def main():
# データセットの取得
datasets = get_datasets(DATASETS_CSV,
test_size=TEST_SIZE,
image_size=IMAGE_SIZE)
train_data, _, _, _ = datasets
# TODO
# 29 0.9961215257644653
# 30回目で過学習により5割減
# 30回に分けてトレーニング実行
max_steps = 25
run_training(
datasets,
tensorboard_path=TENSORBOARD_PATH,
checkpoint_path=MODEL_PATH,
# 分類数
num_classes=len(get_labels(DATASETS_CSV)),
# 画像サイズ
image_size=IMAGE_SIZE,
# ピクセルのベクトル数 3=カラー,1=モノクロ
channel=CHANNELS,
# 学習実行回数
max_steps=max_steps,
# 1エポックで学習するデータサイズ
batch_size=int(len(train_data) / max_steps),
# 学習率
learning_rate=1e-4)
def main(mode, speaker, modality, epochs, batch_size, lr, step_size, gamma):
dataloaders, dataset_sizes = get_datasets(speaker,
modality=modality,
batch_size=batch_size)
model = define_model(modality, speaker)
model = model.to(device)
optimizer = optim.Adam(model.parameters(), lr=lr)
exp_lr_scheduler = lr_scheduler.StepLR(optimizer,
step_size=step_size,
gamma=gamma)
if mode == 'train':
model = train_model(model,
optimizer,
exp_lr_scheduler,
device,
dataloaders,
dataset_sizes,
speaker,
num_epochs=epochs,
modality=modality)
torch.save(model.state_dict(),
'data/' + speaker + '/' + modality + '/model')
else:
print('test time!')
# features_list, labels_list, preds_list = predict_model(model, device, dataloaders, dataset_sizes)
pass
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 test_one_shot_model(ckpath, use_train):
from models import get_cell_based_tiny_net, get_search_spaces
from datasets import get_datasets, SearchDataset
from config_utils import load_config, dict2config
from utils.nas_utils import evaluate_one_shot
use_train = int(use_train) > 0
#ckpath = 'output/search-cell-nas-bench-201/DARTS-V1-cifar10/checkpoint/seed-11416-basic.pth'
#ckpath = 'output/search-cell-nas-bench-201/DARTS-V1-cifar10/checkpoint/seed-28640-basic.pth'
print ('ckpath : {:}'.format(ckpath))
ckp = torch.load(ckpath)
xargs = ckp['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}, None)
config = load_config('./configs/nas-benchmark/algos/DARTS.config', {'class_num': class_num, 'xshape': xshape}, None)
if xargs.dataset == 'cifar10':
cifar_split = load_config('configs/nas-benchmark/cifar-split.txt', None, None)
xvalid_data = deepcopy(train_data)
xvalid_data.transform = valid_data.transform
valid_loader= torch.utils.data.DataLoader(xvalid_data, batch_size=2048, sampler=torch.utils.data.sampler.SubsetRandomSampler(cifar_split.valid), num_workers=12, pin_memory=True)
else: raise ValueError('invalid dataset : {:}'.format(xargs.dataseet))
search_space = get_search_spaces('cell', xargs.search_space_name)
model_config = dict2config({'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': True}, None)
search_model = get_cell_based_tiny_net(model_config)
search_model.load_state_dict( ckp['search_model'] )
search_model = search_model.cuda()
api = API('/home/dxy/.torch/NAS-Bench-201-v1_0-e61699.pth')
archs, probs, accuracies = evaluate_one_shot(search_model, valid_loader, api, use_train)
def test_one_shot_model(ckpath, use_train):
from models import get_cell_based_tiny_net, get_search_spaces
from datasets import get_datasets, SearchDataset
from config_utils import load_config, dict2config
from utils.nas_utils import evaluate_one_shot
use_train = int(use_train) > 0
# ckpath = 'output/search-cell-nas-bench-201/DARTS-V1-cifar10/checkpoint/seed-11416-basic.pth'
# ckpath = 'output/search-cell-nas-bench-201/DARTS-V1-cifar10/checkpoint/seed-28640-basic.pth'
print("ckpath : {:}".format(ckpath))
ckp = torch.load(ckpath)
xargs = ckp["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}, None)
config = load_config(
"./configs/nas-benchmark/algos/DARTS.config",
{
"class_num": class_num,
"xshape": xshape
},
None,
)
if xargs.dataset == "cifar10":
cifar_split = load_config("configs/nas-benchmark/cifar-split.txt",
None, None)
xvalid_data = deepcopy(train_data)
xvalid_data.transform = valid_data.transform
valid_loader = torch.utils.data.DataLoader(
xvalid_data,
batch_size=2048,
sampler=torch.utils.data.sampler.SubsetRandomSampler(
cifar_split.valid),
num_workers=12,
pin_memory=True,
)
else:
raise ValueError("invalid dataset : {:}".format(xargs.dataseet))
search_space = get_search_spaces("cell", xargs.search_space_name)
model_config = dict2config(
{
"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": True,
},
None,
)
search_model = get_cell_based_tiny_net(model_config)
search_model.load_state_dict(ckp["search_model"])
search_model = search_model.cuda()
api = API("/home/dxy/.torch/NAS-Bench-201-v1_0-e61699.pth")
archs, probs, accuracies = evaluate_one_shot(search_model, valid_loader,
api, use_train)
def get_train_loader(args):
tr_dataset, _ = get_datasets(args)
loader = torch.utils.data.DataLoader(dataset=tr_dataset,
batch_size=args.batch_size,
shuffle=False,
num_workers=0,
pin_memory=True,
drop_last=False)
return loader
def test_get_datasets(tmp_path):
dataset, _ = get_datasets("cifar10", tmp_path)
assert isinstance(dataset, Dataset)
batch = next(iter(dataset))
assert isinstance(batch, Iterable)
assert isinstance(batch[0], Tensor)
assert isinstance(batch[1], Number)
assert batch[0].ndim == 3
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 )
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 get_test_loader(args):
_, te_dataset = get_datasets(args)
if args.resume_dataset_mean is not None and args.resume_dataset_std is not None:
mean = np.load(args.resume_dataset_mean)
std = np.load(args.resume_dataset_std)
te_dataset.renormalize(mean, std)
loader = torch.utils.data.DataLoader(
dataset=te_dataset, batch_size=args.batch_size, shuffle=False,
num_workers=0, pin_memory=True, drop_last=False)
return loader
def show_imagenet_16_120(dataset_dir=None):
if dataset_dir is None:
torch_home_dir = os.environ['TORCH_HOME'] if 'TORCH_HOME' in os.environ else os.path.join(os.environ['HOME'], '.torch')
dataset_dir = os.path.join(torch_home_dir, 'cifar.python', 'ImageNet16')
train_data, valid_data, xshape, class_num = get_datasets('ImageNet16-120', dataset_dir, -1)
split_info = load_config('configs/nas-benchmark/ImageNet16-120-split.txt', None, None)
print('=' * 10 + ' ImageNet-16-120 ' + '=' * 10)
print('Training Data: {:}'.format(train_data))
print('Evaluation Data: {:}'.format(valid_data))
print('Hold-out training: {:} images.'.format(len(split_info.train)))
print('Hold-out valid : {:} images.'.format(len(split_info.valid)))
def run(args):
device = 'cuda' if torch.cuda.is_available() else 'cpu'
_, _, test = get_datasets(args.data)
test_loader = DataLoader(test, batch_size=2, num_workers=0)
model_factory = {
'fcn-resnet50': lambda: torchvision.models.segmentation.fcn_resnet50(num_classes=NUM_CLASSES,
pretrained=False),
'fcn-resnet101': lambda: torchvision.models.segmentation.fcn_resnet101(num_classes=NUM_CLASSES,
pretrained=False),
'deeplab-resnet50': lambda: torchvision.models.segmentation.deeplabv3_resnet50(num_classes=NUM_CLASSES,
pretrained=False),
'deeplab-resnet101': lambda: torchvision.models.segmentation.deeplabv3_resnet101(num_classes=NUM_CLASSES,
pretrained=False)
}
model = model_factory[args.model]()
# model.load_state_dict(torch.load(args.weights))
model.to(device)
print(colored("Evaluating", "white"))
images, gt_seg, pd_seg = [], [], []
for i, batch in enumerate(test_loader):
image, seg = prepare_batch(batch, device=device, non_blocking=True)
image = image.to(device)
predicted_seg = model(image)
images.append(image.detach().cpu())
gt_seg.append(seg.detach().cpu())
pd_seg.append(predicted_seg['out'].detach().cpu())
images = torch.cat(images, dim=0)
gt_seg = torch.cat(gt_seg, dim=0)
pd_seg = torch.cat(pd_seg, dim=0)
print(colored("Plotting", "white"))
fig, axs = plt.subplots(ncols=3, nrows=4, figsize=(14, 14))
axs[0][0].set_title("Image")
axs[0][1].set_title("Expected segmentation")
axs[0][2].set_title("Predicted segmentation")
for i in range(images.shape[0]):
# plot image
axs[i][0].imshow(images[i].permute(1, 2, 0).cpu())
axs[i][0].axis('off')
# plot gt seg
axs[i][1].imshow(gt_seg[i].permute(1, 2, 0).cpu(), cmap='gray')
axs[i][1].axis('off')
# plot pd seg
axs[i][2].imshow(pd_seg[i].permute(1, 2, 0).cpu(), cmap='gray')
axs[i][2].axis('off')
plt.tight_layout()
plt.savefig("output.svg")
plt.show()
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)
assert xargs.dataset == 'cifar10', 'currently only support CIFAR-10'
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}
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)
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)
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.'.format(time_string(), best_arch, best_acc, len(history), total_time_cost))
info = nas_bench.query_by_arch( best_arch )
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 make_dataset(config: dict, train_transform: object = None, val_transform: object = None, mode='train'):
''' make train, val or test datasets '''
datasets = get_datasets(config)
train_data = datasets[config.dataset + '_train'](transform=train_transform)
val_data = datasets[config.dataset + '_val'](transform=val_transform)
if config.test_dataset.type == 'LCC_FASD' and config.dataset == 'celeba_spoof':
test_data = datasets['LCC_FASD_combined'](transform=val_transform)
else:
test_data = datasets[config.test_dataset.type + '_test'](transform=val_transform)
if mode == 'train':
return train_data, val_data, test_data
else:
assert mode == 'eval'
return test_data
def test_get_datasets(tmp_path):
train_ds, eval_ds = get_datasets(tmp_path)
assert isinstance(train_ds, Dataset)
assert isinstance(eval_ds, Dataset)
train_batch = next(iter(train_ds))
assert isinstance(train_batch, Iterable)
assert isinstance(train_batch[0], Tensor)
assert isinstance(train_batch[1], Number)
assert train_batch[0].ndim == 3
eval_batch = next(iter(eval_ds))
assert isinstance(eval_batch, Iterable)
assert isinstance(eval_batch[0], Tensor)
assert isinstance(eval_batch[1], Number)
assert eval_batch[0].ndim == 3
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': 'SPOS', '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)
logger.log('search space : {:}'.format(search_space))
model = get_cell_based_tiny_net(model_config)
flop, param = get_model_infos(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))
checkpoint_path_template = '{}/checkpoint/seed-{}_epoch-{}.pth'
logger.log("=> loading checkpoint from {}".format(checkpoint_path_template.format(args.save_dir, args.rand_seed, 0)))
load(checkpoint_path_template.format(args.save_dir, args.rand_seed, 0), model)
init_model = deepcopy(model)
angles = []
for epoch in range(xargs.epochs):
genotype = load(checkpoint_path_template.format(args.save_dir, args.rand_seed, epoch), model)
logger.log("=> loading checkpoint from {}".format(checkpoint_path_template.format(args.dataset, args.rand_seed, epoch)))
cur_model = deepcopy(model)
angle = get_arch_angle(init_model, cur_model, genotype, search_space)
logger.log('[{:}] cal angle : angle={}'.format(epoch, angle))
angle = round(angle,2)
angles.append(angle)
print(angles)
def show_imagenet_16_120(dataset_dir=None):
if dataset_dir is None:
torch_home_dir = (os.environ["TORCH_HOME"]
if "TORCH_HOME" in os.environ else os.path.join(
os.environ["HOME"], ".torch"))
dataset_dir = os.path.join(torch_home_dir, "cifar.python",
"ImageNet16")
train_data, valid_data, xshape, class_num = get_datasets(
"ImageNet16-120", dataset_dir, -1)
split_info = load_config("configs/nas-benchmark/ImageNet16-120-split.txt",
None, None)
print("=" * 10 + " ImageNet-16-120 " + "=" * 10)
print("Training Data: {:}".format(train_data))
print("Evaluation Data: {:}".format(valid_data))
print("Hold-out training: {:} images.".format(len(split_info.train)))
print("Hold-out valid : {:} images.".format(len(split_info.valid)))
def main(args):
model = PointFlow(args)
def _transform_(m):
return nn.DataParallel(m)
model = model.cuda()
model.multi_gpu_wrapper(_transform_)
print("Resume Path:%s" % args.resume_checkpoint)
checkpoint = torch.load(args.resume_checkpoint)
model.load_state_dict(checkpoint)
model.eval()
_, te_dataset = get_datasets(args)
if args.resume_dataset_mean is not None and args.resume_dataset_std is not None:
mean = np.load(args.resume_dataset_mean)
std = np.load(args.resume_dataset_std)
te_dataset.renormalize(mean, std)
ds_mean = torch.from_numpy(te_dataset.all_points_mean).cuda()
ds_std = torch.from_numpy(te_dataset.all_points_std).cuda()
all_sample = []
with torch.no_grad():
for i in range(0, args.num_sample_shapes, args.batch_size):
B = len(range(i, min(i + args.batch_size, args.num_sample_shapes)))
N = args.num_sample_points
_, out_pc = model.sample(B, N)
out_pc = out_pc * ds_std + ds_mean
all_sample.append(out_pc)
sample_pcs = torch.cat(all_sample, dim=0).cpu().detach().numpy()
print("Generation sample size:(%s, %s, %s)" % sample_pcs.shape)
# Save the generative output
os.makedirs("demo", exist_ok=True)
np.save(os.path.join("demo", "model_out_smp.npy"), sample_pcs)
# Visualize the demo
pcl = o3d.geometry.PointCloud()
for i in range(int(sample_pcs.shape[0])):
print("Visualizing: %03d/%03d" % (i, sample_pcs.shape[0]))
pts = sample_pcs[i].reshape(-1, 3)
pcl.points = o3d.utility.Vector3dVector(pts)
o3d.visualization.draw_geometries([pcl])
def evaluate_all_datasets(arch, datasets, xpaths, splits, seed, arch_config, workers, logger):
machine_info, arch_config = get_machine_info(), deepcopy(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 configurature
if dataset == 'cifar10' or dataset == 'cifar100':
config_path = 'configs/nas-benchmark/CIFAR.config'
split_info = load_config('configs/nas-benchmark/cifar-split.txt', None, None)
elif dataset.startswith('ImageNet16'):
config_path = 'configs/nas-benchmark/ImageNet-16.config'
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, \
{'class_num': class_num,
'xshape' : xshape}, \
logger)
# check whether use splited validation set
if bool(split):
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)
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)
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))
results = evaluate_for_seed(arch_config, config, arch, train_loader, valid_loader, 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 run(args):
device = 'cuda' if torch.cuda.is_available() else 'cpu'
print(colored("Using device: ", "white") + colored(device, "green"))
print(colored("Initializing test dataset...", color="white"))
_, _, test_dataset = get_datasets(args.data)
test_loader = DataLoader(test_dataset, batch_size=args.batch_size, shuffle=True)
model_factory = {
'fcn-resnet50': lambda: torchvision.models.segmentation.fcn_resnet50(num_classes=NUM_CLASSES,
pretrained=False),
'fcn-resnet101': lambda: torchvision.models.segmentation.fcn_resnet101(num_classes=NUM_CLASSES,
pretrained=False),
'deeplab-resnet50': lambda: torchvision.models.segmentation.deeplabv3_resnet50(num_classes=NUM_CLASSES,
pretrained=False),
'deeplab-resnet101': lambda: torchvision.models.segmentation.deeplabv3_resnet101(num_classes=NUM_CLASSES,
pretrained=False)
}
model = model_factory[args.model]()
model.load_state_dict(torch.load(args.weights))
model.to(device)
cm_metric = ConfusionMatrix(num_classes=NUM_CLASSES, output_transform=output_transform_seg)
metrics = {'dice': MetricsLambda(lambda x: torch.mean(x).item(), DiceCoefficient(cm_metric)),
'iou': MetricsLambda(lambda x: torch.mean(x).item(), IoU(cm_metric)),
'dice_background': MetricsLambda(lambda x: x[0].item(), DiceCoefficient(cm_metric)),
'dice_head': MetricsLambda(lambda x: x[1].item(), DiceCoefficient(cm_metric)),
'dice_mid': MetricsLambda(lambda x: x[2].item(), DiceCoefficient(cm_metric)),
'dice_tail': MetricsLambda(lambda x: x[3].item(), DiceCoefficient(cm_metric)),
'iou_background': MetricsLambda(lambda x: x[0].item(), IoU(cm_metric)),
'iou_head': MetricsLambda(lambda x: x[1].item(), IoU(cm_metric)),
'iou_mid': MetricsLambda(lambda x: x[2].item(), IoU(cm_metric)),
'iou_tail': MetricsLambda(lambda x: x[3].item(), IoU(cm_metric))
}
print(colored("Evaluating...\n", color="white"))
test_evaluator = create_supervised_evaluator(model, metrics=metrics, device=device, prepare_batch=prepare_batch)
@test_evaluator.on(Events.COMPLETED)
def log_training_loss(engine):
for k, v in engine.state.metrics.items():
print(f"{k}: {v:.4f}")
test_evaluator.run(test_loader)
def get_test_loader(args):
train_dataset, val_dataset, _, _ = get_datasets(args)
if args.resume_dataset_mean is not None and args.resume_dataset_std is not None:
mean = np.load(args.resume_dataset_mean)
std = np.load(args.resume_dataset_std)
val_dataset.renormalize(mean, std)
train_loader = torch.utils.data.DataLoader(dataset=train_dataset,
batch_size=args.batch_size,
shuffle=True,
collate_fn=collate_fn,
num_workers=0,
pin_memory=True,
drop_last=True)
val_loader = torch.utils.data.DataLoader(dataset=val_dataset,
batch_size=args.batch_size,
shuffle=False,
collate_fn=collate_fn,
num_workers=0,
pin_memory=True,
drop_last=False)
return train_loader, val_loader
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(args):
assert torch.cuda.is_available(), "CUDA is not available."
torch.backends.cudnn.enabled = True
torch.backends.cudnn.benchmark = True
# torch.backends.cudnn.deterministic = True
torch.set_num_threads(args.workers)
prepare_seed(args.rand_seed)
logger = prepare_logger(args)
# prepare dataset
train_data, valid_data, xshape, class_num = get_datasets(
args.dataset, args.data_path, args.cutout_length)
# train_loader = torch.utils.data.DataLoader(train_data, batch_size=args.batch_size, shuffle=True , num_workers=args.workers, pin_memory=True)
valid_loader = torch.utils.data.DataLoader(
valid_data,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.workers,
pin_memory=True,
)
split_file_path = Path(args.split_path)
assert split_file_path.exists(), "{:} does not exist".format(
split_file_path)
split_info = torch.load(split_file_path)
train_split, valid_split = split_info["train"], split_info["valid"]
assert (len(set(train_split).intersection(set(valid_split))) == 0
), "There should be 0 element that belongs to both train and valid"
assert len(train_split) + len(valid_split) == len(
train_data), "{:} + {:} vs {:}".format(len(train_split),
len(valid_split),
len(train_data))
search_dataset = SearchDataset(args.dataset, train_data, train_split,
valid_split)
search_train_loader = torch.utils.data.DataLoader(
train_data,
batch_size=args.batch_size,
sampler=torch.utils.data.sampler.SubsetRandomSampler(train_split),
pin_memory=True,
num_workers=args.workers,
)
search_valid_loader = torch.utils.data.DataLoader(
train_data,
batch_size=args.batch_size,
sampler=torch.utils.data.sampler.SubsetRandomSampler(valid_split),
pin_memory=True,
num_workers=args.workers,
)
search_loader = torch.utils.data.DataLoader(
search_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.workers,
pin_memory=True,
sampler=None,
)
# get configures
model_config = load_config(
args.model_config,
{
"class_num": class_num,
"search_mode": args.search_shape
},
logger,
)
# obtain the model
search_model = obtain_search_model(model_config)
MAX_FLOP, param = get_model_infos(search_model, xshape)
optim_config = load_config(args.optim_config, {
"class_num": class_num,
"FLOP": MAX_FLOP
}, logger)
logger.log("Model Information : {:}".format(search_model.get_message()))
logger.log("MAX_FLOP = {:} M".format(MAX_FLOP))
logger.log("Params = {:} M".format(param))
logger.log("train_data : {:}".format(train_data))
logger.log("search-data: {:}".format(search_dataset))
logger.log("search_train_loader : {:} samples".format(len(train_split)))
logger.log("search_valid_loader : {:} samples".format(len(valid_split)))
base_optimizer, scheduler, criterion = get_optim_scheduler(
search_model.base_parameters(), optim_config)
arch_optimizer = torch.optim.Adam(
search_model.arch_parameters(),
lr=optim_config.arch_LR,
betas=(0.5, 0.999),
weight_decay=optim_config.arch_decay,
)
logger.log("base-optimizer : {:}".format(base_optimizer))
logger.log("arch-optimizer : {:}".format(arch_optimizer))
logger.log("scheduler : {:}".format(scheduler))
logger.log("criterion : {:}".format(criterion))
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()
# load checkpoint
if last_info.exists() or (args.resume is not None and osp.isfile(
args.resume)): # automatically resume from previous checkpoint
if args.resume is not None and osp.isfile(args.resume):
resume_path = Path(args.resume)
elif last_info.exists():
resume_path = last_info
else:
raise ValueError("Something is wrong.")
logger.log("=> loading checkpoint of the last-info '{:}' start".format(
resume_path))
checkpoint = torch.load(resume_path)
if "last_checkpoint" in checkpoint:
last_checkpoint_path = checkpoint["last_checkpoint"]
if not last_checkpoint_path.exists():
logger.log("Does not find {:}, try another path".format(
last_checkpoint_path))
last_checkpoint_path = (resume_path.parent /
last_checkpoint_path.parent.name /
last_checkpoint_path.name)
assert (last_checkpoint_path.exists()
), "can not find the checkpoint from {:}".format(
last_checkpoint_path)
checkpoint = torch.load(last_checkpoint_path)
start_epoch = checkpoint["epoch"] + 1
search_model.load_state_dict(checkpoint["search_model"])
scheduler.load_state_dict(checkpoint["scheduler"])
base_optimizer.load_state_dict(checkpoint["base_optimizer"])
arch_optimizer.load_state_dict(checkpoint["arch_optimizer"])
valid_accuracies = checkpoint["valid_accuracies"]
arch_genotypes = checkpoint["arch_genotypes"]
discrepancies = checkpoint["discrepancies"]
logger.log(
"=> loading checkpoint of the last-info '{:}' start with {:}-th epoch."
.format(resume_path, start_epoch))
else:
logger.log(
"=> do not find the last-info file : {:} or resume : {:}".format(
last_info, args.resume))
start_epoch, valid_accuracies, arch_genotypes, discrepancies = (
0,
{
"best": -1
},
{},
{},
)
# main procedure
train_func, valid_func = get_procedures(args.procedure)
total_epoch = optim_config.epochs + optim_config.warmup
start_time, epoch_time = time.time(), AverageMeter()
for epoch in range(start_epoch, total_epoch):
scheduler.update(epoch, 0.0)
search_model.set_tau(args.gumbel_tau_max, args.gumbel_tau_min,
epoch * 1.0 / total_epoch)
need_time = "Time Left: {:}".format(
convert_secs2time(epoch_time.avg * (total_epoch - epoch), True))
epoch_str = "epoch={:03d}/{:03d}".format(epoch, total_epoch)
LRs = scheduler.get_lr()
find_best = False
logger.log(
"\n***{:s}*** start {:s} {:s}, LR=[{:.6f} ~ {:.6f}], scheduler={:}, tau={:}, FLOP={:.2f}"
.format(
time_string(),
epoch_str,
need_time,
min(LRs),
max(LRs),
scheduler,
search_model.tau,
MAX_FLOP,
))
# train for one epoch
train_base_loss, train_arch_loss, train_acc1, train_acc5 = train_func(
search_loader,
network,
criterion,
scheduler,
base_optimizer,
arch_optimizer,
optim_config,
{
"epoch-str": epoch_str,
"FLOP-exp": MAX_FLOP * args.FLOP_ratio,
"FLOP-weight": args.FLOP_weight,
"FLOP-tolerant": MAX_FLOP * args.FLOP_tolerant,
},
args.print_freq,
logger,
)
# log the results
logger.log(
"***{:s}*** TRAIN [{:}] base-loss = {:.6f}, arch-loss = {:.6f}, accuracy-1 = {:.2f}, accuracy-5 = {:.2f}"
.format(
time_string(),
epoch_str,
train_base_loss,
train_arch_loss,
train_acc1,
train_acc5,
))
cur_FLOP, genotype = search_model.get_flop("genotype",
model_config._asdict(),
None)
arch_genotypes[epoch] = genotype
arch_genotypes["last"] = genotype
logger.log("[{:}] genotype : {:}".format(epoch_str, genotype))
arch_info, discrepancy = search_model.get_arch_info()
logger.log(arch_info)
discrepancies[epoch] = discrepancy
logger.log(
"[{:}] FLOP : {:.2f} MB, ratio : {:.4f}, Expected-ratio : {:.4f}, Discrepancy : {:.3f}"
.format(
epoch_str,
cur_FLOP,
cur_FLOP / MAX_FLOP,
args.FLOP_ratio,
np.mean(discrepancy),
))
# if cur_FLOP/MAX_FLOP > args.FLOP_ratio:
# init_flop_weight = init_flop_weight * args.FLOP_decay
# else:
# init_flop_weight = init_flop_weight / args.FLOP_decay
# evaluate the performance
if (epoch % args.eval_frequency == 0) or (epoch + 1 == total_epoch):
logger.log("-" * 150)
valid_loss, valid_acc1, valid_acc5 = valid_func(
search_valid_loader,
network,
criterion,
epoch_str,
args.print_freq_eval,
logger,
)
valid_accuracies[epoch] = valid_acc1
logger.log(
"***{:s}*** VALID [{:}] loss = {:.6f}, accuracy@1 = {:.2f}, accuracy@5 = {:.2f} | Best-Valid-Acc@1={:.2f}, Error@1={:.2f}"
.format(
time_string(),
epoch_str,
valid_loss,
valid_acc1,
valid_acc5,
valid_accuracies["best"],
100 - valid_accuracies["best"],
))
if valid_acc1 > valid_accuracies["best"]:
valid_accuracies["best"] = valid_acc1
arch_genotypes["best"] = genotype
find_best = True
logger.log(
"Currently, the best validation accuracy found at {:03d}-epoch :: acc@1={:.2f}, acc@5={:.2f}, error@1={:.2f}, error@5={:.2f}, save into {:}."
.format(
epoch,
valid_acc1,
valid_acc5,
100 - valid_acc1,
100 - valid_acc5,
model_best_path,
))
# save checkpoint
save_path = save_checkpoint(
{
"epoch": epoch,
"args": deepcopy(args),
"valid_accuracies": deepcopy(valid_accuracies),
"model-config": model_config._asdict(),
"optim-config": optim_config._asdict(),
"search_model": search_model.state_dict(),
"scheduler": scheduler.state_dict(),
"base_optimizer": base_optimizer.state_dict(),
"arch_optimizer": arch_optimizer.state_dict(),
"arch_genotypes": arch_genotypes,
"discrepancies": discrepancies,
},
model_base_path,
logger,
)
if find_best:
copy_checkpoint(model_base_path, model_best_path, logger)
last_info = save_checkpoint(
{
"epoch": epoch,
"args": deepcopy(args),
"last_checkpoint": save_path,
},
logger.path("info"),
logger,
)
# measure elapsed time
epoch_time.update(time.time() - start_time)
start_time = time.time()
logger.log("")
logger.log("-" * 100)
last_config_path = logger.path("log") / "seed-{:}-last.config".format(
args.rand_seed)
configure2str(arch_genotypes["last"], str(last_config_path))
logger.log("save the last config int {:} :\n{:}".format(
last_config_path, arch_genotypes["last"]))
best_arch, valid_acc = arch_genotypes["best"], valid_accuracies["best"]
for key, config in arch_genotypes.items():
if key == "last":
continue
FLOP_ratio = config["estimated_FLOP"] / MAX_FLOP
if abs(FLOP_ratio - args.FLOP_ratio) <= args.FLOP_tolerant:
if valid_acc < valid_accuracies[key]:
best_arch, valid_acc = config, valid_accuracies[key]
print("Best-Arch : {:}\nRatio={:}, Valid-ACC={:}".format(
best_arch, best_arch["estimated_FLOP"] / MAX_FLOP, valid_acc))
best_config_path = logger.path("log") / "seed-{:}-best.config".format(
args.rand_seed)
configure2str(best_arch, str(best_config_path))
logger.log("save the last config int {:} :\n{:}".format(
best_config_path, best_arch))
logger.log("\n" + "-" * 200)
logger.log(
"Finish training/validation in {:}, and save final checkpoint into {:}"
.format(convert_secs2time(epoch_time.sum, True), logger.path("info")))
logger.close()
print('Number of parameters: {} \n'.format(num_param))
if opts.resume is None:
model.initialize()
ep0 = -1
total_iter = 0
else:
ep0, total_iter = model.resume(opts.resume)
model.set_scheduler(opts, ep0)
ep0 += 1
print('Start training at epoch {} \n'.format(ep0))
# select dataset
train_set, val_set = get_datasets(opts)
train_loader = DataLoader(dataset=train_set,
num_workers=opts.num_workers,
batch_size=opts.batch_size,
shuffle=True)
val_loader = DataLoader(dataset=val_set,
num_workers=opts.num_workers,
batch_size=1,
shuffle=False)
# Setup directories
output_directory = os.path.join(opts.output_path, 'outputs',
opts.experiment_name)
checkpoint_directory, image_directory = prepare_sub_folder(output_directory)
# evaluation
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)
if os.path.isdir(xargs.save_dir):
if click.confirm(
'\nSave directory already exists in {}. Erase?'.format(
xargs.save_dir, default=False)):
os.system('rm -r ' + xargs.save_dir)
assert not os.path.exists(xargs.save_dir)
os.mkdir(xargs.save_dir)
logger = prepare_logger(args)
writer = SummaryWriter(xargs.save_dir)
perturb_alpha = None
if xargs.perturb:
perturb_alpha = random_alpha
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': xargs.model,
'C': xargs.channel,
'N': xargs.num_cells,
'max_nodes': xargs.max_nodes,
'num_classes': class_num,
'space': search_space,
'affine': bool(xargs.affine),
'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': bool(xargs.affine),
'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, xargs.weight_learning_rate)
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
start_time, search_time, epoch_time = time.time(), AverageMeter(
), AverageMeter()
total_epoch = config.epochs + config.warmup
assert 0 < xargs.early_stop_epoch <= total_epoch - 1
for epoch in range(start_epoch, total_epoch):
if epoch >= xargs.early_stop_epoch:
logger.log(f"Early stop @ {epoch} epoch.")
break
if xargs.perturb:
epsilon_alpha = 0.03 + (xargs.epsilon_alpha -
0.03) * epoch / total_epoch
logger.log(f'epoch {epoch} epsilon_alpha {epsilon_alpha}')
else:
epsilon_alpha = None
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,
xargs.gradient_clip, perturb_alpha, epsilon_alpha)
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)
writer.add_scalar('search/weight_loss', search_w_loss, epoch)
writer.add_scalar('search/weight_top1_acc', search_w_top1, epoch)
writer.add_scalar('search/weight_top5_acc', search_w_top5, epoch)
writer.add_scalar('search/arch_loss', search_a_loss, epoch)
writer.add_scalar('search/arch_top1_acc', search_a_top1, epoch)
writer.add_scalar('search/arch_top5_acc', search_a_top5, epoch)
writer.add_scalar('evaluate/loss', valid_a_loss, epoch)
writer.add_scalar('evaluate/top1_acc', valid_a_top1, epoch)
writer.add_scalar('evaluate/top5_acc', valid_a_top5, epoch)
logger.log(
'[{:}] evaluate : loss={:.2f}, accuracy@1={:.2f}%, accuracy@5={:.2f}%'
.format(epoch_str, valid_a_loss, valid_a_top1, valid_a_top5))
writer.add_scalar('entropy', search_model.entropy, epoch)
per_edge_dict = get_per_egde_value_dict(search_model.arch_parameters)
for edge_name, edge_val in per_edge_dict.items():
writer.add_scalars(f"cell/{edge_name}", edge_val, epoch)
# 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)
save_checkpoint(
{
'epoch': epoch + 1,
'args': deepcopy(args),
'last_checkpoint': save_path,
}, logger.path('info'), logger)
if xargs.snapshoot > 0 and epoch % xargs.snapshoot == 0:
save_checkpoint(
{
'epoch': epoch + 1,
'args': deepcopy(args),
'search_model': search_model.state_dict(),
},
os.path.join(str(logger.model_dir),
f"checkpoint_epoch{epoch}.pth"), logger)
if find_best:
logger.log(
'<<<--->>> The {:}-th epoch : find the highest validation accuracy : {:.2f}%.'
.format(epoch_str, valid_a_top1))
copy_checkpoint(model_base_path, model_best_path, logger)
with torch.no_grad():
logger.log('{:}'.format(search_model.show_alphas()))
if api is not None:
logger.log('{:}'.format(api.query_by_arch(genotypes[epoch])))
index = api.query_index_by_arch(genotypes[epoch])
info = api.query_meta_info_by_index(
index) # This is an instance of `ArchResults`
res_metrics = info.get_metrics(
f'{xargs.dataset}',
'ori-test') # This is a dict with metric names as keys
# cost_metrics = info.get_comput_costs('cifar10')
writer.add_scalar(f'{xargs.dataset}_ground_acc_ori-test',
res_metrics['accuracy'], epoch)
writer.add_scalar(f'{xargs.dataset}_search_acc', valid_a_top1,
epoch)
if xargs.dataset.lower() != 'cifar10':
writer.add_scalar(
f'{xargs.dataset}_ground_acc_x-test',
info.get_metrics(f'{xargs.dataset}', 'x-test')['accuracy'],
epoch)
if find_best:
valid_accuracies['best_gt'] = res_metrics['accuracy']
writer.add_scalar(f"{xargs.dataset}_cur_best_gt_acc_ori-test",
valid_accuracies['best_gt'], epoch)
# measure elapsed time
epoch_time.update(time.time() - start_time)
start_time = time.time()
logger.log('\n' + '-' * 100)
logger.log('{:} : run {:} epochs, cost {:.1f} s, last-geno is {:}.'.format(
args.model, xargs.early_stop_epoch, search_time.sum,
genotypes[xargs.early_stop_epoch - 1]))
if api is not None:
logger.log('{:}'.format(
api.query_by_arch(genotypes[xargs.early_stop_epoch - 1])))
logger.close()
def _setup(self, config):
if not WITH_RAY:
search_space = {
"dataset_name": "cryo_exp_class_2d",
"class_2d": 39,
"lr": 0.005,
"latent_dim": 4,
"n_blocks": 6,
"lambda_regu": 0.2,
"lambda_adv": 0.2,
}
config = Config(search_space)
train_params = TRAIN_PARAMS
train_params["lr"] = config.get("lr")
train_params["lambda_regu"] = config.get("lambda_regu")
train_params["lambda_adv"] = config.get("lambda_adv")
train_params["dataset_name"] = config.get("dataset_name")
train_params["class_2d"] = config.get("class_2d")
nn_architecture = NN_ARCHITECTURE
nn_architecture["latent_dim"] = config.get("latent_dim")
nn_architecture["n_blocks"] = config.get("n_blocks")
train_dataset, val_dataset = datasets.get_datasets(
dataset_name=train_params["dataset_name"],
frac_val=train_params["frac_val"],
batch_size=train_params["batch_size"],
img_shape=nn_architecture["img_shape"],
class_2d=train_params["class_2d"],
)
logging.info("Train: %s", train_utils.get_logging_shape(train_dataset))
logging.info("Val: %s", train_utils.get_logging_shape(val_dataset))
logging.info("NN architecture: ")
logging.info(nn_architecture)
logging.info("Training parameters:")
logging.info(train_params)
train_loader = torch.utils.data.DataLoader(
train_dataset, batch_size=train_params["batch_size"], shuffle=True, **KWARGS
)
val_loader = torch.utils.data.DataLoader(
val_dataset, batch_size=train_params["batch_size"], shuffle=True, **KWARGS
)
m, o, s, t, v = train_utils.init_training(
self.logdir, nn_architecture, train_params
)
modules, optimizers, start_epoch = m, o, s
train_losses_all_epochs, val_losses_all_epochs = t, v
self.train_loader = train_loader
self.val_loader = val_loader
self.modules = modules
self.optimizers = optimizers
self.start_epoch = start_epoch
self.train_losses_all_epochs = train_losses_all_epochs
self.val_losses_all_epochs = val_losses_all_epochs
self.train_params = train_params
self.nn_architecture = nn_architecture
import numpy as np
from sklearn.ensemble import RandomForestRegressor, RandomForestClassifier
import datasets
from sklearn.cross_validation import cross_val_score, KFold
from sklearn.metrics import make_scorer, mean_absolute_error
from wrappers.correlation import ace,rcorrs,rcorrp, bcdcor
from collections import defaultdict
MAD = make_scorer(mean_absolute_error, greater_is_better=False)
sc=defaultdict(lambda:defaultdict(lambda:defaultdict(int)))
for ds in datasets.get_datasets():
Xc, yc = datasets.get_data(ds, convert='numbers', standardize=False)
kf = KFold(Xc.shape[0], 5, shuffle=True)
for vi in (bcdcor, rcorrp, rcorrs, ace):
cor = vi(Xc, yc, datasets.get_column_index(ds))
#print zip(ds['category']+ds['numeric'],cor)
for drop in range(10):
vi_name = vi.__name__
X, y = datasets.get_data(ds, drop=np.argsort(cor)[:drop])
if ds['rtype'] == 'Binary':
metric = 'log_loss'
clf = RandomForestClassifier(n_estimators=200,min_samples_leaf=5,n_jobs=1)
else:
metric = MAD
clf = RandomForestRegressor(n_estimators=200,min_samples_leaf=5,n_jobs=1)
sc[ds['name']][vi_name][drop] = cross_val_score(clf, X, y, cv=kf, n_jobs=3, scoring=metric).mean()
print ds['name']+' '+vi_name+" "+str(drop)+' '+str(sc[ds['name']][vi_name][drop])
import sys
sys.path.append('/home/glen/workspace/DataRobot')
from ModelingMachine.engine.eda_multi import _get_report_level_one
from ModelingMachine.engine.metrics import direction_by_name, metric_by_name
import pandas
import numpy as np
import datasets
def get_accuracy_metric(col):
eda_report = {}
rp = _get_report_level_one(eda_report, col, col.dtype)
return rp['metric_options']['accuracy']['short_name']
#for ds in datasets.get_datasets(name=['fastiron-train-30k','kickcars_train_full','census_1990_small','french_damage_cost','allstate_nonzero_small','amazon_small_no-c','credit_full','bank_marketing_small','bio_grid_small','mets','trainingDataWithoutNegativeWeights','census_2012h_small']):
#for ds in datasets.get_datasets(name='trainingDataWithoutNegativeWeights'):
for ds in datasets.get_datasets(name='fastiron-train-30k'):
#for ds in datasets.get_datasets(name='fastiron_small'):
#for ds in datasets.get_datasets(name='kickcars_train_full'):
#for ds in datasets.get_datasets(name='french_damage_cost'):
#for ds in datasets.get_datasets(name='census_1990_small'):
X, y = datasets.get_data(ds,standardize=False,convert='numbers')
cols = datasets.get_columns(ds)
cats = datasets.get_columns(ds,'category')
sc1={}
sc2={}
klist=(0,)
#klist = (1,5,10,20,40,80)
df=pandas.DataFrame(np.hstack((X,np.reshape(y,(-1,1)))),columns=cols+[ds['target']])
rm = get_accuracy_metric(df[ds['target']])
mdir = direction_by_name(rm)
mfunc = metric_by_name(rm)
def main(xargs):
assert torch.cuda.is_available(), 'CUDA is not available.'
torch.backends.cudnn.enabled = True
torch.backends.cudnn.benchmark = False
torch.backends.cudnn.deterministic = True
torch.set_num_threads(xargs.workers)
prepare_seed(xargs.rand_seed)
logger = prepare_logger(args)
train_data, valid_data, xshape, class_num = get_datasets(
xargs.dataset, xargs.data_path, -1)
if xargs.overwite_epochs is None:
extra_info = {'class_num': class_num, 'xshape': xshape}
else:
extra_info = {
'class_num': class_num,
'xshape': xshape,
'epochs': xargs.overwite_epochs
}
config = load_config(xargs.config_path, extra_info, logger)
search_loader, train_loader, valid_loader = get_nas_search_loaders(
train_data, valid_data, xargs.dataset, 'configs/nas-benchmark/',
(config.batch_size, config.test_batch_size), xargs.workers)
logger.log(
'||||||| {:10s} ||||||| Search-Loader-Num={:}, Valid-Loader-Num={:}, batch size={:}'
.format(xargs.dataset, len(search_loader), len(valid_loader),
config.batch_size))
logger.log('||||||| {:10s} ||||||| Config={:}'.format(
xargs.dataset, config))
search_space = get_search_spaces(xargs.search_space, 'nas-bench-301')
model_config = dict2config(
dict(name='generic',
C=xargs.channel,
N=xargs.num_cells,
max_nodes=xargs.max_nodes,
num_classes=class_num,
space=search_space,
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, 'topology', 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']
baseline = checkpoint['baseline']
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.return_topK(1, True)[0]
}
baseline = 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={:}'.format(
epoch_str, need_time, min(w_scheduler.get_lr())))
network.set_drop_path(
float(epoch + 1) / total_epoch, xargs.drop_path_rate)
if xargs.algo == 'gdas':
network.set_tau(xargs.tau_max -
(xargs.tau_max - xargs.tau_min) * epoch /
(total_epoch - 1))
logger.log('[RESET tau as : {:} and drop_path as {:}]'.format(
network.tau, network.drop_path))
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, xargs.algo, 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))
if xargs.algo == 'enas':
ctl_loss, ctl_acc, baseline, ctl_reward \
= train_controller(valid_loader, network, criterion, a_optimizer, baseline, epoch_str, xargs.print_freq, logger)
logger.log(
'[{:}] controller : loss={:}, acc={:}, baseline={:}, reward={:}'
.format(epoch_str, ctl_loss, ctl_acc, baseline, ctl_reward))
genotype, temp_accuracy = get_best_arch(valid_loader, network,
xargs.eval_candidate_num,
xargs.algo)
if xargs.algo == 'setn' or xargs.algo == 'enas':
network.set_cal_mode('dynamic', genotype)
elif xargs.algo == 'gdas':
network.set_cal_mode('gdas', None)
elif xargs.algo.startswith('darts'):
network.set_cal_mode('joint', None)
elif xargs.algo == 'random':
network.set_cal_mode('urs', None)
else:
raise ValueError('Invalid algorithm name : {:}'.format(xargs.algo))
logger.log('[{:}] - [get_best_arch] : {:} -> {:}'.format(
epoch_str, genotype, temp_accuracy))
valid_a_loss, valid_a_top1, valid_a_top5 = valid_func(
valid_loader, network, criterion, xargs.algo, 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),
'baseline': baseline,
'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.eval_candidate_num,
xargs.algo)
if xargs.algo == 'setn' or xargs.algo == 'enas':
network.set_cal_mode('dynamic', genotype)
elif xargs.algo == 'gdas':
network.set_cal_mode('gdas', None)
elif xargs.algo.startswith('darts'):
network.set_cal_mode('joint', None)
elif xargs.algo == 'random':
network.set_cal_mode('urs', None)
else:
raise ValueError('Invalid algorithm name : {:}'.format(xargs.algo))
search_time.update(time.time() - start_time)
valid_a_loss, valid_a_top1, valid_a_top5 = valid_func(
valid_loader, network, criterion, xargs.algo, 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, '200')))
logger.close()
def evaluate_all_datasets(channels: Text, datasets: List[Text], xpaths: List[Text],
splits: List[Text], config_path: Text, seed: int, workers: int, logger):
machine_info = get_machine_info()
all_infos = {'info': machine_info}
all_dataset_keys = []
# look all the dataset
for dataset, xpath, split in zip(datasets, xpaths, splits):
# the train and 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 the splitted 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-index= 9930, arch=|nor_conv_3x3~0|+|nor_conv_3x3~0|nor_conv_3x3~1|+|skip_connect~0|nor_conv_3x3~1|nor_conv_3x3~2|
# this genotype is the architecture with the highest accuracy on CIFAR-100 validation set
genotype = '|nor_conv_3x3~0|+|nor_conv_3x3~0|nor_conv_3x3~1|+|skip_connect~0|nor_conv_3x3~1|nor_conv_3x3~2|'
arch_config = dict2config(dict(name='infer.shape.tiny', channels=channels, genotype=genotype, num_classes=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
sc = 0
for i in range(act.shape[0]):
if act[i]==0:
sc += 2 * pred[i]
else:
if pred[i]<=0:
sc += 2 * (act[i] * math.log(act[i]/0.1) - (act[i] - pred[i]))
else:
sc += 2 * (act[i] * math.log(act[i]/pred[i]) - (act[i] - pred[i]))
return sc/act.shape[0]
RMSLE = make_scorer(rmsle, greater_is_better=False)
PSDEV = make_scorer(psdev, greater_is_better=False)
#for ds in datasets.get_datasets(rtype='Positive'): #,name='census_1990_small'):
for ds in datasets.get_datasets(name='census_2012h_small'):
#if 'census' in ds['name']: continue
X, y = datasets.get_data(ds,standardize=False)
kf = KFold(X.shape[0], 2, shuffle=True, random_state=1234)
for a in (0.1, 0.3, 0.6, 0.9):
for lm in (1, 0.1, 0.01):
clf1 = CGLM(distribution='Poisson',trace=False)
clf2 = Ridge(alpha=lm)
clf3 = ElasticNetC(distribution='Poisson', alpha=lm, l1_ratio=a, tolerance=0.001)
#clf3t = ElasticNetC(distribution='Tweedie', alpha=lm, l1_ratio=a, tolerance=0.001,p=1.5)
#clf3g = ElasticNetC(distribution='Gamma', alpha=lm, l1_ratio=a, tolerance=0.001)
clf4 = GlmnetWrapper(**{'family': 'gaussian', 'alpha': a, 'lambda': lm, 'maxit': 300})
#clf4 = GlmnetWrapper(**{'family': 'poisson', 'alpha': a, 'lambda': lm, 'maxit': 300})
score_func = PSDEV #'roc_auc'
for clf in (clf1, clf2, clf3, clf4):
#for clf in (clf1, clf2, clf3, clf3t, clf3g, clf4):
from analysis import generate_prototype_activation_matrix, generate_prototype_activation_matrices
from datasets import get_datasets
from helpers import makedir, load_or_create_experiment, SAVED_MODELS_PATH, LOGS_DIR, \
CHECKPOINT_FREQUENCY_STEPS
from model import construct_PPNet, construct_PPNet_for_config
from push import push_prototypes
from save import load_train_state, save_train_state, get_state_path_for_prefix, load_model_from_train_state, \
snapshot_code
from train_and_test import warm_only, train, joint, test, last_only, valid, TrainMode
args, config, seed, DEBUG, load_state_path, checkpoint_file_prefix = load_or_create_experiment(
)
workers = 0 if DEBUG else 4
train_loader, train_push_loader, valid_loader, valid_push_loader, test_loader, test_push_loader = get_datasets(
args.dataset, seed, workers, config)
print('Dataset loaded.')
ppnet = construct_PPNet_for_config(config).cuda()
if config.base_architecture != 'noop':
summary(ppnet, (10, 3, config.img_size, config.img_size),
col_names=("input_size", "output_size", "num_params"),
depth=4)
else:
summary(ppnet, (10, *config.noop_features_size),
col_names=("input_size", "output_size", "num_params"),
depth=4)
joint_optimizer_specs = [{
import datasets
from wrappers.elastic_net_c import ElasticNetC
import time
from wrappers.glmnet import GlmnetWrapper
from wrappers.zeroinfl import ZeroInflWrapper
from wrappers.glm_c import CGLM
from sklearn.cross_validation import cross_val_score, KFold
from sklearn.metrics import make_scorer, mean_absolute_error
from sklearn.linear_model import SGDRegressor, Ridge, LogisticRegression, SGDClassifier, ElasticNet
MAD = make_scorer(mean_absolute_error, greater_is_better=False)
score_func='r2'
for ds in datasets.get_datasets(name='mets_short'):
X, y = datasets.get_data(ds)
kf = KFold(X.shape[0], 15, shuffle=True, random_state=1234)
for a in (0.1, 0.5, 1.0, 1.5):
clf1 = Ridge(alpha=a)
clf2 = CGLM(distribution='Poisson',trace=False)
clf3 = ZeroInflWrapper()
for clf in (clf1,clf2,clf3):
if a!=0.1 and clf!=clf1:
continue
st = time.time()
print '%-20s %20s a %3.1f sc %9.4f tm %5.2f' % \
(ds['name'], clf.__class__.__name__, a,
cross_val_score(clf, X, y, cv=kf, scoring=score_func, n_jobs=3).mean(), time.time()-st)
from tesla.ensemble import GradientBoostingClassifier
from sklearn.cross_validation import cross_val_score, KFold
from math import sqrt
import datasets
import sys
score_func='roc_auc'
#for ds in datasets.get_datasets(name='kickcars_small'):
for ds in datasets.get_datasets(name='kickcars_train_full'):
X, y = datasets.get_data(ds,standardize=False)
kf = KFold(X.shape[0], 2, shuffle=True, random_state=1234)
#for lr in (0.1,0.05,0.01):
#for lr in (0.05,):
#for md in (2,3,4):
lr=0.05
for md in (2,3,4):
for ms in (1,):
t=100
td=100
bestsc = 0
bestt = 40
incr=0
while True:
clf = GradientBoostingClassifier(min_samples_leaf=ms,learning_rate=0.05,max_depth=md,n_estimators=t)
sc = cross_val_score(clf, X, y, cv=kf, scoring=score_func, n_jobs=-1).mean()
print '%f %d %d %d %f' % (lr,md,ms,t,sc)
if sc > bestsc:
bestsc=sc
bestt=t
incr=0
td *= 2
def main(xargs):
assert torch.cuda.is_available(), 'CUDA is not available.'
torch.backends.cudnn.enabled = True
torch.backends.cudnn.benchmark = False
torch.backends.cudnn.deterministic = True
torch.set_num_threads(xargs.workers)
prepare_seed(xargs.rand_seed)
logger = prepare_logger(args)
train_data, valid_data, xshape, class_num = get_datasets(
xargs.dataset, xargs.data_path, -1)
#config_path = 'configs/nas-benchmark/algos/GDAS.config'
config = load_config(xargs.config_path, {
'class_num': class_num,
'xshape': xshape
}, logger)
search_loader, _, valid_loader = get_nas_search_loaders(
train_data, valid_data, xargs.dataset, 'configs/nas-benchmark/',
config.batch_size, xargs.workers)
logger.log(
'||||||| {:10s} ||||||| Search-Loader-Num={:}, batch size={:}'.format(
xargs.dataset, len(search_loader), config.batch_size))
logger.log('||||||| {:10s} ||||||| Config={:}'.format(
xargs.dataset, config))
search_space = get_search_spaces('cell', xargs.search_space_name)
if xargs.model_config is None:
model_config = dict2config(
{
'name': 'GDAS',
'C': xargs.channel,
'N': xargs.num_cells,
'max_nodes': xargs.max_nodes,
'num_classes': class_num,
'space': search_space,
'affine': False,
'track_running_stats': bool(xargs.track_running_stats)
}, None)
else:
model_config = load_config(
xargs.model_config, {
'num_classes': class_num,
'space': search_space,
'affine': False,
'track_running_stats': bool(xargs.track_running_stats)
}, None)
search_model = get_cell_based_tiny_net(model_config)
logger.log('search-model :\n{:}'.format(search_model))
logger.log('model-config : {:}'.format(model_config))
w_optimizer, w_scheduler, criterion = get_optim_scheduler(
search_model.get_weights(), config)
a_optimizer = torch.optim.Adam(search_model.get_alphas(),
lr=xargs.arch_learning_rate,
betas=(0.5, 0.999),
weight_decay=xargs.arch_weight_decay)
logger.log('w-optimizer : {:}'.format(w_optimizer))
logger.log('a-optimizer : {:}'.format(a_optimizer))
logger.log('w-scheduler : {:}'.format(w_scheduler))
logger.log('criterion : {:}'.format(criterion))
flop, param = get_model_infos(search_model, xshape)
logger.log('FLOP = {:.2f} M, Params = {:.2f} MB'.format(flop, param))
logger.log('search-space [{:} ops] : {:}'.format(len(search_space),
search_space))
if xargs.arch_nas_dataset is None:
api = None
else:
api = API(xargs.arch_nas_dataset)
logger.log('{:} create API = {:} done'.format(time_string(), api))
last_info, model_base_path, model_best_path = logger.path(
'info'), logger.path('model'), logger.path('best')
network, criterion = torch.nn.DataParallel(
search_model).cuda(), criterion.cuda()
if last_info.exists(): # automatically resume from previous checkpoint
logger.log("=> loading checkpoint of the last-info '{:}' start".format(
last_info))
last_info = torch.load(last_info)
start_epoch = last_info['epoch']
checkpoint = torch.load(last_info['last_checkpoint'])
genotypes = checkpoint['genotypes']
valid_accuracies = checkpoint['valid_accuracies']
search_model.load_state_dict(checkpoint['search_model'])
w_scheduler.load_state_dict(checkpoint['w_scheduler'])
w_optimizer.load_state_dict(checkpoint['w_optimizer'])
a_optimizer.load_state_dict(checkpoint['a_optimizer'])
logger.log(
"=> loading checkpoint of the last-info '{:}' start with {:}-th epoch."
.format(last_info, start_epoch))
else:
logger.log("=> do not find the last-info file : {:}".format(last_info))
start_epoch, valid_accuracies, genotypes = 0, {
'best': -1
}, {
-1: search_model.genotype()
}
# start training
start_time, search_time, epoch_time, total_epoch = time.time(
), AverageMeter(), AverageMeter(), config.epochs + config.warmup
for epoch in range(start_epoch, total_epoch):
w_scheduler.update(epoch, 0.0)
need_time = 'Time Left: {:}'.format(
convert_secs2time(epoch_time.val * (total_epoch - epoch), True))
epoch_str = '{:03d}-{:03d}'.format(epoch, total_epoch)
search_model.set_tau(xargs.tau_max -
(xargs.tau_max - xargs.tau_min) * epoch /
(total_epoch - 1))
logger.log('\n[Search the {:}-th epoch] {:}, tau={:}, LR={:}'.format(
epoch_str, need_time, search_model.get_tau(),
min(w_scheduler.get_lr())))
search_w_loss, search_w_top1, search_w_top5, valid_a_loss , valid_a_top1 , valid_a_top5 \
= search_func(search_loader, network, criterion, w_scheduler, w_optimizer, a_optimizer, epoch_str, xargs.print_freq, logger)
search_time.update(time.time() - start_time)
logger.log(
'[{:}] searching : loss={:.2f}, accuracy@1={:.2f}%, accuracy@5={:.2f}%, time-cost={:.1f} s'
.format(epoch_str, search_w_loss, search_w_top1, search_w_top5,
search_time.sum))
logger.log(
'[{:}] evaluate : loss={:.2f}, accuracy@1={:.2f}%, accuracy@5={:.2f}%'
.format(epoch_str, valid_a_loss, valid_a_top1, valid_a_top5))
# check the best accuracy
valid_accuracies[epoch] = valid_a_top1
if valid_a_top1 > valid_accuracies['best']:
valid_accuracies['best'] = valid_a_top1
genotypes['best'] = search_model.genotype()
find_best = True
else:
find_best = False
genotypes[epoch] = search_model.genotype()
logger.log('<<<--->>> The {:}-th epoch : {:}'.format(
epoch_str, genotypes[epoch]))
# save checkpoint
save_path = save_checkpoint(
{
'epoch': epoch + 1,
'args': deepcopy(xargs),
'search_model': search_model.state_dict(),
'w_optimizer': w_optimizer.state_dict(),
'a_optimizer': a_optimizer.state_dict(),
'w_scheduler': w_scheduler.state_dict(),
'genotypes': genotypes,
'valid_accuracies': valid_accuracies
}, model_base_path, logger)
last_info = save_checkpoint(
{
'epoch': epoch + 1,
'args': deepcopy(args),
'last_checkpoint': save_path,
}, logger.path('info'), logger)
if find_best:
logger.log(
'<<<--->>> The {:}-th epoch : find the highest validation accuracy : {:.2f}%.'
.format(epoch_str, valid_a_top1))
copy_checkpoint(model_base_path, model_best_path, logger)
with torch.no_grad():
logger.log('{:}'.format(search_model.show_alphas()))
if api is not None:
logger.log('{:}'.format(api.query_by_arch(genotypes[epoch],
'200')))
# measure elapsed time
epoch_time.update(time.time() - start_time)
start_time = time.time()
logger.log('\n' + '-' * 100)
# check the performance from the architecture dataset
logger.log(
'GDAS : run {:} epochs, cost {:.1f} s, last-geno is {:}.'.format(
total_epoch, search_time.sum, genotypes[total_epoch - 1]))
if api is not None:
logger.log('{:}'.format(
api.query_by_arch(genotypes[total_epoch - 1], '200')))
logger.close()
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()