch].append(valid_acc)
test_accs[iepoch].append(test_acc)
otest_accs[iepoch].append(otest_acc)
if iepoch == 0:
res = info.get_comput_costs('cifar10')
flop, param = res['flops'], res['params']
flops.append(flop)
params.append(param)
info = {
'params': params,
'flops': flops,
'train_accs': train_accs,
'valid_accs': valid_accs,
'test_accs': test_accs,
'otest_accs': otest_accs
}
torch.save(info, cache_file_path)
else:
print('Find cache file : {:}'.format(cache_file_path))
info = torch.load(cache_file_path)
params, flops, train_accs, valid_accs, test_accs, otest_accs = info[
'params'], info['flops'], info['train_accs'], info[
'valid_accs'], info['test_accs'], info['otest_accs']
print('{:} collect data done.'.format(time_string()))
#selected_epochs = [0, 100, 150, 180, 190, 191, 192, 193, 194, 195, 196, 197, 198, 199]
selected_epochs = list(range(200))
x_xtests = test_accs[199]
indexes = list(range(len(x_xtests)))
ord_idxs = sorted(indexes, key=lambda i: x_xtests[i])
for sepoch in selected_epochs:
x_valids = valid_accs[sepoch]
valid_ord_idxs = sorted(indexes, key=lambda i: x_valids[i])
valid_ord_lbls = []
for idx in ord_idxs:
valid_ord_lbls.append(valid_ord_idxs.index(idx))
# labeled data
dpi, width, height = 300, 2600, 2600
figsize = width / float(dpi), height / float(dpi)
LabelSize, LegendFontsize = 18, 18
fig = plt.figure(figsize=figsize)
ax = fig.add_subplot(111)
plt.xlim(min(indexes), max(indexes))
plt.ylim(min(indexes), max(indexes))
plt.yticks(np.arange(min(indexes), max(indexes),
max(indexes) // 6),
fontsize=LegendFontsize,
rotation='vertical')
plt.xticks(np.arange(min(indexes), max(indexes),
max(indexes) // 6),
fontsize=LegendFontsize)
ax.scatter(indexes,
valid_ord_lbls,
marker='^',
s=0.5,
c='tab:green',
alpha=0.8)
ax.scatter(indexes,
indexes,
marker='o',
s=0.5,
c='tab:blue',
alpha=0.8)
ax.scatter([-1], [-1],
marker='^',
s=100,
c='tab:green',
label='CIFAR-10 validation')
ax.scatter([-1], [-1],
marker='o',
s=100,
c='tab:blue',
label='CIFAR-10 test')
plt.grid(zorder=0)
ax.set_axisbelow(True)
plt.legend(loc='upper left', fontsize=LegendFontsize)
ax.set_xlabel('architecture ranking in the final test accuracy',
fontsize=LabelSize)
ax.set_ylabel('architecture ranking in the validation set',
fontsize=LabelSize)
save_path = (vis_save_dir / 'time-{:03d}.pdf'.format(sepoch)).resolve()
fig.savefig(save_path, dpi=dpi, bbox_inches='tight', format='pdf')
save_path = (vis_save_dir / 'time-{:03d}.png'.format(sepoch)).resolve()
fig.savefig(save_path, dpi=dpi, bbox_inches='tight', format='png')
print('{:} save into {:}'.format(time_string(), save_path))
plt.close('all')
default="./output/search-cell-nas-bench-201/visuals",
help="The base-name of folder to save checkpoints and log.",
)
parser.add_argument(
"--api_path",
type=str,
default=None,
help="The path to the NAS-Bench-201 benchmark file.",
)
args = parser.parse_args()
vis_save_dir = Path(args.save_dir)
vis_save_dir.mkdir(parents=True, exist_ok=True)
meta_file = Path(args.api_path)
assert meta_file.exists(), "invalid path for api : {:}".format(meta_file)
# check_unique_arch(meta_file)
api = API(str(meta_file))
# for iepoch in [11, 25, 50, 100, 150, 175, 200]:
# check_cor_for_bandit(api, 6, iepoch)
# check_cor_for_bandit(api, 12, iepoch)
check_cor_for_bandit_v2(api, 6, True, True)
check_cor_for_bandit_v2(api, 12, True, True)
check_cor_for_bandit_v2(api, 12, False, True)
check_cor_for_bandit_v2(api, 24, False, True)
check_cor_for_bandit_v2(api, 100, False, True)
check_cor_for_bandit_v2(api, 150, False, True)
check_cor_for_bandit_v2(api, 175, False, True)
check_cor_for_bandit_v2(api, 200, False, True)
print("----")
def main(xargs):
assert torch.cuda.is_available(), "CUDA is not available."
torch.backends.cudnn.enabled = True
torch.backends.cudnn.benchmark = False
torch.backends.cudnn.deterministic = True
torch.set_num_threads(xargs.workers)
prepare_seed(xargs.rand_seed)
logger = prepare_logger(args)
train_data, valid_data, xshape, class_num = get_datasets(
xargs.dataset, xargs.data_path, -1)
config = load_config(xargs.config_path, {
"class_num": class_num,
"xshape": xshape
}, logger)
search_loader, _, valid_loader = get_nas_search_loaders(
train_data,
valid_data,
xargs.dataset,
"configs/nas-benchmark/",
(config.batch_size, config.test_batch_size),
xargs.workers,
)
logger.log(
"||||||| {:10s} ||||||| Search-Loader-Num={:}, Valid-Loader-Num={:}, batch size={:}"
.format(xargs.dataset, len(search_loader), len(valid_loader),
config.batch_size))
logger.log("||||||| {:10s} ||||||| Config={:}".format(
xargs.dataset, config))
search_space = get_search_spaces("cell", xargs.search_space_name)
model_config = dict2config(
{
"name": "RANDOM",
"C": xargs.channel,
"N": xargs.num_cells,
"max_nodes": xargs.max_nodes,
"num_classes": class_num,
"space": search_space,
"affine": False,
"track_running_stats": bool(xargs.track_running_stats),
},
None,
)
search_model = get_cell_based_tiny_net(model_config)
w_optimizer, w_scheduler, criterion = get_optim_scheduler(
search_model.parameters(), config)
logger.log("w-optimizer : {:}".format(w_optimizer))
logger.log("w-scheduler : {:}".format(w_scheduler))
logger.log("criterion : {:}".format(criterion))
if xargs.arch_nas_dataset is None:
api = None
else:
api = API(xargs.arch_nas_dataset)
logger.log("{:} create API = {:} done".format(time_string(), api))
last_info, model_base_path, model_best_path = (
logger.path("info"),
logger.path("model"),
logger.path("best"),
)
network, criterion = torch.nn.DataParallel(
search_model).cuda(), criterion.cuda()
if last_info.exists(): # automatically resume from previous checkpoint
logger.log("=> loading checkpoint of the last-info '{:}' start".format(
last_info))
last_info = torch.load(last_info)
start_epoch = last_info["epoch"]
checkpoint = torch.load(last_info["last_checkpoint"])
genotypes = checkpoint["genotypes"]
valid_accuracies = checkpoint["valid_accuracies"]
search_model.load_state_dict(checkpoint["search_model"])
w_scheduler.load_state_dict(checkpoint["w_scheduler"])
w_optimizer.load_state_dict(checkpoint["w_optimizer"])
logger.log(
"=> loading checkpoint of the last-info '{:}' start with {:}-th epoch."
.format(last_info, start_epoch))
else:
logger.log("=> do not find the last-info file : {:}".format(last_info))
start_epoch, valid_accuracies, genotypes = 0, {"best": -1}, {}
# start training
start_time, search_time, epoch_time, total_epoch = (
time.time(),
AverageMeter(),
AverageMeter(),
config.epochs + config.warmup,
)
for epoch in range(start_epoch, total_epoch):
w_scheduler.update(epoch, 0.0)
need_time = "Time Left: {:}".format(
convert_secs2time(epoch_time.val * (total_epoch - epoch), True))
epoch_str = "{:03d}-{:03d}".format(epoch, total_epoch)
logger.log("\n[Search the {:}-th epoch] {:}, LR={:}".format(
epoch_str, need_time, min(w_scheduler.get_lr())))
# selected_arch = search_find_best(valid_loader, network, criterion, xargs.select_num)
search_w_loss, search_w_top1, search_w_top5 = search_func(
search_loader,
network,
criterion,
w_scheduler,
w_optimizer,
epoch_str,
xargs.print_freq,
logger,
)
search_time.update(time.time() - start_time)
logger.log(
"[{:}] searching : loss={:.2f}, accuracy@1={:.2f}%, accuracy@5={:.2f}%, time-cost={:.1f} s"
.format(epoch_str, search_w_loss, search_w_top1, search_w_top5,
search_time.sum))
valid_a_loss, valid_a_top1, valid_a_top5 = valid_func(
valid_loader, network, criterion)
logger.log(
"[{:}] evaluate : loss={:.2f}, accuracy@1={:.2f}%, accuracy@5={:.2f}%"
.format(epoch_str, valid_a_loss, valid_a_top1, valid_a_top5))
cur_arch, cur_valid_acc = search_find_best(valid_loader, network,
xargs.select_num)
logger.log("[{:}] find-the-best : {:}, accuracy@1={:.2f}%".format(
epoch_str, cur_arch, cur_valid_acc))
genotypes[epoch] = cur_arch
# check the best accuracy
valid_accuracies[epoch] = valid_a_top1
if valid_a_top1 > valid_accuracies["best"]:
valid_accuracies["best"] = valid_a_top1
find_best = True
else:
find_best = False
# save checkpoint
save_path = save_checkpoint(
{
"epoch": epoch + 1,
"args": deepcopy(xargs),
"search_model": search_model.state_dict(),
"w_optimizer": w_optimizer.state_dict(),
"w_scheduler": w_scheduler.state_dict(),
"genotypes": genotypes,
"valid_accuracies": valid_accuracies,
},
model_base_path,
logger,
)
last_info = save_checkpoint(
{
"epoch": epoch + 1,
"args": deepcopy(args),
"last_checkpoint": save_path,
},
logger.path("info"),
logger,
)
if find_best:
logger.log(
"<<<--->>> The {:}-th epoch : find the highest validation accuracy : {:.2f}%."
.format(epoch_str, valid_a_top1))
copy_checkpoint(model_base_path, model_best_path, logger)
if api is not None:
logger.log("{:}".format(api.query_by_arch(genotypes[epoch],
"200")))
# measure elapsed time
epoch_time.update(time.time() - start_time)
start_time = time.time()
logger.log("\n" + "-" * 200)
logger.log("Pre-searching costs {:.1f} s".format(search_time.sum))
start_time = time.time()
best_arch, best_acc = search_find_best(valid_loader, network,
xargs.select_num)
search_time.update(time.time() - start_time)
logger.log(
"RANDOM-NAS finds the best one : {:} with accuracy={:.2f}%, with {:.1f} s."
.format(best_arch, best_acc, search_time.sum))
if api is not None:
logger.log("{:}".format(api.query_by_arch(best_arch, "200")))
logger.close()
def main(xargs):
assert torch.cuda.is_available(), 'CUDA is not available.'
torch.backends.cudnn.enabled = True
torch.backends.cudnn.benchmark = False
torch.backends.cudnn.deterministic = True
torch.set_num_threads(xargs.workers)
prepare_seed(xargs.rand_seed)
logger = prepare_logger(args)
train_data, valid_data, xshape, class_num = get_datasets(
xargs.dataset, xargs.data_path, -1)
config = load_config(xargs.config_path, {
'class_num': class_num,
'xshape': xshape
}, logger)
search_loader, _, valid_loader = get_nas_search_loaders(train_data, valid_data, xargs.dataset, 'configs/nas-benchmark/', \
(config.batch_size, config.test_batch_size), xargs.workers)
logger.log(
'||||||| {:10s} ||||||| Search-Loader-Num={:}, Valid-Loader-Num={:}, batch size={:}'
.format(xargs.dataset, len(search_loader), len(valid_loader),
config.batch_size))
logger.log('||||||| {:10s} ||||||| Config={:}'.format(
xargs.dataset, config))
search_space = get_search_spaces('cell', xargs.search_space_name)
model_config = dict2config(
{
'name': 'RANDOM',
'C': xargs.channel,
'N': xargs.num_cells,
'max_nodes': xargs.max_nodes,
'num_classes': class_num,
'space': search_space,
'affine': False,
'track_running_stats': bool(xargs.track_running_stats)
}, None)
search_model = get_cell_based_tiny_net(model_config)
w_optimizer, w_scheduler, criterion = get_optim_scheduler(
search_model.parameters(), config)
logger.log('w-optimizer : {:}'.format(w_optimizer))
logger.log('w-scheduler : {:}'.format(w_scheduler))
logger.log('criterion : {:}'.format(criterion))
if xargs.arch_nas_dataset is None: api = None
else: api = API(xargs.arch_nas_dataset)
logger.log('{:} create API = {:} done'.format(time_string(), api))
last_info, model_base_path, model_best_path = logger.path(
'info'), logger.path('model'), logger.path('best')
network, criterion = torch.nn.DataParallel(
search_model).cuda(), criterion.cuda()
if last_info.exists(): # automatically resume from previous checkpoint
logger.log("=> loading checkpoint of the last-info '{:}' start".format(
last_info))
last_info = torch.load(last_info)
start_epoch = last_info['epoch']
checkpoint = torch.load(last_info['last_checkpoint'])
genotypes = checkpoint['genotypes']
valid_accuracies = checkpoint['valid_accuracies']
search_model.load_state_dict(checkpoint['search_model'])
w_scheduler.load_state_dict(checkpoint['w_scheduler'])
w_optimizer.load_state_dict(checkpoint['w_optimizer'])
logger.log(
"=> loading checkpoint of the last-info '{:}' start with {:}-th epoch."
.format(last_info, start_epoch))
else:
logger.log("=> do not find the last-info file : {:}".format(last_info))
start_epoch, valid_accuracies, genotypes = 0, {'best': -1}, {}
# start training
start_time, search_time, epoch_time, total_epoch = time.time(
), AverageMeter(), AverageMeter(), config.epochs + config.warmup
for epoch in range(start_epoch, total_epoch):
w_scheduler.update(epoch, 0.0)
need_time = 'Time Left: {:}'.format(
convert_secs2time(epoch_time.val * (total_epoch - epoch), True))
epoch_str = '{:03d}-{:03d}'.format(epoch, total_epoch)
logger.log('\n[Search the {:}-th epoch] {:}, LR={:}'.format(
epoch_str, need_time, min(w_scheduler.get_lr())))
# selected_arch = search_find_best(valid_loader, network, criterion, xargs.select_num)
search_w_loss, search_w_top1, search_w_top5 = search_func(
search_loader, network, criterion, w_scheduler, w_optimizer,
epoch_str, xargs.print_freq, logger)
search_time.update(time.time() - start_time)
logger.log(
'[{:}] searching : loss={:.2f}, accuracy@1={:.2f}%, accuracy@5={:.2f}%, time-cost={:.1f} s'
.format(epoch_str, search_w_loss, search_w_top1, search_w_top5,
search_time.sum))
valid_a_loss, valid_a_top1, valid_a_top5 = valid_func(
valid_loader, network, criterion)
logger.log(
'[{:}] evaluate : loss={:.2f}, accuracy@1={:.2f}%, accuracy@5={:.2f}%'
.format(epoch_str, valid_a_loss, valid_a_top1, valid_a_top5))
cur_arch, cur_valid_acc = search_find_best(valid_loader, network,
xargs.select_num)
logger.log('[{:}] find-the-best : {:}, accuracy@1={:.2f}%'.format(
epoch_str, cur_arch, cur_valid_acc))
genotypes[epoch] = cur_arch
# check the best accuracy
valid_accuracies[epoch] = valid_a_top1
if valid_a_top1 > valid_accuracies['best']:
valid_accuracies['best'] = valid_a_top1
find_best = True
else:
find_best = False
# save checkpoint
save_path = save_checkpoint(
{
'epoch': epoch + 1,
'args': deepcopy(xargs),
'search_model': search_model.state_dict(),
'w_optimizer': w_optimizer.state_dict(),
'w_scheduler': w_scheduler.state_dict(),
'genotypes': genotypes,
'valid_accuracies': valid_accuracies
}, model_base_path, logger)
last_info = save_checkpoint(
{
'epoch': epoch + 1,
'args': deepcopy(args),
'last_checkpoint': save_path,
}, logger.path('info'), logger)
if find_best:
logger.log(
'<<<--->>> The {:}-th epoch : find the highest validation accuracy : {:.2f}%.'
.format(epoch_str, valid_a_top1))
copy_checkpoint(model_base_path, model_best_path, logger)
if api is not None:
logger.log('{:}'.format(api.query_by_arch(genotypes[epoch],
'200')))
# measure elapsed time
epoch_time.update(time.time() - start_time)
start_time = time.time()
logger.log('\n' + '-' * 200)
logger.log('Pre-searching costs {:.1f} s'.format(search_time.sum))
start_time = time.time()
best_arch, best_acc = search_find_best(valid_loader, network,
xargs.select_num)
search_time.update(time.time() - start_time)
logger.log(
'RANDOM-NAS finds the best one : {:} with accuracy={:.2f}%, with {:.1f} s.'
.format(best_arch, best_acc, search_time.sum))
if api is not None:
logger.log('{:}'.format(api.query_by_arch(best_arch, '200')))
logger.close()
def main(xargs):
assert torch.cuda.is_available(), "CUDA is not available."
torch.backends.cudnn.enabled = True
torch.backends.cudnn.benchmark = False
torch.backends.cudnn.deterministic = True
torch.set_num_threads(xargs.workers)
prepare_seed(xargs.rand_seed)
logger = prepare_logger(args)
train_data, test_data, xshape, class_num = get_datasets(
xargs.dataset, xargs.data_path, -1)
logger.log("use config from : {:}".format(xargs.config_path))
config = load_config(xargs.config_path, {
"class_num": class_num,
"xshape": xshape
}, logger)
_, train_loader, valid_loader = get_nas_search_loaders(
train_data,
test_data,
xargs.dataset,
"configs/nas-benchmark/",
config.batch_size,
xargs.workers,
)
# since ENAS will train the controller on valid-loader, we need to use train transformation for valid-loader
valid_loader.dataset.transform = deepcopy(train_loader.dataset.transform)
if hasattr(valid_loader.dataset, "transforms"):
valid_loader.dataset.transforms = deepcopy(
train_loader.dataset.transforms)
# data loader
logger.log(
"||||||| {:10s} ||||||| Train-Loader-Num={:}, Valid-Loader-Num={:}, batch size={:}"
.format(xargs.dataset, len(train_loader), len(valid_loader),
config.batch_size))
logger.log("||||||| {:10s} ||||||| Config={:}".format(
xargs.dataset, config))
search_space = get_search_spaces("cell", xargs.search_space_name)
model_config = dict2config(
{
"name": "ENAS",
"C": xargs.channel,
"N": xargs.num_cells,
"max_nodes": xargs.max_nodes,
"num_classes": class_num,
"space": search_space,
"affine": False,
"track_running_stats": bool(xargs.track_running_stats),
},
None,
)
shared_cnn = get_cell_based_tiny_net(model_config)
controller = shared_cnn.create_controller()
w_optimizer, w_scheduler, criterion = get_optim_scheduler(
shared_cnn.parameters(), config)
a_optimizer = torch.optim.Adam(
controller.parameters(),
lr=config.controller_lr,
betas=config.controller_betas,
eps=config.controller_eps,
)
logger.log("w-optimizer : {:}".format(w_optimizer))
logger.log("a-optimizer : {:}".format(a_optimizer))
logger.log("w-scheduler : {:}".format(w_scheduler))
logger.log("criterion : {:}".format(criterion))
# flop, param = get_model_infos(shared_cnn, xshape)
# logger.log('{:}'.format(shared_cnn))
# logger.log('FLOP = {:.2f} M, Params = {:.2f} MB'.format(flop, param))
logger.log("search-space : {:}".format(search_space))
if xargs.arch_nas_dataset is None:
api = None
else:
api = API(xargs.arch_nas_dataset)
logger.log("{:} create API = {:} done".format(time_string(), api))
shared_cnn, controller, criterion = (
torch.nn.DataParallel(shared_cnn).cuda(),
controller.cuda(),
criterion.cuda(),
)
last_info, model_base_path, model_best_path = (
logger.path("info"),
logger.path("model"),
logger.path("best"),
)
if last_info.exists(): # automatically resume from previous checkpoint
logger.log("=> loading checkpoint of the last-info '{:}' start".format(
last_info))
last_info = torch.load(last_info)
start_epoch = last_info["epoch"]
checkpoint = torch.load(last_info["last_checkpoint"])
genotypes = checkpoint["genotypes"]
baseline = checkpoint["baseline"]
valid_accuracies = checkpoint["valid_accuracies"]
shared_cnn.load_state_dict(checkpoint["shared_cnn"])
controller.load_state_dict(checkpoint["controller"])
w_scheduler.load_state_dict(checkpoint["w_scheduler"])
w_optimizer.load_state_dict(checkpoint["w_optimizer"])
a_optimizer.load_state_dict(checkpoint["a_optimizer"])
logger.log(
"=> loading checkpoint of the last-info '{:}' start with {:}-th epoch."
.format(last_info, start_epoch))
else:
logger.log("=> do not find the last-info file : {:}".format(last_info))
start_epoch, valid_accuracies, genotypes, baseline = 0, {
"best": -1
}, {}, None
# start training
start_time, search_time, epoch_time, total_epoch = (
time.time(),
AverageMeter(),
AverageMeter(),
config.epochs + config.warmup,
)
for epoch in range(start_epoch, total_epoch):
w_scheduler.update(epoch, 0.0)
need_time = "Time Left: {:}".format(
convert_secs2time(epoch_time.val * (total_epoch - epoch), True))
epoch_str = "{:03d}-{:03d}".format(epoch, total_epoch)
logger.log(
"\n[Search the {:}-th epoch] {:}, LR={:}, baseline={:}".format(
epoch_str, need_time, min(w_scheduler.get_lr()), baseline))
cnn_loss, cnn_top1, cnn_top5 = train_shared_cnn(
train_loader,
shared_cnn,
controller,
criterion,
w_scheduler,
w_optimizer,
epoch_str,
xargs.print_freq,
logger,
)
logger.log(
"[{:}] shared-cnn : loss={:.2f}, accuracy@1={:.2f}%, accuracy@5={:.2f}%"
.format(epoch_str, cnn_loss, cnn_top1, cnn_top5))
ctl_loss, ctl_acc, ctl_baseline, ctl_reward, baseline = train_controller(
valid_loader,
shared_cnn,
controller,
criterion,
a_optimizer,
dict2config(
{
"baseline": baseline,
"ctl_train_steps": xargs.controller_train_steps,
"ctl_num_aggre": xargs.controller_num_aggregate,
"ctl_entropy_w": xargs.controller_entropy_weight,
"ctl_bl_dec": xargs.controller_bl_dec,
},
None,
),
epoch_str,
xargs.print_freq,
logger,
)
search_time.update(time.time() - start_time)
logger.log(
"[{:}] controller : loss={:.2f}, accuracy={:.2f}%, baseline={:.2f}, reward={:.2f}, current-baseline={:.4f}, time-cost={:.1f} s"
.format(
epoch_str,
ctl_loss,
ctl_acc,
ctl_baseline,
ctl_reward,
baseline,
search_time.sum,
))
best_arch, _ = get_best_arch(controller, shared_cnn, valid_loader)
shared_cnn.module.update_arch(best_arch)
_, best_valid_acc, _ = valid_func(valid_loader, shared_cnn, criterion)
genotypes[epoch] = best_arch
# check the best accuracy
valid_accuracies[epoch] = best_valid_acc
if best_valid_acc > valid_accuracies["best"]:
valid_accuracies["best"] = best_valid_acc
genotypes["best"] = best_arch
find_best = True
else:
find_best = False
logger.log("<<<--->>> The {:}-th epoch : {:}".format(
epoch_str, genotypes[epoch]))
# save checkpoint
save_path = save_checkpoint(
{
"epoch": epoch + 1,
"args": deepcopy(xargs),
"baseline": baseline,
"shared_cnn": shared_cnn.state_dict(),
"controller": controller.state_dict(),
"w_optimizer": w_optimizer.state_dict(),
"a_optimizer": a_optimizer.state_dict(),
"w_scheduler": w_scheduler.state_dict(),
"genotypes": genotypes,
"valid_accuracies": valid_accuracies,
},
model_base_path,
logger,
)
last_info = save_checkpoint(
{
"epoch": epoch + 1,
"args": deepcopy(args),
"last_checkpoint": save_path,
},
logger.path("info"),
logger,
)
if find_best:
logger.log(
"<<<--->>> The {:}-th epoch : find the highest validation accuracy : {:.2f}%."
.format(epoch_str, best_valid_acc))
copy_checkpoint(model_base_path, model_best_path, logger)
if api is not None:
logger.log("{:}".format(api.query_by_arch(genotypes[epoch],
"200")))
# measure elapsed time
epoch_time.update(time.time() - start_time)
start_time = time.time()
logger.log("\n" + "-" * 100)
logger.log("During searching, the best architecture is {:}".format(
genotypes["best"]))
logger.log("Its accuracy is {:.2f}%".format(valid_accuracies["best"]))
logger.log("Randomly select {:} architectures and select the best.".format(
xargs.controller_num_samples))
start_time = time.time()
final_arch, _ = get_best_arch(controller, shared_cnn, valid_loader,
xargs.controller_num_samples)
search_time.update(time.time() - start_time)
shared_cnn.module.update_arch(final_arch)
final_loss, final_top1, final_top5 = valid_func(valid_loader, shared_cnn,
criterion)
logger.log("The Selected Final Architecture : {:}".format(final_arch))
logger.log("Loss={:.3f}, Accuracy@1={:.2f}%, Accuracy@5={:.2f}%".format(
final_loss, final_top1, final_top5))
logger.log(
"ENAS : run {:} epochs, cost {:.1f} s, last-geno is {:}.".format(
total_epoch, search_time.sum, final_arch))
if api is not None:
logger.log("{:}".format(api.query_by_arch(final_arch)))
logger.close()
def main(xargs):
assert torch.cuda.is_available(), "CUDA is not available."
torch.backends.cudnn.enabled = True
torch.backends.cudnn.benchmark = False
torch.backends.cudnn.deterministic = True
torch.set_num_threads(xargs.workers)
prepare_seed(xargs.rand_seed)
logger = prepare_logger(args)
train_data, valid_data, xshape, class_num = get_datasets(
xargs.dataset, xargs.data_path, -1
)
# config_path = 'configs/nas-benchmark/algos/GDAS.config'
config = load_config(
xargs.config_path, {"class_num": class_num, "xshape": xshape}, logger
)
search_loader, _, valid_loader = get_nas_search_loaders(
train_data,
valid_data,
xargs.dataset,
"configs/nas-benchmark/",
config.batch_size,
xargs.workers,
)
logger.log(
"||||||| {:10s} ||||||| Search-Loader-Num={:}, batch size={:}".format(
xargs.dataset, len(search_loader), config.batch_size
)
)
logger.log("||||||| {:10s} ||||||| Config={:}".format(xargs.dataset, config))
search_space = get_search_spaces("cell", xargs.search_space_name)
if xargs.model_config is None:
model_config = dict2config(
{
"name": "GDAS",
"C": xargs.channel,
"N": xargs.num_cells,
"max_nodes": xargs.max_nodes,
"num_classes": class_num,
"space": search_space,
"affine": False,
"track_running_stats": bool(xargs.track_running_stats),
},
None,
)
else:
model_config = load_config(
xargs.model_config,
{
"num_classes": class_num,
"space": search_space,
"affine": False,
"track_running_stats": bool(xargs.track_running_stats),
},
None,
)
search_model = get_cell_based_tiny_net(model_config)
logger.log("search-model :\n{:}".format(search_model))
logger.log("model-config : {:}".format(model_config))
w_optimizer, w_scheduler, criterion = get_optim_scheduler(
search_model.get_weights(), config
)
a_optimizer = torch.optim.Adam(
search_model.get_alphas(),
lr=xargs.arch_learning_rate,
betas=(0.5, 0.999),
weight_decay=xargs.arch_weight_decay,
)
logger.log("w-optimizer : {:}".format(w_optimizer))
logger.log("a-optimizer : {:}".format(a_optimizer))
logger.log("w-scheduler : {:}".format(w_scheduler))
logger.log("criterion : {:}".format(criterion))
flop, param = get_model_infos(search_model, xshape)
logger.log("FLOP = {:.2f} M, Params = {:.2f} MB".format(flop, param))
logger.log("search-space [{:} ops] : {:}".format(len(search_space), search_space))
if xargs.arch_nas_dataset is None:
api = None
else:
api = API(xargs.arch_nas_dataset)
logger.log("{:} create API = {:} done".format(time_string(), api))
last_info, model_base_path, model_best_path = (
logger.path("info"),
logger.path("model"),
logger.path("best"),
)
network, criterion = torch.nn.DataParallel(search_model).cuda(), criterion.cuda()
if last_info.exists(): # automatically resume from previous checkpoint
logger.log(
"=> loading checkpoint of the last-info '{:}' start".format(last_info)
)
last_info = torch.load(last_info)
start_epoch = last_info["epoch"]
checkpoint = torch.load(last_info["last_checkpoint"])
genotypes = checkpoint["genotypes"]
valid_accuracies = checkpoint["valid_accuracies"]
search_model.load_state_dict(checkpoint["search_model"])
w_scheduler.load_state_dict(checkpoint["w_scheduler"])
w_optimizer.load_state_dict(checkpoint["w_optimizer"])
a_optimizer.load_state_dict(checkpoint["a_optimizer"])
logger.log(
"=> loading checkpoint of the last-info '{:}' start with {:}-th epoch.".format(
last_info, start_epoch
)
)
else:
logger.log("=> do not find the last-info file : {:}".format(last_info))
start_epoch, valid_accuracies, genotypes = (
0,
{"best": -1},
{-1: search_model.genotype()},
)
# start training
start_time, search_time, epoch_time, total_epoch = (
time.time(),
AverageMeter(),
AverageMeter(),
config.epochs + config.warmup,
)
for epoch in range(start_epoch, total_epoch):
w_scheduler.update(epoch, 0.0)
need_time = "Time Left: {:}".format(
convert_secs2time(epoch_time.val * (total_epoch - epoch), True)
)
epoch_str = "{:03d}-{:03d}".format(epoch, total_epoch)
search_model.set_tau(
xargs.tau_max - (xargs.tau_max - xargs.tau_min) * epoch / (total_epoch - 1)
)
logger.log(
"\n[Search the {:}-th epoch] {:}, tau={:}, LR={:}".format(
epoch_str, need_time, search_model.get_tau(), min(w_scheduler.get_lr())
)
)
(
search_w_loss,
search_w_top1,
search_w_top5,
valid_a_loss,
valid_a_top1,
valid_a_top5,
) = search_func(
search_loader,
network,
criterion,
w_scheduler,
w_optimizer,
a_optimizer,
epoch_str,
xargs.print_freq,
logger,
)
search_time.update(time.time() - start_time)
logger.log(
"[{:}] searching : loss={:.2f}, accuracy@1={:.2f}%, accuracy@5={:.2f}%, time-cost={:.1f} s".format(
epoch_str, search_w_loss, search_w_top1, search_w_top5, search_time.sum
)
)
logger.log(
"[{:}] evaluate : loss={:.2f}, accuracy@1={:.2f}%, accuracy@5={:.2f}%".format(
epoch_str, valid_a_loss, valid_a_top1, valid_a_top5
)
)
# check the best accuracy
valid_accuracies[epoch] = valid_a_top1
if valid_a_top1 > valid_accuracies["best"]:
valid_accuracies["best"] = valid_a_top1
genotypes["best"] = search_model.genotype()
find_best = True
else:
find_best = False
genotypes[epoch] = search_model.genotype()
logger.log(
"<<<--->>> The {:}-th epoch : {:}".format(epoch_str, genotypes[epoch])
)
# save checkpoint
save_path = save_checkpoint(
{
"epoch": epoch + 1,
"args": deepcopy(xargs),
"search_model": search_model.state_dict(),
"w_optimizer": w_optimizer.state_dict(),
"a_optimizer": a_optimizer.state_dict(),
"w_scheduler": w_scheduler.state_dict(),
"genotypes": genotypes,
"valid_accuracies": valid_accuracies,
},
model_base_path,
logger,
)
last_info = save_checkpoint(
{
"epoch": epoch + 1,
"args": deepcopy(args),
"last_checkpoint": save_path,
},
logger.path("info"),
logger,
)
if find_best:
logger.log(
"<<<--->>> The {:}-th epoch : find the highest validation accuracy : {:.2f}%.".format(
epoch_str, valid_a_top1
)
)
copy_checkpoint(model_base_path, model_best_path, logger)
with torch.no_grad():
logger.log("{:}".format(search_model.show_alphas()))
if api is not None:
logger.log("{:}".format(api.query_by_arch(genotypes[epoch], "200")))
# measure elapsed time
epoch_time.update(time.time() - start_time)
start_time = time.time()
logger.log("\n" + "-" * 100)
# check the performance from the architecture dataset
logger.log(
"GDAS : run {:} epochs, cost {:.1f} s, last-geno is {:}.".format(
total_epoch, search_time.sum, genotypes[total_epoch - 1]
)
)
if api is not None:
logger.log("{:}".format(api.query_by_arch(genotypes[total_epoch - 1], "200")))
logger.close()
args = parser.parse_args()
args.verbose = True if args.verbose == 'True' else False
args.fix_seed = True if args.fix_seed == 'True' else False
max_budget = args.max_budget
dataset = args.dataset
# Directory where files will be written
if args. async is None:
folder = "de_pop{}".format(args.pop_size)
else:
folder = "ade_{}_pop{}".format(args. async, args.pop_size)
output_path = os.path.join(args.output_path, args.dataset, folder)
os.makedirs(output_path, exist_ok=True)
# Loading NAS-201
api = API(args.data_dir)
search_space = get_search_spaces('cell', 'nas-bench-201')
# Parameter space to be used by DE
cs = get_configuration_space(args.max_nodes, search_space)
dimensions = len(cs.get_hyperparameters())
config2structure = config2structure_func(args.max_nodes)
y_star_valid, y_star_test = find_nas201_best(api, dataset)
inc_config = cs.get_default_configuration().get_array().tolist()
# Custom objective function for DE to interface NASBench-201
def f(config, budget=max_budget):
global dataset, api
structure = config2structure(config)
def main(xargs, myargs):
assert torch.cuda.is_available(), 'CUDA is not available.'
torch.backends.cudnn.enabled = True
torch.backends.cudnn.benchmark = False
torch.backends.cudnn.deterministic = True
torch.set_num_threads(xargs.workers)
prepare_seed(xargs.rand_seed)
logger = prepare_logger(xargs)
train_data, valid_data, xshape, class_num = get_datasets(
xargs.dataset, xargs.data_path, -1)
config = load_config(xargs.config_path, {
'class_num': class_num,
'xshape': xshape
}, logger)
search_loader, _, valid_loader = get_nas_search_loaders(
train_data, valid_data, xargs.dataset,
'AutoDL-Projects/configs/nas-benchmark/',
(config.batch_size, config.test_batch_size), xargs.num_worker)
logger.log(
'||||||| {:10s} ||||||| Search-Loader-Num={:}, Valid-Loader-Num={:}, batch size={:}'
.format(xargs.dataset, len(search_loader), len(valid_loader),
config.batch_size))
logger.log('||||||| {:10s} ||||||| Config={:}'.format(
xargs.dataset, config))
search_space = get_search_spaces('cell', xargs.search_space_name)
if not hasattr(xargs, 'model_config') or xargs.model_config is None:
model_config = dict2config(
dict(name='SETN',
C=xargs.channel,
N=xargs.num_cells,
max_nodes=xargs.max_nodes,
num_classes=class_num,
space=search_space,
affine=False,
track_running_stats=bool(xargs.track_running_stats)), None)
else:
model_config = load_config(
xargs.model_config,
dict(num_classes=class_num,
space=search_space,
affine=False,
track_running_stats=bool(xargs.track_running_stats)), None)
logger.log('search space : {:}'.format(search_space))
search_model = get_cell_based_tiny_net(model_config)
w_optimizer, w_scheduler, criterion = get_optim_scheduler(
search_model.get_weights(), config)
a_optimizer = torch.optim.Adam(search_model.get_alphas(),
lr=xargs.arch_learning_rate,
betas=(0.5, 0.999),
weight_decay=xargs.arch_weight_decay)
logger.log('w-optimizer : {:}'.format(w_optimizer))
logger.log('a-optimizer : {:}'.format(a_optimizer))
logger.log('w-scheduler : {:}'.format(w_scheduler))
logger.log('criterion : {:}'.format(criterion))
flop, param = get_model_infos(search_model, xshape)
logger.log('FLOP = {:.2f} M, Params = {:.2f} MB'.format(flop, param))
logger.log('search-space : {:}'.format(search_space))
if xargs.arch_nas_dataset is None:
api = None
else:
api = API(xargs.arch_nas_dataset)
logger.log('{:} create API = {:} done'.format(time_string(), api))
last_info, model_base_path, model_best_path = logger.path(
'info'), logger.path('model'), logger.path('best')
network, criterion = torch.nn.DataParallel(
search_model).cuda(), criterion.cuda()
if last_info.exists(): # automatically resume from previous checkpoint
logger.log("=> loading checkpoint of the last-info '{:}' start".format(
last_info))
last_info = torch.load(last_info)
start_epoch = last_info['epoch']
checkpoint = torch.load(last_info['last_checkpoint'])
genotypes = checkpoint['genotypes']
valid_accuracies = checkpoint['valid_accuracies']
search_model.load_state_dict(checkpoint['search_model'])
w_scheduler.load_state_dict(checkpoint['w_scheduler'])
w_optimizer.load_state_dict(checkpoint['w_optimizer'])
a_optimizer.load_state_dict(checkpoint['a_optimizer'])
logger.log(
"=> loading checkpoint of the last-info '{:}' start with {:}-th epoch."
.format(last_info, start_epoch))
else:
logger.log("=> do not find the last-info file : {:}".format(last_info))
init_genotype, _ = get_best_arch(valid_loader, network,
xargs.select_num)
start_epoch, valid_accuracies, genotypes = 0, {
'best': -1
}, {
-1: init_genotype
}
# start training
start_time, search_time, epoch_time, total_epoch = time.time(
), AverageMeter(), AverageMeter(), config.epochs + config.warmup
for epoch in range(start_epoch, total_epoch):
w_scheduler.update(epoch, 0.0)
need_time = 'Time Left: {:}'.format(
convert_secs2time(epoch_time.val * (total_epoch - epoch), True))
epoch_str = '{:03d}-{:03d}'.format(epoch, total_epoch)
logger.log('\n[Search the {:}-th epoch] {:}, LR={:}'.format(
epoch_str, need_time, min(w_scheduler.get_lr())))
search_w_loss, search_w_top1, search_w_top5, search_a_loss, search_a_top1, search_a_top5 \
= search_func(search_loader, network, criterion, w_scheduler, w_optimizer, a_optimizer, epoch_str, xargs.print_freq, logger)
search_time.update(time.time() - start_time)
logger.log(
'[{:}] search [base] : loss={:.2f}, accuracy@1={:.2f}%, accuracy@5={:.2f}%, time-cost={:.1f} s'
.format(epoch_str, search_w_loss, search_w_top1, search_w_top5,
search_time.sum))
logger.log(
'[{:}] search [arch] : loss={:.2f}, accuracy@1={:.2f}%, accuracy@5={:.2f}%'
.format(epoch_str, search_a_loss, search_a_top1, search_a_top5))
genotype, temp_accuracy = get_best_arch(valid_loader, network,
xargs.select_num)
network.module.set_cal_mode('dynamic', genotype)
valid_a_loss, valid_a_top1, valid_a_top5 = valid_func(
valid_loader, network, criterion)
logger.log(
'[{:}] evaluate : loss={:.2f}, accuracy@1={:.2f}%, accuracy@5={:.2f}% | {:}'
.format(epoch_str, valid_a_loss, valid_a_top1, valid_a_top5,
genotype))
#search_model.set_cal_mode('urs')
#valid_a_loss , valid_a_top1 , valid_a_top5 = valid_func(valid_loader, network, criterion)
#logger.log('[{:}] URS---evaluate : loss={:.2f}, accuracy@1={:.2f}%, accuracy@5={:.2f}%'.format(epoch_str, valid_a_loss, valid_a_top1, valid_a_top5))
#search_model.set_cal_mode('joint')
#valid_a_loss , valid_a_top1 , valid_a_top5 = valid_func(valid_loader, network, criterion)
#logger.log('[{:}] JOINT-evaluate : loss={:.2f}, accuracy@1={:.2f}%, accuracy@5={:.2f}%'.format(epoch_str, valid_a_loss, valid_a_top1, valid_a_top5))
#search_model.set_cal_mode('select')
#valid_a_loss , valid_a_top1 , valid_a_top5 = valid_func(valid_loader, network, criterion)
#logger.log('[{:}] Selec-evaluate : loss={:.2f}, accuracy@1={:.2f}%, accuracy@5={:.2f}%'.format(epoch_str, valid_a_loss, valid_a_top1, valid_a_top5))
# check the best accuracy
valid_accuracies[epoch] = valid_a_top1
genotypes[epoch] = genotype
logger.log('<<<--->>> The {:}-th epoch : {:}'.format(
epoch_str, genotypes[epoch]))
# save checkpoint
save_path = save_checkpoint(
{
'epoch': epoch + 1,
'args': deepcopy(xargs),
'search_model': search_model.state_dict(),
'w_optimizer': w_optimizer.state_dict(),
'a_optimizer': a_optimizer.state_dict(),
'w_scheduler': w_scheduler.state_dict(),
'genotypes': genotypes,
'valid_accuracies': valid_accuracies
}, model_base_path, logger)
last_info = save_checkpoint(
{
'epoch': epoch + 1,
'args': deepcopy(xargs),
'last_checkpoint': save_path,
}, logger.path('info'), logger)
with torch.no_grad():
logger.log('{:}'.format(search_model.show_alphas()))
if api is not None:
logger.log('{:}'.format(api.query_by_arch(genotypes[epoch],
'200')))
# measure elapsed time
epoch_time.update(time.time() - start_time)
start_time = time.time()
# the final post procedure : count the time
start_time = time.time()
genotype, temp_accuracy = get_best_arch(valid_loader, network,
xargs.select_num)
search_time.update(time.time() - start_time)
network.module.set_cal_mode('dynamic', genotype)
valid_a_loss, valid_a_top1, valid_a_top5 = valid_func(
valid_loader, network, criterion)
logger.log(
'Last : the gentotype is : {:}, with the validation accuracy of {:.3f}%.'
.format(genotype, valid_a_top1))
logger.log('\n' + '-' * 100)
# check the performance from the architecture dataset
logger.log(
'SETN : run {:} epochs, cost {:.1f} s, last-geno is {:}.'.format(
total_epoch, search_time.sum, genotype))
if api is not None:
logger.log('{:}'.format(api.query_by_arch(genotype, '200')))
logger.close()
def main(xargs):
assert torch.cuda.is_available(), 'CUDA is not available.'
torch.backends.cudnn.enabled = True
torch.backends.cudnn.benchmark = False
torch.backends.cudnn.deterministic = True
torch.set_num_threads( xargs.workers )
prepare_seed(xargs.rand_seed)
logger = prepare_logger(args)
train_data, test_data, xshape, class_num = get_datasets(xargs.dataset, xargs.data_path, -1)
logger.log('use config from : {:}'.format(xargs.config_path))
config = load_config(xargs.config_path, {'class_num': class_num, 'xshape': xshape}, logger)
_, train_loader, valid_loader = get_nas_search_loaders(train_data, test_data, xargs.dataset, 'configs/nas-benchmark/', config.batch_size, xargs.workers)
# since ENAS will train the controller on valid-loader, we need to use train transformation for valid-loader
valid_loader.dataset.transform = deepcopy(train_loader.dataset.transform)
if hasattr(valid_loader.dataset, 'transforms'):
valid_loader.dataset.transforms = deepcopy(train_loader.dataset.transforms)
# data loader
logger.log('||||||| {:10s} ||||||| Train-Loader-Num={:}, Valid-Loader-Num={:}, batch size={:}'.format(xargs.dataset, len(train_loader), len(valid_loader), config.batch_size))
logger.log('||||||| {:10s} ||||||| Config={:}'.format(xargs.dataset, config))
search_space = get_search_spaces('cell', xargs.search_space_name)
model_config = dict2config({'name': 'ENAS', 'C': xargs.channel, 'N': xargs.num_cells,
'max_nodes': xargs.max_nodes, 'num_classes': class_num,
'space' : search_space,
'affine' : False, 'track_running_stats': bool(xargs.track_running_stats)}, None)
shared_cnn = get_cell_based_tiny_net(model_config)
controller = shared_cnn.create_controller()
w_optimizer, w_scheduler, criterion = get_optim_scheduler(shared_cnn.parameters(), config)
a_optimizer = torch.optim.Adam(controller.parameters(), lr=config.controller_lr, betas=config.controller_betas, eps=config.controller_eps)
logger.log('w-optimizer : {:}'.format(w_optimizer))
logger.log('a-optimizer : {:}'.format(a_optimizer))
logger.log('w-scheduler : {:}'.format(w_scheduler))
logger.log('criterion : {:}'.format(criterion))
#flop, param = get_model_infos(shared_cnn, xshape)
#logger.log('{:}'.format(shared_cnn))
#logger.log('FLOP = {:.2f} M, Params = {:.2f} MB'.format(flop, param))
logger.log('search-space : {:}'.format(search_space))
if xargs.arch_nas_dataset is None:
api = None
else:
api = API(xargs.arch_nas_dataset)
logger.log('{:} create API = {:} done'.format(time_string(), api))
shared_cnn, controller, criterion = torch.nn.DataParallel(shared_cnn).cuda(), controller.cuda(), criterion.cuda()
last_info, model_base_path, model_best_path = logger.path('info'), logger.path('model'), logger.path('best')
if last_info.exists(): # automatically resume from previous checkpoint
logger.log("=> loading checkpoint of the last-info '{:}' start".format(last_info))
last_info = torch.load(last_info)
start_epoch = last_info['epoch']
checkpoint = torch.load(last_info['last_checkpoint'])
genotypes = checkpoint['genotypes']
baseline = checkpoint['baseline']
valid_accuracies = checkpoint['valid_accuracies']
shared_cnn.load_state_dict( checkpoint['shared_cnn'] )
controller.load_state_dict( checkpoint['controller'] )
w_scheduler.load_state_dict ( checkpoint['w_scheduler'] )
w_optimizer.load_state_dict ( checkpoint['w_optimizer'] )
a_optimizer.load_state_dict ( checkpoint['a_optimizer'] )
logger.log("=> loading checkpoint of the last-info '{:}' start with {:}-th epoch.".format(last_info, start_epoch))
else:
logger.log("=> do not find the last-info file : {:}".format(last_info))
start_epoch, valid_accuracies, genotypes, baseline = 0, {'best': -1}, {}, None
# start training
start_time, search_time, epoch_time, total_epoch = time.time(), AverageMeter(), AverageMeter(), config.epochs + config.warmup
for epoch in range(start_epoch, total_epoch):
w_scheduler.update(epoch, 0.0)
need_time = 'Time Left: {:}'.format( convert_secs2time(epoch_time.val * (total_epoch-epoch), True) )
epoch_str = '{:03d}-{:03d}'.format(epoch, total_epoch)
logger.log('\n[Search the {:}-th epoch] {:}, LR={:}, baseline={:}'.format(epoch_str, need_time, min(w_scheduler.get_lr()), baseline))
cnn_loss, cnn_top1, cnn_top5 = train_shared_cnn(train_loader, shared_cnn, controller, criterion, w_scheduler, w_optimizer, epoch_str, xargs.print_freq, logger)
logger.log('[{:}] shared-cnn : loss={:.2f}, accuracy@1={:.2f}%, accuracy@5={:.2f}%'.format(epoch_str, cnn_loss, cnn_top1, cnn_top5))
ctl_loss, ctl_acc, ctl_baseline, ctl_reward, baseline \
= train_controller(valid_loader, shared_cnn, controller, criterion, a_optimizer, \
dict2config({'baseline': baseline,
'ctl_train_steps': xargs.controller_train_steps, 'ctl_num_aggre': xargs.controller_num_aggregate,
'ctl_entropy_w': xargs.controller_entropy_weight,
'ctl_bl_dec' : xargs.controller_bl_dec}, None), \
epoch_str, xargs.print_freq, logger)
search_time.update(time.time() - start_time)
logger.log('[{:}] controller : loss={:.2f}, accuracy={:.2f}%, baseline={:.2f}, reward={:.2f}, current-baseline={:.4f}, time-cost={:.1f} s'.format(epoch_str, ctl_loss, ctl_acc, ctl_baseline, ctl_reward, baseline, search_time.sum))
best_arch, _ = get_best_arch(controller, shared_cnn, valid_loader)
shared_cnn.module.update_arch(best_arch)
_, best_valid_acc, _ = valid_func(valid_loader, shared_cnn, criterion)
genotypes[epoch] = best_arch
# check the best accuracy
valid_accuracies[epoch] = best_valid_acc
if best_valid_acc > valid_accuracies['best']:
valid_accuracies['best'] = best_valid_acc
genotypes['best'] = best_arch
find_best = True
else: find_best = False
logger.log('<<<--->>> The {:}-th epoch : {:}'.format(epoch_str, genotypes[epoch]))
# save checkpoint
save_path = save_checkpoint({'epoch' : epoch + 1,
'args' : deepcopy(xargs),
'baseline' : baseline,
'shared_cnn' : shared_cnn.state_dict(),
'controller' : controller.state_dict(),
'w_optimizer' : w_optimizer.state_dict(),
'a_optimizer' : a_optimizer.state_dict(),
'w_scheduler' : w_scheduler.state_dict(),
'genotypes' : genotypes,
'valid_accuracies' : valid_accuracies},
model_base_path, logger)
last_info = save_checkpoint({
'epoch': epoch + 1,
'args' : deepcopy(args),
'last_checkpoint': save_path,
}, logger.path('info'), logger)
if find_best:
logger.log('<<<--->>> The {:}-th epoch : find the highest validation accuracy : {:.2f}%.'.format(epoch_str, best_valid_acc))
copy_checkpoint(model_base_path, model_best_path, logger)
if api is not None: logger.log('{:}'.format(api.query_by_arch( genotypes[epoch] )))
# measure elapsed time
epoch_time.update(time.time() - start_time)
start_time = time.time()
logger.log('\n' + '-'*100)
logger.log('During searching, the best architecture is {:}'.format(genotypes['best']))
logger.log('Its accuracy is {:.2f}%'.format(valid_accuracies['best']))
logger.log('Randomly select {:} architectures and select the best.'.format(xargs.controller_num_samples))
start_time = time.time()
final_arch, _ = get_best_arch(controller, shared_cnn, valid_loader, xargs.controller_num_samples)
search_time.update(time.time() - start_time)
shared_cnn.module.update_arch(final_arch)
final_loss, final_top1, final_top5 = valid_func(valid_loader, shared_cnn, criterion)
logger.log('The Selected Final Architecture : {:}'.format(final_arch))
logger.log('Loss={:.3f}, Accuracy@1={:.2f}%, Accuracy@5={:.2f}%'.format(final_loss, final_top1, final_top5))
logger.log('ENAS : run {:} epochs, cost {:.1f} s, last-geno is {:}.'.format(total_epoch, search_time.sum, final_arch))
if api is not None: logger.log('{:}'.format( api.query_by_arch(final_arch) ))
logger.close()