def generate_meta_info(save_dir, max_node, divide=40):
aa_nas_bench_ss = get_search_spaces("cell", "nas-bench-201")
archs = CellStructure.gen_all(aa_nas_bench_ss, max_node, False)
print("There are {:} archs vs {:}.".format(
len(archs),
len(aa_nas_bench_ss)**((max_node - 1) * max_node / 2)))
random.seed(88) # please do not change this line for reproducibility
random.shuffle(archs)
# to test fixed-random shuffle
# print ('arch [0] : {:}\n---->>>> {:}'.format( archs[0], archs[0].tostr() ))
# print ('arch [9] : {:}\n---->>>> {:}'.format( archs[9], archs[9].tostr() ))
assert (
archs[0].tostr() ==
"|avg_pool_3x3~0|+|nor_conv_1x1~0|skip_connect~1|+|nor_conv_1x1~0|skip_connect~1|skip_connect~2|"
), "please check the 0-th architecture : {:}".format(archs[0])
assert (
archs[9].tostr() ==
"|avg_pool_3x3~0|+|none~0|none~1|+|skip_connect~0|none~1|nor_conv_3x3~2|"
), "please check the 9-th architecture : {:}".format(archs[9])
assert (
archs[123].tostr() ==
"|avg_pool_3x3~0|+|avg_pool_3x3~0|nor_conv_1x1~1|+|none~0|avg_pool_3x3~1|nor_conv_3x3~2|"
), "please check the 123-th architecture : {:}".format(archs[123])
total_arch = len(archs)
num = 50000
indexes_5W = list(range(num))
random.seed(1021)
random.shuffle(indexes_5W)
train_split = sorted(list(set(indexes_5W[:num // 2])))
valid_split = sorted(list(set(indexes_5W[num // 2:])))
assert len(train_split) + len(valid_split) == num
assert (train_split[0] == 0 and train_split[10] == 26
and train_split[111] == 203 and valid_split[0] == 1
and valid_split[10] == 18
and valid_split[111] == 242), "{:} {:} {:} - {:} {:} {:}".format(
train_split[0],
train_split[10],
train_split[111],
valid_split[0],
valid_split[10],
valid_split[111],
)
splits = {num: {"train": train_split, "valid": valid_split}}
info = {
"archs": [x.tostr() for x in archs],
"total": total_arch,
"max_node": max_node,
"splits": splits,
}
save_dir = Path(save_dir)
save_dir.mkdir(parents=True, exist_ok=True)
save_name = save_dir / "meta-node-{:}.pth".format(max_node)
assert not save_name.exists(), "{:} already exist".format(save_name)
torch.save(info, save_name)
print("save the meta file into {:}".format(save_name))
def main(xargs, api):
torch.set_num_threads(4)
prepare_seed(xargs.rand_seed)
logger = prepare_logger(args)
logger.log("{:} use api : {:}".format(time_string(), api))
api.reset_time()
search_space = get_search_spaces(xargs.search_space, "nats-bench")
if xargs.search_space == "tss":
random_arch = random_topology_func(search_space)
else:
random_arch = random_size_func(search_space)
best_arch, best_acc, total_time_cost, history = None, -1, [], []
current_best_index = []
while len(total_time_cost) == 0 or total_time_cost[-1] < xargs.time_budget:
arch = random_arch()
accuracy, _, _, total_cost = api.simulate_train_eval(
arch, xargs.dataset, hp="12"
)
total_time_cost.append(total_cost)
history.append(arch)
if best_arch is None or best_acc < accuracy:
best_acc, best_arch = accuracy, arch
logger.log(
"[{:03d}] : {:} : accuracy = {:.2f}%".format(len(history), arch, accuracy)
)
current_best_index.append(api.query_index_by_arch(best_arch))
logger.log(
"{:} best arch is {:}, accuracy = {:.2f}%, visit {:} archs with {:.1f} s.".format(
time_string(), best_arch, best_acc, len(history), total_time_cost[-1]
)
)
info = api.query_info_str_by_arch(
best_arch, "200" if xargs.search_space == "tss" else "90"
)
logger.log("{:}".format(info))
logger.log("-" * 100)
logger.close()
return logger.log_dir, current_best_index, total_time_cost
def main(xargs, api):
torch.set_num_threads(4)
prepare_seed(xargs.rand_seed)
logger = prepare_logger(args)
search_space = get_search_spaces(xargs.search_space, "nats-bench")
if xargs.search_space == "tss":
random_arch = random_topology_func(search_space)
mutate_arch = mutate_topology_func(search_space)
else:
random_arch = random_size_func(search_space)
mutate_arch = mutate_size_func(search_space)
x_start_time = time.time()
logger.log("{:} use api : {:}".format(time_string(), api))
logger.log("-" * 30 +
" start searching with the time budget of {:} s".format(
xargs.time_budget))
history, current_best_index, total_times = regularized_evolution(
xargs.ea_cycles,
xargs.ea_population,
xargs.ea_sample_size,
xargs.time_budget,
random_arch,
mutate_arch,
api,
xargs.use_proxy > 0,
xargs.dataset,
)
logger.log(
"{:} regularized_evolution finish with history of {:} arch with {:.1f} s (real-cost={:.2f} s)."
.format(time_string(), len(history), total_times[-1],
time.time() - x_start_time))
best_arch = max(history, key=lambda x: x[0])[1]
logger.log("{:} best arch is {:}".format(time_string(), best_arch))
info = api.query_info_str_by_arch(
best_arch, "200" if xargs.search_space == "tss" else "90")
logger.log("{:}".format(info))
logger.log("-" * 100)
logger.close()
return logger.log_dir, current_best_index, total_times
def traverse_net(max_node):
aa_nas_bench_ss = get_search_spaces("cell", "nats-bench")
archs = CellStructure.gen_all(aa_nas_bench_ss, max_node, False)
print("There are {:} archs vs {:}.".format(
len(archs),
len(aa_nas_bench_ss)**((max_node - 1) * max_node / 2)))
random.seed(88) # please do not change this line for reproducibility
random.shuffle(archs)
assert (
archs[0].tostr() ==
"|avg_pool_3x3~0|+|nor_conv_1x1~0|skip_connect~1|+|nor_conv_1x1~0|skip_connect~1|skip_connect~2|"
), "please check the 0-th architecture : {:}".format(archs[0])
assert (
archs[9].tostr() ==
"|avg_pool_3x3~0|+|none~0|none~1|+|skip_connect~0|none~1|nor_conv_3x3~2|"
), "please check the 9-th architecture : {:}".format(archs[9])
assert (
archs[123].tostr() ==
"|avg_pool_3x3~0|+|avg_pool_3x3~0|nor_conv_1x1~1|+|none~0|avg_pool_3x3~1|nor_conv_3x3~2|"
), "please check the 123-th architecture : {:}".format(archs[123])
return [x.tostr() for x in archs]
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)
if 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(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],
"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],
"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, nas_bench):
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)
if xargs.dataset == "cifar10":
dataname = "cifar10-valid"
else:
dataname = xargs.dataset
if xargs.data_path is not None:
train_data, valid_data, xshape, class_num = get_datasets(
xargs.dataset, xargs.data_path, -1)
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))
config_path = "configs/nas-benchmark/algos/R-EA.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
train_loader = torch.utils.data.DataLoader(
train_data,
batch_size=config.batch_size,
sampler=torch.utils.data.sampler.SubsetRandomSampler(train_split),
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} ||||||| 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))
extra_info = {
"config": config,
"train_loader": train_loader,
"valid_loader": valid_loader,
}
else:
config_path = "configs/nas-benchmark/algos/R-EA.config"
config = load_config(config_path, None, logger)
logger.log("||||||| {:10s} ||||||| Config={:}".format(
xargs.dataset, config))
extra_info = {
"config": config,
"train_loader": None,
"valid_loader": None
}
search_space = get_search_spaces("cell", xargs.search_space_name)
random_arch = random_architecture_func(xargs.max_nodes, search_space)
# x =random_arch() ; y = mutate_arch(x)
x_start_time = time.time()
logger.log("{:} use nas_bench : {:}".format(time_string(), nas_bench))
best_arch, best_acc, total_time_cost, history = None, -1, 0, []
# for idx in range(xargs.random_num):
while total_time_cost < xargs.time_budget:
arch = random_arch()
accuracy, cost_time = train_and_eval(arch, nas_bench, extra_info,
dataname)
if total_time_cost + cost_time > xargs.time_budget:
break
else:
total_time_cost += cost_time
history.append(arch)
if best_arch is None or best_acc < accuracy:
best_acc, best_arch = accuracy, arch
logger.log("[{:03d}] : {:} : accuracy = {:.2f}%".format(
len(history), arch, accuracy))
logger.log(
"{:} best arch is {:}, accuracy = {:.2f}%, visit {:} archs with {:.1f} s (real-cost = {:.3f} s)."
.format(
time_string(),
best_arch,
best_acc,
len(history),
total_time_cost,
time.time() - x_start_time,
))
info = nas_bench.query_by_arch(best_arch, "200")
if info is None:
logger.log("Did not find this architecture : {:}.".format(best_arch))
else:
logger.log("{:}".format(info))
logger.log("-" * 100)
logger.close()
return logger.log_dir, nas_bench.query_index_by_arch(best_arch)
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, "nats-bench")
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 = create(None, "size", fast_mode=True, 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)
if (
xargs.warmup_ratio is None
or xargs.warmup_ratio <= float(epoch) / total_epoch
):
enable_controller = True
network.set_warmup_ratio(None)
else:
enable_controller = False
network.set_warmup_ratio(
1.0 - float(epoch) / total_epoch / xargs.warmup_ratio
)
logger.log(
"\n[Search the {:}-th epoch] {:}, LR={:}, controller-warmup={:}, enable_controller={:}".format(
epoch_str,
need_time,
min(w_scheduler.get_lr()),
network.warmup_ratio,
enable_controller,
)
)
if xargs.algo == "mask_gumbel" 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,
enable_controller,
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, nas_bench):
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)
if xargs.dataset == "cifar10":
dataname = "cifar10-valid"
else:
dataname = xargs.dataset
if xargs.data_path is not None:
train_data, valid_data, xshape, class_num = get_datasets(
xargs.dataset, xargs.data_path, -1)
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))
config_path = "configs/nas-benchmark/algos/R-EA.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
train_loader = torch.utils.data.DataLoader(
train_data,
batch_size=config.batch_size,
sampler=torch.utils.data.sampler.SubsetRandomSampler(train_split),
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} ||||||| 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))
extra_info = {
"config": config,
"train_loader": train_loader,
"valid_loader": valid_loader,
}
else:
config_path = "configs/nas-benchmark/algos/R-EA.config"
config = load_config(config_path, None, logger)
extra_info = {
"config": config,
"train_loader": None,
"valid_loader": None
}
logger.log("||||||| {:10s} ||||||| Config={:}".format(
xargs.dataset, config))
search_space = get_search_spaces("cell", xargs.search_space_name)
policy = Policy(xargs.max_nodes, search_space)
optimizer = torch.optim.Adam(policy.parameters(), lr=xargs.learning_rate)
# optimizer = torch.optim.SGD(policy.parameters(), lr=xargs.learning_rate)
eps = np.finfo(np.float32).eps.item()
baseline = ExponentialMovingAverage(xargs.EMA_momentum)
logger.log("policy : {:}".format(policy))
logger.log("optimizer : {:}".format(optimizer))
logger.log("eps : {:}".format(eps))
# nas dataset load
logger.log("{:} use nas_bench : {:}".format(time_string(), nas_bench))
# REINFORCE
# attempts = 0
x_start_time = time.time()
logger.log("Will start searching with time budget of {:} s.".format(
xargs.time_budget))
total_steps, total_costs, trace = 0, 0, []
# for istep in range(xargs.RL_steps):
while total_costs < xargs.time_budget:
start_time = time.time()
log_prob, action = select_action(policy)
arch = policy.generate_arch(action)
reward, cost_time = train_and_eval(arch, nas_bench, extra_info,
dataname)
trace.append((reward, arch))
# accumulate time
if total_costs + cost_time < xargs.time_budget:
total_costs += cost_time
else:
break
baseline.update(reward)
# calculate loss
policy_loss = (-log_prob * (reward - baseline.value())).sum()
optimizer.zero_grad()
policy_loss.backward()
optimizer.step()
# accumulate time
total_costs += time.time() - start_time
total_steps += 1
logger.log(
"step [{:3d}] : average-reward={:.3f} : policy_loss={:.4f} : {:}".
format(total_steps, baseline.value(), policy_loss.item(),
policy.genotype()))
# logger.log('----> {:}'.format(policy.arch_parameters))
# logger.log('')
# best_arch = policy.genotype() # first version
best_arch = max(trace, key=lambda x: x[0])[1]
logger.log(
"REINFORCE finish with {:} steps and {:.1f} s (real cost={:.3f}).".
format(total_steps, total_costs,
time.time() - x_start_time))
info = nas_bench.query_by_arch(best_arch, "200")
if info is None:
logger.log("Did not find this architecture : {:}.".format(best_arch))
else:
logger.log("{:}".format(info))
logger.log("-" * 100)
logger.close()
return logger.log_dir, nas_bench.query_index_by_arch(best_arch)
def main(xargs, nas_bench):
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)
if xargs.dataset == "cifar10":
dataname = "cifar10-valid"
else:
dataname = xargs.dataset
if xargs.data_path is not None:
train_data, valid_data, xshape, class_num = get_datasets(
xargs.dataset, xargs.data_path, -1)
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))
config_path = "configs/nas-benchmark/algos/R-EA.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
train_loader = torch.utils.data.DataLoader(
train_data,
batch_size=config.batch_size,
sampler=torch.utils.data.sampler.SubsetRandomSampler(train_split),
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} ||||||| 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))
extra_info = {
"config": config,
"train_loader": train_loader,
"valid_loader": valid_loader,
}
else:
config_path = "configs/nas-benchmark/algos/R-EA.config"
config = load_config(config_path, None, logger)
logger.log("||||||| {:10s} ||||||| Config={:}".format(
xargs.dataset, config))
extra_info = {
"config": config,
"train_loader": None,
"valid_loader": None
}
search_space = get_search_spaces("cell", xargs.search_space_name)
random_arch = random_architecture_func(xargs.max_nodes, search_space)
mutate_arch = mutate_arch_func(search_space)
# x =random_arch() ; y = mutate_arch(x)
x_start_time = time.time()
logger.log("{:} use nas_bench : {:}".format(time_string(), nas_bench))
logger.log("-" * 30 +
" start searching with the time budget of {:} s".format(
xargs.time_budget))
history, total_cost = regularized_evolution(
xargs.ea_cycles,
xargs.ea_population,
xargs.ea_sample_size,
xargs.time_budget,
random_arch,
mutate_arch,
nas_bench if args.ea_fast_by_api else None,
extra_info,
dataname,
)
logger.log(
"{:} regularized_evolution finish with history of {:} arch with {:.1f} s (real-cost={:.2f} s)."
.format(time_string(), len(history), total_cost,
time.time() - x_start_time))
best_arch = max(history, key=lambda i: i.accuracy)
best_arch = best_arch.arch
logger.log("{:} best arch is {:}".format(time_string(), best_arch))
info = nas_bench.query_by_arch(best_arch, "200")
if info is None:
logger.log("Did not find this architecture : {:}.".format(best_arch))
else:
logger.log("{:}".format(info))
logger.log("-" * 100)
logger.close()
return logger.log_dir, nas_bench.query_index_by_arch(best_arch)
def generate_meta_info(save_dir, max_node, divide=40):
aa_nas_bench_ss = get_search_spaces("cell", "nas-bench-201")
archs = CellStructure.gen_all(aa_nas_bench_ss, max_node, False)
print("There are {:} archs vs {:}.".format(
len(archs),
len(aa_nas_bench_ss)**((max_node - 1) * max_node / 2)))
random.seed(88) # please do not change this line for reproducibility
random.shuffle(archs)
# to test fixed-random shuffle
# print ('arch [0] : {:}\n---->>>> {:}'.format( archs[0], archs[0].tostr() ))
# print ('arch [9] : {:}\n---->>>> {:}'.format( archs[9], archs[9].tostr() ))
assert (
archs[0].tostr() ==
"|avg_pool_3x3~0|+|nor_conv_1x1~0|skip_connect~1|+|nor_conv_1x1~0|skip_connect~1|skip_connect~2|"
), "please check the 0-th architecture : {:}".format(archs[0])
assert (
archs[9].tostr() ==
"|avg_pool_3x3~0|+|none~0|none~1|+|skip_connect~0|none~1|nor_conv_3x3~2|"
), "please check the 9-th architecture : {:}".format(archs[9])
assert (
archs[123].tostr() ==
"|avg_pool_3x3~0|+|avg_pool_3x3~0|nor_conv_1x1~1|+|none~0|avg_pool_3x3~1|nor_conv_3x3~2|"
), "please check the 123-th architecture : {:}".format(archs[123])
total_arch = len(archs)
num = 50000
indexes_5W = list(range(num))
random.seed(1021)
random.shuffle(indexes_5W)
train_split = sorted(list(set(indexes_5W[:num // 2])))
valid_split = sorted(list(set(indexes_5W[num // 2:])))
assert len(train_split) + len(valid_split) == num
assert (train_split[0] == 0 and train_split[10] == 26
and train_split[111] == 203 and valid_split[0] == 1
and valid_split[10] == 18
and valid_split[111] == 242), "{:} {:} {:} - {:} {:} {:}".format(
train_split[0],
train_split[10],
train_split[111],
valid_split[0],
valid_split[10],
valid_split[111],
)
splits = {num: {"train": train_split, "valid": valid_split}}
info = {
"archs": [x.tostr() for x in archs],
"total": total_arch,
"max_node": max_node,
"splits": splits,
}
save_dir = Path(save_dir)
save_dir.mkdir(parents=True, exist_ok=True)
save_name = save_dir / "meta-node-{:}.pth".format(max_node)
assert not save_name.exists(), "{:} already exist".format(save_name)
torch.save(info, save_name)
print("save the meta file into {:}".format(save_name))
script_name_full = save_dir / "BENCH-201-N{:}.opt-full.script".format(
max_node)
script_name_less = save_dir / "BENCH-201-N{:}.opt-less.script".format(
max_node)
full_file = open(str(script_name_full), "w")
less_file = open(str(script_name_less), "w")
gaps = total_arch // divide
for start in range(0, total_arch, gaps):
xend = min(start + gaps, total_arch)
full_file.write(
"bash ./scripts-search/NAS-Bench-201/train-models.sh 0 {:5d} {:5d} -1 '777 888 999'\n"
.format(start, xend - 1))
less_file.write(
"bash ./scripts-search/NAS-Bench-201/train-models.sh 1 {:5d} {:5d} -1 '777 888 999'\n"
.format(start, xend - 1))
print("save the training script into {:} and {:}".format(
script_name_full, script_name_less))
full_file.close()
less_file.close()
script_name = save_dir / "meta-node-{:}.cal-script.txt".format(max_node)
macro = "OMP_NUM_THREADS=6 CUDA_VISIBLE_DEVICES=0"
with open(str(script_name), "w") as cfile:
for start in range(0, total_arch, gaps):
xend = min(start + gaps, total_arch)
cfile.write(
"{:} python exps/NAS-Bench-201/statistics.py --mode cal --target_dir {:06d}-{:06d}-C16-N5\n"
.format(macro, start, xend - 1))
print("save the post-processing script into {:}".format(script_name))
def train_single_model(save_dir, workers, datasets, xpaths, splits, use_less,
seeds, model_str, arch_config):
assert torch.cuda.is_available(), "CUDA is not available."
torch.backends.cudnn.enabled = True
torch.backends.cudnn.deterministic = True
# torch.backends.cudnn.benchmark = True
torch.set_num_threads(workers)
save_dir = (Path(save_dir) / "specifics" / "{:}-{:}-{:}-{:}".format(
"LESS" if use_less else "FULL",
model_str,
arch_config["channel"],
arch_config["num_cells"],
))
logger = Logger(str(save_dir), 0, False)
if model_str in CellArchitectures:
arch = CellArchitectures[model_str]
logger.log(
"The model string is found in pre-defined architecture dict : {:}".
format(model_str))
else:
try:
arch = CellStructure.str2structure(model_str)
except:
raise ValueError(
"Invalid model string : {:}. It can not be found or parsed.".
format(model_str))
assert arch.check_valid_op(get_search_spaces(
"cell", "full")), "{:} has the invalid op.".format(arch)
logger.log("Start train-evaluate {:}".format(arch.tostr()))
logger.log("arch_config : {:}".format(arch_config))
start_time, seed_time = time.time(), AverageMeter()
for _is, seed in enumerate(seeds):
logger.log(
"\nThe {:02d}/{:02d}-th seed is {:} ----------------------<.>----------------------"
.format(_is, len(seeds), seed))
to_save_name = save_dir / "seed-{:04d}.pth".format(seed)
if to_save_name.exists():
logger.log("Find the existing file {:}, directly load!".format(
to_save_name))
checkpoint = torch.load(to_save_name)
else:
logger.log(
"Does not find the existing file {:}, train and evaluate!".
format(to_save_name))
checkpoint = evaluate_all_datasets(
arch,
datasets,
xpaths,
splits,
use_less,
seed,
arch_config,
workers,
logger,
)
torch.save(checkpoint, to_save_name)
# log information
logger.log("{:}".format(checkpoint["info"]))
all_dataset_keys = checkpoint["all_dataset_keys"]
for dataset_key in all_dataset_keys:
logger.log("\n{:} dataset : {:} {:}".format(
"-" * 15, dataset_key, "-" * 15))
dataset_info = checkpoint[dataset_key]
# logger.log('Network ==>\n{:}'.format( dataset_info['net_string'] ))
logger.log("Flops = {:} MB, Params = {:} MB".format(
dataset_info["flop"], dataset_info["param"]))
logger.log("config : {:}".format(dataset_info["config"]))
logger.log("Training State (finish) = {:}".format(
dataset_info["finish-train"]))
last_epoch = dataset_info["total_epoch"] - 1
train_acc1es, train_acc5es = (
dataset_info["train_acc1es"],
dataset_info["train_acc5es"],
)
valid_acc1es, valid_acc5es = (
dataset_info["valid_acc1es"],
dataset_info["valid_acc5es"],
)
logger.log(
"Last Info : Train = Acc@1 {:.2f}% Acc@5 {:.2f}% Error@1 {:.2f}%, Test = Acc@1 {:.2f}% Acc@5 {:.2f}% Error@1 {:.2f}%"
.format(
train_acc1es[last_epoch],
train_acc5es[last_epoch],
100 - train_acc1es[last_epoch],
valid_acc1es[last_epoch],
valid_acc5es[last_epoch],
100 - valid_acc1es[last_epoch],
))
# measure elapsed time
seed_time.update(time.time() - start_time)
start_time = time.time()
need_time = "Time Left: {:}".format(
convert_secs2time(seed_time.avg * (len(seeds) - _is - 1), True))
logger.log(
"\n<<<***>>> The {:02d}/{:02d}-th seed is {:} <finish> other procedures need {:}"
.format(_is, len(seeds), seed, need_time))
logger.close()
def main(xargs, api):
torch.set_num_threads(4)
prepare_seed(xargs.rand_seed)
logger = prepare_logger(args)
logger.log("{:} use api : {:}".format(time_string(), api))
api.reset_time()
search_space = get_search_spaces(xargs.search_space, "nats-bench")
if xargs.search_space == "tss":
cs = get_topology_config_space(search_space)
config2structure = config2topology_func()
else:
cs = get_size_config_space(search_space)
config2structure = config2size_func(search_space)
hb_run_id = "0"
NS = hpns.NameServer(run_id=hb_run_id, host="localhost", port=0)
ns_host, ns_port = NS.start()
num_workers = 1
workers = []
for i in range(num_workers):
w = MyWorker(
nameserver=ns_host,
nameserver_port=ns_port,
convert_func=config2structure,
dataset=xargs.dataset,
api=api,
run_id=hb_run_id,
id=i,
)
w.run(background=True)
workers.append(w)
start_time = time.time()
bohb = BOHB(
configspace=cs,
run_id=hb_run_id,
eta=3,
min_budget=1,
max_budget=12,
nameserver=ns_host,
nameserver_port=ns_port,
num_samples=xargs.num_samples,
random_fraction=xargs.random_fraction,
bandwidth_factor=xargs.bandwidth_factor,
ping_interval=10,
min_bandwidth=xargs.min_bandwidth,
)
results = bohb.run(xargs.n_iters, min_n_workers=num_workers)
bohb.shutdown(shutdown_workers=True)
NS.shutdown()
# print('There are {:} runs.'.format(len(results.get_all_runs())))
# workers[0].total_times
# workers[0].trajectory
current_best_index = []
for idx in range(len(workers[0].trajectory)):
trajectory = workers[0].trajectory[:idx + 1]
arch = max(trajectory, key=lambda x: x[0])[1]
current_best_index.append(api.query_index_by_arch(arch))
best_arch = max(workers[0].trajectory, key=lambda x: x[0])[1]
logger.log("Best found configuration: {:} within {:.3f} s".format(
best_arch, workers[0].total_times[-1]))
info = api.query_info_str_by_arch(
best_arch, "200" if xargs.search_space == "tss" else "90")
logger.log("{:}".format(info))
logger.log("-" * 100)
logger.close()
return logger.log_dir, current_best_index, workers[0].total_times
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_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], "200")))
# 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], "200")))
logger.close()
def main(xargs, nas_bench):
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)
if xargs.dataset == "cifar10":
dataname = "cifar10-valid"
else:
dataname = xargs.dataset
if xargs.data_path is not None:
train_data, valid_data, xshape, class_num = get_datasets(
xargs.dataset, xargs.data_path, -1
)
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))
config_path = "configs/nas-benchmark/algos/R-EA.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
train_loader = torch.utils.data.DataLoader(
train_data,
batch_size=config.batch_size,
sampler=torch.utils.data.sampler.SubsetRandomSampler(train_split),
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} ||||||| 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))
extra_info = {
"config": config,
"train_loader": train_loader,
"valid_loader": valid_loader,
}
else:
config_path = "configs/nas-benchmark/algos/R-EA.config"
config = load_config(config_path, None, logger)
logger.log("||||||| {:10s} ||||||| Config={:}".format(xargs.dataset, config))
extra_info = {"config": config, "train_loader": None, "valid_loader": None}
# nas dataset load
assert xargs.arch_nas_dataset is not None and os.path.isfile(xargs.arch_nas_dataset)
search_space = get_search_spaces("cell", xargs.search_space_name)
cs = get_configuration_space(xargs.max_nodes, search_space)
config2structure = config2structure_func(xargs.max_nodes)
hb_run_id = "0"
NS = hpns.NameServer(run_id=hb_run_id, host="localhost", port=0)
ns_host, ns_port = NS.start()
num_workers = 1
# nas_bench = AANASBenchAPI(xargs.arch_nas_dataset)
# logger.log('{:} Create NAS-BENCH-API DONE'.format(time_string()))
workers = []
for i in range(num_workers):
w = MyWorker(
nameserver=ns_host,
nameserver_port=ns_port,
convert_func=config2structure,
dataname=dataname,
nas_bench=nas_bench,
time_budget=xargs.time_budget,
run_id=hb_run_id,
id=i,
)
w.run(background=True)
workers.append(w)
start_time = time.time()
bohb = BOHB(
configspace=cs,
run_id=hb_run_id,
eta=3,
min_budget=12,
max_budget=200,
nameserver=ns_host,
nameserver_port=ns_port,
num_samples=xargs.num_samples,
random_fraction=xargs.random_fraction,
bandwidth_factor=xargs.bandwidth_factor,
ping_interval=10,
min_bandwidth=xargs.min_bandwidth,
)
results = bohb.run(xargs.n_iters, min_n_workers=num_workers)
bohb.shutdown(shutdown_workers=True)
NS.shutdown()
real_cost_time = time.time() - start_time
id2config = results.get_id2config_mapping()
incumbent = results.get_incumbent_id()
logger.log(
"Best found configuration: {:} within {:.3f} s".format(
id2config[incumbent]["config"], real_cost_time
)
)
best_arch = config2structure(id2config[incumbent]["config"])
info = nas_bench.query_by_arch(best_arch, "200")
if info is None:
logger.log("Did not find this architecture : {:}.".format(best_arch))
else:
logger.log("{:}".format(info))
logger.log("-" * 100)
logger.log(
"workers : {:.1f}s with {:} archs".format(
workers[0].time_budget, len(workers[0].seen_archs)
)
)
logger.close()
return logger.log_dir, nas_bench.query_index_by_arch(best_arch), real_cost_time
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 = {}
w_optimizer = {}
a_optimizer = {}
w_scheduler = {}
a_scheduler = {}
valid_accuracies, genotypes = {}, {}
search_globle_model = get_cell_based_tiny_net(model_config).cuda()
for one in search_loader:
search_model[one] = get_cell_based_tiny_net(model_config).cuda()
search_model[one].load_state_dict(search_globle_model.state_dict())
w_optimizer[one], w_scheduler[one], criterion = get_optim_scheduler(search_model[one].parameters(), config)
if args.baseline == "dl":
w_optimizer[one] = dlOptimizer(search_model[one].get_weights(), xargs.arch_learning_rate, 0.1)
a_optimizer[one] = torch.optim.Adam(search_model[one].get_alphas(), lr=xargs.arch_learning_rate, betas=(0.5, 0.999), weight_decay=xargs.arch_weight_decay,)
valid_accuracies[one], genotypes[one] = (
{"best": -1},
{-1: search_model[one].genotype()},
)
criterion = criterion.cuda()
logger.log("search-model :\n{:}".format(search_globle_model))
logger.log("model-config : {:}".format(model_config))
# 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_globle_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"),
)
# 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 = 0
# start training
start_time, search_time, epoch_time, total_epoch = (
time.time(),
AverageMeter(),
AverageMeter(),
config.epochs + config.warmup,
)
local_epoch = args.local_epoch
for epoch in range(start_epoch, total_epoch):
for user in w_scheduler:
w_scheduler[user].update(epoch, 0.0)
search_model[user].set_tau(
xargs.tau_max - (xargs.tau_max - xargs.tau_min) * epoch / (total_epoch - 1)
)
# 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] {:}, tau={:}, LR={:}".format(
# epoch_str, need_time, search_model.get_tau(), min(w_scheduler.get_lr())
# )
# )
weight_list = []
acc_list = []
test_acc_list = []
for user in search_loader:
( search_w_loss,
search_w_top1,
search_w_top5,
valid_a_loss,
valid_a_top1,
valid_a_top5,
weight
) = search_func(
search_loader[user],
search_model[user],
search_globle_model,
criterion,
w_scheduler[user],
w_optimizer[user],
a_optimizer[user],
epoch_str,
xargs.print_freq,
logger,
local_epoch
)
logger.log(
"User {} : [{:}] searching : loss={:.2f}, accuracy@1={:.2f}%, accuracy@5={:.2f}%, time-cost={:.1f} s".format(
user, epoch_str, search_w_loss, search_w_top1, search_w_top5, search_time.sum
)
)
logger.log(
"User {} : [{:}] evaluate : loss={:.2f}, accuracy@1={:.2f}%, accuracy@5={:.2f}%".format(
user, epoch_str, valid_a_loss, valid_a_top1, valid_a_top5
)
)
weight_list.append(weight)
acc_list.append(valid_a_top1)
valid_accuracies[user][epoch] = valid_a_top1
genotypes[user][epoch] = search_model[user].genotype()
# loss, top1acc, top5acc = test_func(valid_loader[user], search_model[user], criterion)
# test_acc_list.append(top1acc)
# logger.log(
# "||||---|||| The {epoch:}-th epoch, user {user}, valid loss={loss:.3f}, valid_top1={top1:.2f}%, valid_top5={top5:.2f}%".format(
# epoch=epoch_str, user=user, loss=loss, top1=top1acc, top5=top5acc, )
# )
info_dict = {
"{}user_w_loss".format(user): search_w_loss,
"{}user_w_top1".format(user): search_w_top1,
"{}user_w_top5".format(user): search_w_top5,
"{}user_a_loss".format(user): valid_a_loss,
"{}user_a_top1".format(user): valid_a_top1,
"{}user_a_top5".format(user): valid_a_top5,
# "{}user_test_loss".format(user): search_w_loss,
# "{}user_test_top1".format(user): search_w_loss,
# "{}user_test_top5".format(user): search_w_loss,
}
wandb.log(info_dict)
info_dict = {
"epoch": epoch,
"average_valid_acc": np.average(acc_list),
"average_test_acc": np.average(test_acc_list)
}
wandb.log(info_dict)
arch_personalize = args.personalize_arch
weight_average, arch_normal_list, arch_reduce_list = average_weights(weight_list, arch_personalize)
for user in search_model:
if arch_personalize:
tep = copy.deepcopy(weight_average)
tep['arch_normal_parameters'] = arch_normal_list[user]
tep['arch_reduce_parameters'] = arch_reduce_list[user]
search_model[user].load_state_dict(tep)
else:
search_model[user].load_state_dict(weight_average)
logger.log(
"<<<--->>> The {:}-th epoch : {:}".format(epoch_str, search_model[user].genotype())
)
search_globle_model.load_state_dict(weight_average)
search_time.update(time.time() - start_time)
# check the best accuracy
# 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
# 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_globle_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()
# save checkpoint
for user in search_model:
model_base_path = logger.model_dir / "User{:}-acc-{}-basic-seed-{:}.pth".format(user, valid_accuracies[user][epoch],args.rand_seed)
save_path = save_checkpoint(
{
"epoch": epoch + 1,
"args": deepcopy(xargs),
"search_model": search_model[user].state_dict(),
"w_optimizer": w_optimizer[user].state_dict(),
"a_optimizer": a_optimizer[user].state_dict(),
"w_scheduler": w_scheduler[user].state_dict(),
"genotypes": search_model[user].genotype(),
"valid_accuracies": valid_accuracies[user],
},
model_base_path,
logger,
)
# 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()