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 setup ():
global app
log.debug("Loading global app settings...")
app = config_utils.load_config(paths.config_path(MAINFILE), paths.app_path(MAINSPEC))
log.debug("Loading keymap...")
global keymap
keymap = config_utils.load_config(paths.config_path("keymap.keymap"), paths.app_path("keymaps/"+app['app-settings']['load_keymap']))
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 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 setup():
global app
log.debug("Loading global app settings...")
app = config_utils.load_config(paths.config_path(MAINFILE),
paths.app_path(MAINSPEC))
log.debug("Loading keymap...")
global keymap
keymap = config_utils.load_config(
paths.config_path("keymap.keymap"),
paths.app_path("keymaps/" + app['app-settings']['load_keymap']))
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 main(bam_file, config_file=None, chrom='all', start=0, end=None,
outfile=None, normalize=False, use_tempfile=False):
if config_file:
config = load_config(config_file)
else:
config = {"program": {"ucsc_bigwig": "wigToBigWig"}}
if outfile is None:
outfile = "%s.bigwig" % os.path.splitext(bam_file)[0]
if start > 0:
start = int(start) - 1
if end is not None:
end = int(end)
regions = [(chrom, start, end)]
if os.path.abspath(bam_file) == os.path.abspath(outfile):
sys.stderr.write("Bad arguments, input and output files are the same.\n")
sys.exit(1)
if not (os.path.exists(outfile) and os.path.getsize(outfile) > 0):
if use_tempfile:
# Use a temp file to avoid any possiblity of not having write permission
out_handle = tempfile.NamedTemporaryFile(delete=False)
wig_file = out_handle.name
else:
wig_file = "%s.wig" % os.path.splitext(outfile)[0]
out_handle = open(wig_file, "w")
with closing(out_handle):
chr_sizes, wig_valid = write_bam_track(bam_file, regions, config, out_handle,
normalize)
try:
if wig_valid:
convert_to_bigwig(wig_file, chr_sizes, config, outfile)
finally:
os.remove(wig_file)
def fill_list(self):
sessionsList = []
log.debug("Filling the sessions list.")
self.sessions = []
for i in os.listdir(paths.config_path()):
if os.path.isdir(paths.config_path(i)):
log.debug("Adding session %s" % (i, ))
strconfig = "%s/session.conf" % (paths.config_path(i))
config_test = config_utils.load_config(strconfig)
name = config_test["twitter"]["user_name"]
if config_test["twitter"]["user_key"] != "" and config_test[
"twitter"]["user_secret"] != "":
sessionsList.append(name)
self.sessions.append(i)
else:
try:
log.debug("Deleting session %s" % (i, ))
shutil.rmtree(paths.config_path(i))
except:
output.speak(
"An exception was raised while attempting to clean malformed session data. See the error log for details. If this message persists, contact the developers.",
True)
os.exception(
"Exception thrown while removing malformed session"
)
self.view.fill_list(sessionsList)
def LoadConfiguration(self):
"""
Set the configuration values specified in configuration.json
Return True if configuration sucessfully loaded.
"""
setup_config()
load_config()
try:
log.info("Loaded configuration")
return True
except Exception as e:
log.critical(f"Failed to load configuration, {e}")
def get_config_custom():
# Read the custom configuration file and pass it to AngularJS service config.service.ts
config = load_config(file_name='config.json')
if config is None:
response = json.dumps({'error': 'Unable to read config.json'})
else:
response = json.dumps(config)
return make_json_response(response)
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 get_configuration(self):
""" Gets settings for a session."""
file_ = "%s/session.conf" % (self.session_id,)
# try:
log.debug("Creating config file %s" % (file_,))
self.settings = config_utils.load_config(paths.config_path(file_), paths.app_path("Conf.defaults"))
self.init_sound()
self.deshelve()
def get_configuration(self):
""" Gets settings for a session."""
file_ = "%s/session.conf" % (self.session_id,)
# try:
log.debug("Creating config file %s" % (file_,))
self.settings = config_utils.load_config(paths.config_path(file_), paths.app_path("Conf.defaults"))
self.init_sound()
self.deshelve()
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 get_nas_search_loaders(train_data, valid_data, dataset, config_root, batch_size, workers):
if isinstance(batch_size, (list,tuple)):
batch, test_batch = batch_size
else:
batch, test_batch = batch_size, batch_size
if dataset == 'cifar10':
#split_Fpath = 'configs/nas-benchmark/cifar-split.txt'
cifar_split = load_config('{:}/cifar-split.txt'.format(config_root), None, None)
train_split, valid_split = cifar_split.train, cifar_split.valid # search over the proposed training and validation set
#logger.log('Load split file from {:}'.format(split_Fpath)) # they are two disjoint groups in the original CIFAR-10 training set
# To split data
xvalid_data = deepcopy(train_data)
if hasattr(xvalid_data, 'transforms'): # to avoid a print issue
xvalid_data.transforms = valid_data.transform
xvalid_data.transform = deepcopy( valid_data.transform )
search_data = SearchDataset(dataset, train_data, train_split, valid_split)
# data loader
search_loader = torch.utils.data.DataLoader(search_data, batch_size=batch, shuffle=True , num_workers=workers, pin_memory=True)
train_loader = torch.utils.data.DataLoader(train_data , batch_size=batch, sampler=torch.utils.data.sampler.SubsetRandomSampler(train_split), num_workers=workers, pin_memory=True)
valid_loader = torch.utils.data.DataLoader(xvalid_data, batch_size=test_batch, sampler=torch.utils.data.sampler.SubsetRandomSampler(valid_split), num_workers=workers, pin_memory=True)
elif dataset == 'cifar100':
cifar100_test_split = load_config('{:}/cifar100-test-split.txt'.format(config_root), None, None)
search_train_data = train_data
search_valid_data = deepcopy(valid_data) ; search_valid_data.transform = train_data.transform
search_data = SearchDataset(dataset, [search_train_data,search_valid_data], list(range(len(search_train_data))), cifar100_test_split.xvalid)
search_loader = torch.utils.data.DataLoader(search_data, batch_size=batch, shuffle=True , num_workers=workers, pin_memory=True)
train_loader = torch.utils.data.DataLoader(train_data , batch_size=batch, shuffle=True , num_workers=workers, pin_memory=True)
valid_loader = torch.utils.data.DataLoader(valid_data , batch_size=test_batch, sampler=torch.utils.data.sampler.SubsetRandomSampler(cifar100_test_split.xvalid), num_workers=workers, pin_memory=True)
elif dataset == 'ImageNet16-120':
imagenet_test_split = load_config('{:}/imagenet-16-120-test-split.txt'.format(config_root), None, None)
search_train_data = train_data
search_valid_data = deepcopy(valid_data) ; search_valid_data.transform = train_data.transform
search_data = SearchDataset(dataset, [search_train_data,search_valid_data], list(range(len(search_train_data))), imagenet_test_split.xvalid)
search_loader = torch.utils.data.DataLoader(search_data, batch_size=batch, shuffle=True , num_workers=workers, pin_memory=True)
train_loader = torch.utils.data.DataLoader(train_data , batch_size=batch, shuffle=True , num_workers=workers, pin_memory=True)
valid_loader = torch.utils.data.DataLoader(valid_data , batch_size=test_batch, sampler=torch.utils.data.sampler.SubsetRandomSampler(imagenet_test_split.xvalid), num_workers=workers, pin_memory=True)
else:
raise ValueError('invalid dataset : {:}'.format(dataset))
return search_loader, train_loader, valid_loader
def __init__(self, title):
self.recording=False
wx.Frame.__init__(self, None, title=title, size=(350,200)) # initialize the wx frame
# load config and validate a specification file
self.MAINFILE = "uploader.cfg"
self.MAINSPEC = "app.defaults"
self.appconfig = config_utils.load_config(self.MAINFILE, self.MAINSPEC)
# window events and controls
self.Bind(wx.EVT_CLOSE, self.OnClose)
self.panel = wx.Panel(self)
self.main_box = wx.BoxSizer(wx.VERTICAL)
self.select_file = wx.Button(self.panel, -1, "&Select File")
self.select_file.Bind(wx.EVT_BUTTON, self.SelectFile)
self.main_box.Add(self.select_file, 0, wx.ALL, 10)
self.record = wx.Button(self.panel, -1, "&Record")
self.record.Bind(wx.EVT_BUTTON, self.Record)
self.main_box.Add(self.record, 0, wx.ALL, 10)
self.link_label = wx.StaticText(self.panel, -1, "Audio U&RL")
self.link = wx.TextCtrl(self.panel, -1, "",style=wx.TE_READONLY)
self.link.SetValue("Waiting for audio...")
self.main_box.Add(self.link, 0, wx.ALL, 10)
self.key_label = wx.StaticText(self.panel, -1,"SNDUp API &Key")
self.key = wx.TextCtrl(self.panel, -1, "")
self.main_box.Add(self.key, 0, wx.ALL, 10)
self.key.SetValue(self.appconfig["general"]["APIKey"])
self.services_label=wx.StaticText(self.panel, -1, "Upload to")
self.services = wx.ComboBox(self.panel, -1, choices=services, value=services[0], style=wx.CB_READONLY)
self.services.Bind(wx.EVT_COMBOBOX, self.on_service_change)
self.main_box.Add(self.services, 0, wx.ALL, 10)
self.upload = wx.Button(self.panel, -1, "&Upload")
self.upload.Bind(wx.EVT_BUTTON, self.OnUpload)
self.main_box.Add(self.upload, 0, wx.ALL, 10)
self.upload.Hide()
self.twitter_label = wx.StaticText(self.panel, -1,"Tweet Te&xt")
self.twitter_text = wx.TextCtrl(self.panel, -1, "")
self.main_box.Add(self.twitter_text, 0, wx.ALL, 10)
self.twitter_text.Hide()
self.tweet = wx.Button(self.panel, -1, "&Tweet")
self.tweet.Bind(wx.EVT_BUTTON, self.Tweet)
self.tweet.Hide()
self.main_box.Add(self.tweet, 0, wx.ALL, 10)
self.new = wx.Button(self.panel, -1, "&Attach another file")
self.new.Bind(wx.EVT_BUTTON, self.Reset)
self.main_box.Add(self.new, 0, wx.ALL, 10)
self.new.Hide()
self.close = wx.Button(self.panel, wx.ID_CLOSE, "&Close")
self.close.Bind(wx.EVT_BUTTON, self.OnClose)
self.main_box.Add(self.close, 0, wx.ALL, 10)
self.panel.Layout()
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 send_to_terra():
"""Send manifest data to Terra data framework. Returns URL to UI that will open in new window."""
file_to_retrieve = 'terra' #dummy value to prevent erroring out
config = load_config(file_name='config.json')
module_name = config['cloud-options']['service-name'] + '_handler' # aws_handler, download_handler
class_name = config['cloud-options']['service-name'].title() + 'Handler' # AwsHandler, #DownloadHandler
handler = None
import importlib
module = importlib.import_module(module_name)
handler_class = getattr(module, class_name)
handler = handler_class(file_to_retrieve=file_to_retrieve)
file_data = handler.handle_manifest(request)
metadata = handler.handle_metadata(request)
json_entities = handler.build_json(file_data, metadata)
return handler.post_json_to_terra(json_entities)
def download(date_range):
from utils import download_utils as du
from netCDF4 import Dataset as ncDataset
import os
import cdsapi
config = load_config()
c = cdsapi.Client()
du.check_pandas_datetime_index(date_range)
local_name = os.path.join(config['path']['raw'],
'era5/{year}/era5_{year}{month:02d}_u10_slp.nc')
month_range = du.date_range_to_custom_freq(date_range)
for t in month_range:
local_path = local_name.format(year=t.year, month=t.month)
request_ERA5_U10_SLP(cds_client, t.year, t.month, local_path)
def get_cart_metadata():
file_to_retrieve = 'metadata'
config = load_config(file_name='config.json')
module_name = config['cloud-options']['service-name'] + '_handler' # aws_handler, download_handler
class_name = config['cloud-options']['service-name'].title() + 'Handler' # AwsHandler, #DownloadHandler
handler = None
import importlib
module = importlib.import_module(module_name)
handler_class = getattr(module, class_name)
handler = handler_class(file_to_retrieve=file_to_retrieve)
if config['cloud-options']['service-name'].lower() == 'aws':
# AWS returns a manifest ID {'manifest_id': 'xxx-xxxxxxx-xxxx-xxxxx'}
return handler.handle_metadata(request)
else:
# Download the file
metadata = handler.handle_metadata(request)
return handler.download_file(request, metadata)
def fill_list(self):
sessionsList = []
log.debug("Filling the sessions list.")
self.sessions = []
for i in os.listdir(paths.config_path()):
if os.path.isdir(paths.config_path(i)):
log.debug("Adding session %s" % (i,))
strconfig = "%s/session.conf" % (paths.config_path(i))
config_test = config_utils.load_config(strconfig)
name = config_test["twitter"]["user_name"]
if config_test["twitter"]["user_key"] != "" and config_test["twitter"]["user_secret"] != "":
sessionsList.append(name)
self.sessions.append(i)
else:
try:
log.debug("Deleting session %s" % (i,))
shutil.rmtree(paths.config_path(i))
except:
output.speak("An exception was raised while attempting to clean malformed session data. See the error log for details. If this message persists, contact the developers.",True)
os.exception("Exception thrown while removing malformed session")
self.view.fill_list(sessionsList)
def main(bam_file,
config_file=None,
chrom='all',
start=0,
end=None,
outfile=None,
normalize=False,
use_tempfile=False):
if config_file:
config = load_config(config_file)
else:
config = {"program": {"ucsc_bigwig": "wigToBigWig"}}
if outfile is None:
outfile = "%s.bigwig" % os.path.splitext(bam_file)[0]
if start > 0:
start = int(start) - 1
if end is not None:
end = int(end)
regions = [(chrom, start, end)]
if os.path.abspath(bam_file) == os.path.abspath(outfile):
sys.stderr.write(
"Bad arguments, input and output files are the same.\n")
sys.exit(1)
if not (os.path.exists(outfile) and os.path.getsize(outfile) > 0):
if use_tempfile:
# Use a temp file to avoid any possiblity of not having write permission
out_handle = tempfile.NamedTemporaryFile(delete=False)
wig_file = out_handle.name
else:
wig_file = "%s.wig" % os.path.splitext(outfile)[0]
out_handle = open(wig_file, "w")
with closing(out_handle):
chr_sizes, wig_valid = write_bam_track(bam_file, regions, config,
out_handle, normalize)
try:
if wig_valid:
convert_to_bigwig(wig_file, chr_sizes, config, outfile)
finally:
os.remove(wig_file)
def evaluate_line():
config_path = os.path.join(FLAGS.config_path, 'config')
test_config = load_config(config_path)
_, word_to_id = read_vocab(test_config['vocab_file'])
categorys, cat_to_id = read_category()
contents, labels = read_file('data/cnews.val2.txt')
model = Model(test_config)
session = tf.Session()
session.run(tf.global_variables_initializer())
saver = tf.train.Saver()
# 读取模型
checkpoint_path = os.path.join(FLAGS.checkpoints_path)
checkpoint_file = tf.train.latest_checkpoint(checkpoint_path)
saver.restore(session, checkpoint_file)
while True:
line = input("请输入测试句子:")
x_input = [[word_to_id[x] for x in line if x in word_to_id]]
x_pad = kr.preprocessing.sequence.pad_sequences(x_input, 600)
predict = model.evaluate(session, x_pad)
print(categorys[predict[0][0]])
def __init__(self, cfg):
if 'surface_grasp_logdir_folder' in cfg:
# for sim evaluation
self._contact_grasp_cfg = config_utils.load_config(
cfg['surface_grasp_logdir_folder'],
batch_size=1,
arg_configs=cfg.arg_configs)
self._cfg = cfg
self._num_samples = self._cfg.num_samples
else:
self._contact_grasp_cfg = cfg
self._model_func = importlib.import_module(
self._contact_grasp_cfg['MODEL']['model'])
self._num_input_points = self._contact_grasp_cfg['DATA'][
'raw_num_points'] if 'raw_num_points' in self._contact_grasp_cfg[
'DATA'] else self._contact_grasp_cfg['DATA']['num_point']
self.placeholders = self._model_func.placeholder_inputs(
self._contact_grasp_cfg['OPTIMIZER']['batch_size'],
self._num_input_points,
self._contact_grasp_cfg['DATA']['input_normals'])
self.model_ops = {}
def get_projects_visualization_data():
graph_type = request.args.get('graph_type')
if graph_type == 'bar_graph':
graph_data_length = len(bar_graph_data)
p_str = "{{ \"count\": {0}, \"sort\": \"\", \"from\": 1, \"page\": 1, \"total\": {1}, \"pages\": 1, \"size\": 100 }}".format(graph_data_length, graph_data_length)
counts = {}
hit_list = []
config = load_config(file_name='config.json')
for row in bar_graph_data:
bar_segment_tip = row[config['search']['barchart-config']['data-tooltip-label']]
x_labels = row[config['search']['barchart-config']['x-axis-labels']]
n_files = row["file_count"]
n_cases = row["case_count"]
if x_labels is None:
x_labels = "None"
if x_labels in counts:
counts[x_labels] = counts[x_labels] + 1
else:
counts[x_labels] = 1
hit_list.append({"x_axis" : x_labels, "bar_segment_tip": bar_segment_tip , "summary": { "case_count": n_cases, "file_count": n_files} })
buckets_list = []
for ckey in counts:
ccount = counts[ckey]
buckets_list.append({ "key": ckey, "doc_count": ccount})
buckets_str = json.dumps(buckets_list)
hit_str = json.dumps(hit_list)
agg_str = "{{ \"x_labels\": {{ \"buckets\": {0} }}}}".format(buckets_str)
data = ("{{\"data\" : {{\"aggregations\": {0}, \"hits\" : {1}, \"pagination\": {0}}}, \"warnings\": {{}}}}".format(agg_str, hit_str, p_str))
return make_json_response(data)
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()
from config_utils import load_config
config = load_config()
def main():
import pandas as pd
date_range = pd.date_range(config['mld_holte']['start'],
config['mld_holte']['end'],
freq="1D")
for func_name in config['mld_holte']['functions']:
func = globals()[func_name]
func(date_range)
def download(*args):
from utils import download_utils as dl
from pandas import date_range as pd_date_range
import os
remote_path = "http://mixedlayer.ucsd.edu/data/Argo_mixedlayers_monthlyclim_05092018.nc"
local_path = os.path.join(
config['path']['raw'],
"mld_holte/Argo_mixedlayers_monthlyclim_05092018.nc")
dl.download_wget(remote_path, local_path, verbose=True)
def preprocessor_mld_holte_025d(xds, date=None):
from utils import xarray_tools as xt
import astropy.convolution as conv
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 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, 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 setup ():
global app
log.debug("Loading global app settings...")
app = config_utils.load_config(paths.config_path(MAINFILE), paths.app_path(MAINSPEC))
def setup(): global appconfig appconfig = config_utils.load_config(MAINFILE, MAINSPEC)
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 get_nas_bench_loaders(workers):
torch.set_num_threads(workers)
root_dir = (pathlib.Path(__file__).parent / '..' / '..').resolve()
torch_dir = pathlib.Path(os.environ['TORCH_HOME'])
# cifar
cifar_config_path = root_dir / 'configs' / 'nas-benchmark' / 'CIFAR.config'
cifar_config = load_config(cifar_config_path, None, None)
get_datasets = datasets.get_datasets # a function to return the dataset
break_line = '-' * 150
print('{:} Create data-loader for all datasets'.format(time_string()))
print(break_line)
TRAIN_CIFAR10, VALID_CIFAR10, xshape, class_num = get_datasets(
'cifar10', str(torch_dir / 'cifar.python'), -1)
print(
'original CIFAR-10 : {:} training images and {:} test images : {:} input shape : {:} number of classes'
.format(len(TRAIN_CIFAR10), len(VALID_CIFAR10), xshape, class_num))
cifar10_splits = load_config(
root_dir / 'configs' / 'nas-benchmark' / 'cifar-split.txt', None, None)
assert cifar10_splits.train[:10] == [
0, 5, 7, 11, 13, 15, 16, 17, 20, 24
] and cifar10_splits.valid[:10] == [1, 2, 3, 4, 6, 8, 9, 10, 12, 14]
temp_dataset = copy.deepcopy(TRAIN_CIFAR10)
temp_dataset.transform = VALID_CIFAR10.transform
# data loader
trainval_cifar10_loader = torch.utils.data.DataLoader(
TRAIN_CIFAR10,
batch_size=cifar_config.batch_size,
shuffle=True,
num_workers=workers,
pin_memory=True)
train_cifar10_loader = torch.utils.data.DataLoader(
TRAIN_CIFAR10,
batch_size=cifar_config.batch_size,
sampler=torch.utils.data.sampler.SubsetRandomSampler(
cifar10_splits.train),
num_workers=workers,
pin_memory=True)
valid_cifar10_loader = torch.utils.data.DataLoader(
temp_dataset,
batch_size=cifar_config.batch_size,
sampler=torch.utils.data.sampler.SubsetRandomSampler(
cifar10_splits.valid),
num_workers=workers,
pin_memory=True)
test__cifar10_loader = torch.utils.data.DataLoader(
VALID_CIFAR10,
batch_size=cifar_config.batch_size,
shuffle=False,
num_workers=workers,
pin_memory=True)
print('CIFAR-10 : trval-loader has {:3d} batch with {:} per batch'.format(
len(trainval_cifar10_loader), cifar_config.batch_size))
print('CIFAR-10 : train-loader has {:3d} batch with {:} per batch'.format(
len(train_cifar10_loader), cifar_config.batch_size))
print('CIFAR-10 : valid-loader has {:3d} batch with {:} per batch'.format(
len(valid_cifar10_loader), cifar_config.batch_size))
print('CIFAR-10 : test--loader has {:3d} batch with {:} per batch'.format(
len(test__cifar10_loader), cifar_config.batch_size))
print(break_line)
# CIFAR-100
TRAIN_CIFAR100, VALID_CIFAR100, xshape, class_num = get_datasets(
'cifar100', str(torch_dir / 'cifar.python'), -1)
print(
'original CIFAR-100: {:} training images and {:} test images : {:} input shape : {:} number of classes'
.format(len(TRAIN_CIFAR100), len(VALID_CIFAR100), xshape, class_num))
cifar100_splits = load_config(
root_dir / 'configs' / 'nas-benchmark' / 'cifar100-test-split.txt',
None, None)
assert cifar100_splits.xvalid[:10] == [
1, 3, 4, 5, 8, 10, 13, 14, 15, 16
] and cifar100_splits.xtest[:10] == [0, 2, 6, 7, 9, 11, 12, 17, 20, 24]
train_cifar100_loader = torch.utils.data.DataLoader(
TRAIN_CIFAR100,
batch_size=cifar_config.batch_size,
shuffle=True,
num_workers=workers,
pin_memory=True)
valid_cifar100_loader = torch.utils.data.DataLoader(
VALID_CIFAR100,
batch_size=cifar_config.batch_size,
sampler=torch.utils.data.sampler.SubsetRandomSampler(
cifar100_splits.xvalid),
num_workers=workers,
pin_memory=True)
test__cifar100_loader = torch.utils.data.DataLoader(
VALID_CIFAR100,
batch_size=cifar_config.batch_size,
sampler=torch.utils.data.sampler.SubsetRandomSampler(
cifar100_splits.xtest),
num_workers=workers,
pin_memory=True)
print('CIFAR-100 : train-loader has {:3d} batch'.format(
len(train_cifar100_loader)))
print('CIFAR-100 : valid-loader has {:3d} batch'.format(
len(valid_cifar100_loader)))
print('CIFAR-100 : test--loader has {:3d} batch'.format(
len(test__cifar100_loader)))
print(break_line)
imagenet16_config_path = 'configs/nas-benchmark/ImageNet-16.config'
imagenet16_config = load_config(imagenet16_config_path, None, None)
TRAIN_ImageNet16_120, VALID_ImageNet16_120, xshape, class_num = get_datasets(
'ImageNet16-120', str(torch_dir / 'cifar.python' / 'ImageNet16'), -1)
print(
'original TRAIN_ImageNet16_120: {:} training images and {:} test images : {:} input shape : {:} number of classes'
.format(len(TRAIN_ImageNet16_120), len(VALID_ImageNet16_120), xshape,
class_num))
imagenet_splits = load_config(
root_dir / 'configs' / 'nas-benchmark' /
'imagenet-16-120-test-split.txt', None, None)
assert imagenet_splits.xvalid[:10] == [
1, 2, 3, 6, 7, 8, 9, 12, 16, 18
] and imagenet_splits.xtest[:10] == [0, 4, 5, 10, 11, 13, 14, 15, 17, 20]
train_imagenet_loader = torch.utils.data.DataLoader(
TRAIN_ImageNet16_120,
batch_size=imagenet16_config.batch_size,
shuffle=True,
num_workers=workers,
pin_memory=True)
valid_imagenet_loader = torch.utils.data.DataLoader(
VALID_ImageNet16_120,
batch_size=imagenet16_config.batch_size,
sampler=torch.utils.data.sampler.SubsetRandomSampler(
imagenet_splits.xvalid),
num_workers=workers,
pin_memory=True)
test__imagenet_loader = torch.utils.data.DataLoader(
VALID_ImageNet16_120,
batch_size=imagenet16_config.batch_size,
sampler=torch.utils.data.sampler.SubsetRandomSampler(
imagenet_splits.xtest),
num_workers=workers,
pin_memory=True)
print('ImageNet-16-120 : train-loader has {:3d} batch with {:} per batch'.
format(len(train_imagenet_loader), imagenet16_config.batch_size))
print('ImageNet-16-120 : valid-loader has {:3d} batch with {:} per batch'.
format(len(valid_imagenet_loader), imagenet16_config.batch_size))
print('ImageNet-16-120 : test--loader has {:3d} batch with {:} per batch'.
format(len(test__imagenet_loader), imagenet16_config.batch_size))
# 'cifar10', 'cifar100', 'ImageNet16-120'
loaders = {
'cifar10@trainval': trainval_cifar10_loader,
'cifar10@train': train_cifar10_loader,
'cifar10@valid': valid_cifar10_loader,
'cifar10@test': test__cifar10_loader,
'cifar100@train': train_cifar100_loader,
'cifar100@valid': valid_cifar100_loader,
'cifar100@test': test__cifar100_loader,
'ImageNet16-120@train': train_imagenet_loader,
'ImageNet16-120@valid': valid_imagenet_loader,
'ImageNet16-120@test': test__imagenet_loader
}
return loaders
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('--prefix_path',
type=str,
required=True,
help='Prefix path to the saved model')
parser.add_argument('--in_path',
type=str,
required=True,
help='Path to the input file.')
parser.add_argument('--out_path',
type=str,
default=None,
help='Path to the output file.')
args, unparsed = parser.parse_known_args()
FLAGS = config_utils.load_config(args.prefix_path + ".config.json")
if FLAGS.model_type == 'bert_word':
import zp_datastream
import zp_model
elif FLAGS.model_type == 'bert_char':
import zp_datastream_char as zp_datastream
import zp_model_char as zp_model
else:
assert False, "model_type '{}' not supported".format(FLAGS.model_type)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
n_gpu = torch.cuda.device_count()
print('device: {}, n_gpu: {}, grad_accum_steps: {}'.format(
device, n_gpu, FLAGS.grad_accum_steps))
def setup ():
global app
log.debug("Loading global app settings...")
app = config_utils.load_config(os.path.join(storage.data_directory, MAINFILE), os.path.join(paths.app_path(), MAINSPEC))
