def load_net_from_checkpoint(checkpoint):
assert osp.isfile(checkpoint), "checkpoint {:} does not exist".format(
checkpoint)
checkpoint = torch.load(checkpoint)
model_config = dict2config(checkpoint["model-config"], None)
model = obtain_model(model_config)
model.load_state_dict(checkpoint["base-model"])
return model
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):
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(args):
assert os.path.isdir(args.data_path), "invalid data-path : {:}".format(
args.data_path)
assert os.path.isfile(args.checkpoint), "invalid checkpoint : {:}".format(
args.checkpoint)
checkpoint = torch.load(args.checkpoint)
xargs = checkpoint["args"]
train_data, valid_data, xshape, class_num = get_datasets(
xargs.dataset, args.data_path, xargs.cutout_length)
valid_loader = torch.utils.data.DataLoader(
valid_data,
batch_size=xargs.batch_size,
shuffle=False,
num_workers=xargs.workers,
pin_memory=True,
)
logger = PrintLogger()
model_config = dict2config(checkpoint["model-config"], logger)
base_model = obtain_model(model_config)
flop, param = get_model_infos(base_model, xshape)
logger.log("model ====>>>>:\n{:}".format(base_model))
logger.log("model information : {:}".format(base_model.get_message()))
logger.log("-" * 50)
logger.log("Params={:.2f} MB, FLOPs={:.2f} M ... = {:.2f} G".format(
param, flop, flop / 1e3))
logger.log("-" * 50)
logger.log("valid_data : {:}".format(valid_data))
optim_config = dict2config(checkpoint["optim-config"], logger)
_, _, criterion = get_optim_scheduler(base_model.parameters(),
optim_config)
logger.log("criterion : {:}".format(criterion))
base_model.load_state_dict(checkpoint["base-model"])
_, valid_func = get_procedures(xargs.procedure)
logger.log(
"initialize the CNN done, evaluate it using {:}".format(valid_func))
network = torch.nn.DataParallel(base_model).cuda()
try:
valid_loss, valid_acc1, valid_acc5 = valid_func(
valid_loader,
network,
criterion,
optim_config,
"pure-evaluation",
xargs.print_freq_eval,
logger,
)
except:
_, valid_func = get_procedures("basic")
valid_loss, valid_acc1, valid_acc5 = valid_func(
valid_loader,
network,
criterion,
optim_config,
"pure-evaluation",
xargs.print_freq_eval,
logger,
)
num_bytes = torch.cuda.max_memory_cached(
next(network.parameters()).device) * 1.0
logger.log(
"***{:s}*** EVALUATION loss = {:.6f}, accuracy@1 = {:.2f}, accuracy@5 = {:.2f}, error@1 = {:.2f}, error@5 = {:.2f}"
.format(
time_string(),
valid_loss,
valid_acc1,
valid_acc5,
100 - valid_acc1,
100 - valid_acc5,
))
logger.log(
"[GPU-Memory-Usage on {:} is {:} bytes, {:.2f} KB, {:.2f} MB, {:.2f} GB.]"
.format(
next(network.parameters()).device,
int(num_bytes),
num_bytes / 1e3,
num_bytes / 1e6,
num_bytes / 1e9,
))
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
)
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 get_cell_based_tiny_net(config):
if isinstance(config, dict):
config = dict2config(config,
None) # to support the argument being a dict
super_type = getattr(config, "super_type", "basic")
group_names = [
"DARTS-V1", "DARTS-V2", "GDAS", "SETN", "ENAS", "RANDOM", "generic"
]
if super_type == "basic" and config.name in group_names:
from .cell_searchs import nas201_super_nets as nas_super_nets
try:
return nas_super_nets[config.name](
config.C,
config.N,
config.max_nodes,
config.num_classes,
config.space,
config.affine,
config.track_running_stats,
)
except:
return nas_super_nets[config.name](config.C, config.N,
config.max_nodes,
config.num_classes,
config.space)
elif super_type == "search-shape":
from .shape_searchs import GenericNAS301Model
genotype = CellStructure.str2structure(config.genotype)
return GenericNAS301Model(
config.candidate_Cs,
config.max_num_Cs,
genotype,
config.num_classes,
config.affine,
config.track_running_stats,
)
elif super_type == "nasnet-super":
from .cell_searchs import nasnet_super_nets as nas_super_nets
return nas_super_nets[config.name](
config.C,
config.N,
config.steps,
config.multiplier,
config.stem_multiplier,
config.num_classes,
config.space,
config.affine,
config.track_running_stats,
)
elif config.name == "infer.tiny":
from .cell_infers import TinyNetwork
if hasattr(config, "genotype"):
genotype = config.genotype
elif hasattr(config, "arch_str"):
genotype = CellStructure.str2structure(config.arch_str)
else:
raise ValueError(
"Can not find genotype from this config : {:}".format(config))
return TinyNetwork(config.C, config.N, genotype, config.num_classes)
elif config.name == "infer.shape.tiny":
from .shape_infers import DynamicShapeTinyNet
if isinstance(config.channels, str):
channels = tuple([int(x) for x in config.channels.split(":")])
else:
channels = config.channels
genotype = CellStructure.str2structure(config.genotype)
return DynamicShapeTinyNet(channels, genotype, config.num_classes)
elif config.name == "infer.nasnet-cifar":
from .cell_infers import NASNetonCIFAR
raise NotImplementedError
else:
raise ValueError("invalid network name : {:}".format(config.name))
def evaluate_all_datasets(
arch: Text,
datasets: List[Text],
xpaths: List[Text],
splits: List[Text],
config_path: Text,
seed: int,
raw_arch_config,
workers,
logger,
):
machine_info, raw_arch_config = get_machine_info(), deepcopy(
raw_arch_config)
all_infos = {"info": machine_info}
all_dataset_keys = []
# look all the datasets
for dataset, xpath, split in zip(datasets, xpaths, splits):
# train valid data
train_data, valid_data, xshape, class_num = get_datasets(
dataset, xpath, -1)
# load the configuration
if dataset == "cifar10" or dataset == "cifar100":
split_info = load_config("configs/nas-benchmark/cifar-split.txt",
None, None)
elif dataset.startswith("ImageNet16"):
split_info = load_config(
"configs/nas-benchmark/{:}-split.txt".format(dataset), None,
None)
else:
raise ValueError("invalid dataset : {:}".format(dataset))
config = load_config(config_path,
dict(class_num=class_num, xshape=xshape), logger)
# check whether use splited validation set
if bool(split):
assert dataset == "cifar10"
ValLoaders = {
"ori-test":
torch.utils.data.DataLoader(
valid_data,
batch_size=config.batch_size,
shuffle=False,
num_workers=workers,
pin_memory=True,
)
}
assert len(train_data) == len(split_info.train) + len(
split_info.valid), "invalid length : {:} vs {:} + {:}".format(
len(train_data), len(split_info.train),
len(split_info.valid))
train_data_v2 = deepcopy(train_data)
train_data_v2.transform = valid_data.transform
valid_data = train_data_v2
# data loader
train_loader = torch.utils.data.DataLoader(
train_data,
batch_size=config.batch_size,
sampler=torch.utils.data.sampler.SubsetRandomSampler(
split_info.train),
num_workers=workers,
pin_memory=True,
)
valid_loader = torch.utils.data.DataLoader(
valid_data,
batch_size=config.batch_size,
sampler=torch.utils.data.sampler.SubsetRandomSampler(
split_info.valid),
num_workers=workers,
pin_memory=True,
)
ValLoaders["x-valid"] = valid_loader
else:
# data loader
train_loader = torch.utils.data.DataLoader(
train_data,
batch_size=config.batch_size,
shuffle=True,
num_workers=workers,
pin_memory=True,
)
valid_loader = torch.utils.data.DataLoader(
valid_data,
batch_size=config.batch_size,
shuffle=False,
num_workers=workers,
pin_memory=True,
)
if dataset == "cifar10":
ValLoaders = {"ori-test": valid_loader}
elif dataset == "cifar100":
cifar100_splits = load_config(
"configs/nas-benchmark/cifar100-test-split.txt", None,
None)
ValLoaders = {
"ori-test":
valid_loader,
"x-valid":
torch.utils.data.DataLoader(
valid_data,
batch_size=config.batch_size,
sampler=torch.utils.data.sampler.SubsetRandomSampler(
cifar100_splits.xvalid),
num_workers=workers,
pin_memory=True,
),
"x-test":
torch.utils.data.DataLoader(
valid_data,
batch_size=config.batch_size,
sampler=torch.utils.data.sampler.SubsetRandomSampler(
cifar100_splits.xtest),
num_workers=workers,
pin_memory=True,
),
}
elif dataset == "ImageNet16-120":
imagenet16_splits = load_config(
"configs/nas-benchmark/imagenet-16-120-test-split.txt",
None, None)
ValLoaders = {
"ori-test":
valid_loader,
"x-valid":
torch.utils.data.DataLoader(
valid_data,
batch_size=config.batch_size,
sampler=torch.utils.data.sampler.SubsetRandomSampler(
imagenet16_splits.xvalid),
num_workers=workers,
pin_memory=True,
),
"x-test":
torch.utils.data.DataLoader(
valid_data,
batch_size=config.batch_size,
sampler=torch.utils.data.sampler.SubsetRandomSampler(
imagenet16_splits.xtest),
num_workers=workers,
pin_memory=True,
),
}
else:
raise ValueError("invalid dataset : {:}".format(dataset))
dataset_key = "{:}".format(dataset)
if bool(split):
dataset_key = dataset_key + "-valid"
logger.log(
"Evaluate ||||||| {:10s} ||||||| Train-Num={:}, Valid-Num={:}, Train-Loader-Num={:}, Valid-Loader-Num={:}, batch size={:}"
.format(
dataset_key,
len(train_data),
len(valid_data),
len(train_loader),
len(valid_loader),
config.batch_size,
))
logger.log("Evaluate ||||||| {:10s} ||||||| Config={:}".format(
dataset_key, config))
for key, value in ValLoaders.items():
logger.log("Evaluate ---->>>> {:10s} with {:} batchs".format(
key, len(value)))
arch_config = dict2config(
dict(
name="infer.tiny",
C=raw_arch_config["channel"],
N=raw_arch_config["num_cells"],
genotype=arch,
num_classes=config.class_num,
),
None,
)
results = bench_evaluate_for_seed(arch_config, config, train_loader,
ValLoaders, seed, logger)
all_infos[dataset_key] = results
all_dataset_keys.append(dataset_key)
all_infos["all_dataset_keys"] = all_dataset_keys
return all_infos
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 create_result_count(used_seed, dataset, arch_config, results,
dataloader_dict):
xresult = ResultsCount(
dataset,
results["net_state_dict"],
results["train_acc1es"],
results["train_losses"],
results["param"],
results["flop"],
arch_config,
used_seed,
results["total_epoch"],
None,
)
net_config = dict2config(
{
"name": "infer.tiny",
"C": arch_config["channel"],
"N": arch_config["num_cells"],
"genotype": CellStructure.str2structure(arch_config["arch_str"]),
"num_classes": arch_config["class_num"],
},
None,
)
network = get_cell_based_tiny_net(net_config)
network.load_state_dict(xresult.get_net_param())
if "train_times" in results: # new version
xresult.update_train_info(
results["train_acc1es"],
results["train_acc5es"],
results["train_losses"],
results["train_times"],
)
xresult.update_eval(results["valid_acc1es"], results["valid_losses"],
results["valid_times"])
else:
if dataset == "cifar10-valid":
xresult.update_OLD_eval("x-valid", results["valid_acc1es"],
results["valid_losses"])
loss, top1, top5, latencies = pure_evaluate(
dataloader_dict["{:}@{:}".format("cifar10", "test")],
network.cuda())
xresult.update_OLD_eval(
"ori-test",
{results["total_epoch"] - 1: top1},
{results["total_epoch"] - 1: loss},
)
xresult.update_latency(latencies)
elif dataset == "cifar10":
xresult.update_OLD_eval("ori-test", results["valid_acc1es"],
results["valid_losses"])
loss, top1, top5, latencies = pure_evaluate(
dataloader_dict["{:}@{:}".format(dataset, "test")],
network.cuda())
xresult.update_latency(latencies)
elif dataset == "cifar100" or dataset == "ImageNet16-120":
xresult.update_OLD_eval("ori-test", results["valid_acc1es"],
results["valid_losses"])
loss, top1, top5, latencies = pure_evaluate(
dataloader_dict["{:}@{:}".format(dataset, "valid")],
network.cuda())
xresult.update_OLD_eval(
"x-valid",
{results["total_epoch"] - 1: top1},
{results["total_epoch"] - 1: loss},
)
loss, top1, top5, latencies = pure_evaluate(
dataloader_dict["{:}@{:}".format(dataset, "test")],
network.cuda())
xresult.update_OLD_eval(
"x-test",
{results["total_epoch"] - 1: top1},
{results["total_epoch"] - 1: loss},
)
xresult.update_latency(latencies)
else:
raise ValueError("invalid dataset name : {:}".format(dataset))
return xresult
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()