def train_model():
'''Run for several hyper parameters'''
EPOCH = 300
parser = local_cifar_parser()
args = parser.parse_args()
fim = FileIdManager(ID_STRUCTURE)
FILE_ID = fim.get_id_from_args(args)
SAVE_DIR, LOG_DIR, CSV_DIR, PKL_DIR, BOARD_DIR, ASSET_DIR = subdirs5resultdir(
RESULT_DIR, True)
SAVE_SUBDIR, PKL_SUBDIR, BOARD_SUBDIR, ASSET_SUBDIR, CSV_SUBDIR = muldir2mulsubdir(
[SAVE_DIR, PKL_DIR, BOARD_DIR, ASSET_DIR, CSV_DIR], FILE_ID, True)
# load data
dm_train, dm_val, dm_test = cifar_manager(dm_type=args.ltype,
nsclass=args.nsclass)
for v in [dm_train, dm_val, dm_test]:
v.print_shape()
model = Model(dm_train, dm_val, dm_test, LOG_DIR + FILE_ID + '.log', args)
model.build()
model.set_up_train()
try:
model.restore(save_dir=SAVE_SUBDIR)
except (AttributeError, TypeError):
model.initialize()
model.train(epoch=EPOCH, save_dir=SAVE_SUBDIR, board_dir=BOARD_SUBDIR)
model.restore(save_dir=SAVE_SUBDIR)
model.prepare_test()
content = model.test_metric(K_SET)
write_pkl(content, path=PKL_SUBDIR + 'evaluation.pkl')
write_dict_csv(dict_=content, path=CSV_SUBDIR + 'evaluation.csv')
def integrate_results_and_preprocess():
SAVE_DIR, LOG_DIR, CSV_DIR, PKL_DIR, BOARD_DIR, ASSET_DIR = subdirs5resultdir(
RESULT_DIR, False)
BOUNDARY = 10
FILE_KEY = 'evaluation.pkl'
def get_value(path):
content = read_pkl(path)
return np.sum(content['te_te_precision_at_k'])
max_value = -1
max_file = None
for file in sorted(os.listdir(PKL_DIR)):
PKL_SUBDIR = PKL_DIR + '{}/'.format(file)
path = PKL_SUBDIR + '{}'.format(FILE_KEY)
if os.path.exists(path):
value = get_value(path)
if max_value < value:
max_value = value
max_file = file
# Get the best file id
parser = local_cifar_parser()
args = parser.parse_args()
fim = FileIdManager(ID_STRUCTURE)
fim.update_args_with_id(args, max_file) # update args value
FILE_ID = fim.get_id_from_args(args)
SAVE_SUBDIR, PKL_SUBDIR, BOARD_SUBDIR, ASSET_SUBDIR, CSV_SUBDIR = muldir2mulsubdir(
[SAVE_DIR, PKL_DIR, BOARD_DIR, ASSET_DIR, CSV_DIR], FILE_ID, False)
# load data
dm_train, dm_val, dm_test = cifar_manager(dm_type=args.ltype,
nsclass=args.nsclass)
for v in [dm_train, dm_val, dm_test]:
v.print_shape()
model = Model(dm_train, dm_val, dm_test, './test.log', args)
model.build()
model.set_up_train()
model.restore(save_dir=SAVE_SUBDIR)
model.prepare_test()
model.prepare_test2()
META_DIR = RESULT_DIR + 'meta/'
BESTSAVE_DIR = RESULT_DIR + 'bestsave/'
# copy file
if os.path.isdir(BESTSAVE_DIR): remove_dir(BESTSAVE_DIR)
copy_dir(SAVE_SUBDIR, BESTSAVE_DIR)
# ======================================================#
create_muldir(META_DIR)
store_content = {
'train_embed': model.train_embed,
'test_embed': model.test_embed,
'val_embed': model.val_embed,
'te_te_distance': model.te_te_distance,
'te_tr_distance': model.te_tr_distance,
'val_arg_sort': model.val_arg_sort,
}
for v in store_content.keys():
print("{} : {}".format(v, store_content[v].shape))
write_pkl(store_content, META_DIR + 'meta.pkl')
def train_model():
EPOCH = 300
nk = len(K_SET)
parser = local_cifar_parser()
args = parser.parse_args() # parameter required for model
fim = FileIdManager(ID_STRUCTURE)
FILE_ID = fim.get_id_from_args(args)
SAVE_DIR, LOG_DIR, CSV_DIR, PKL_DIR, BOARD_DIR, ASSET_DIR = subdirs5resultdir(
RESULT_DIR, True)
SAVE_SUBDIR, PKL_SUBDIR, BOARD_SUBDIR, ASSET_SUBDIR, CSV_SUBDIR = muldir2mulsubdir(
[SAVE_DIR, PKL_DIR, BOARD_DIR, ASSET_DIR, CSV_DIR], FILE_ID, True)
dm_train, dm_val, dm_test = cifar_manager(dm_type=args.ltype,
nsclass=args.nsclass)
for v in [dm_train, dm_val, dm_test]:
v.print_shape()
model = Model(dm_train, dm_val, dm_test, LOG_DIR + FILE_ID + '.log', args)
model.build()
meta = read_pkl(args.meta)
model.restore(args.save)
model.set_info(val_arg_sort=meta['val_arg_sort'],
te_te_distance=meta['te_te_distance'],
te_tr_distance=meta['te_tr_distance'])
model.build_hash()
model.set_up_train_hash()
try:
model.restore(save_dir=SAVE_SUBDIR)
except (AttributeError, TypeError):
model.initialize()
model.train_hash(epoch=EPOCH,
save_dir=SAVE_SUBDIR,
board_dir=BOARD_SUBDIR)
model.restore(save_dir=SAVE_SUBDIR)
model.prepare_test_hash()
performance = model.test_hash_metric(K_SET)
model.delete()
del model
del dm_train
del dm_val
del dm_test
write_pkl(performance, path=PKL_SUBDIR + 'evaluation.pkl')
cwrite = CsvWriter2(1)
key_set = [
'train_nmi', 'test_nmi', 'te_tr_suf', 'te_te_suf',
'te_te_precision_at_k', 'te_tr_precision_at_k'
]
for key in key_set:
cwrite.add_header(0, str(key))
content = ''
if 'at_k' in str(key): content = listformat(performance[key])
else: content = performance[key]
cwrite.add_content(0, content)
cwrite.write(CSV_SUBDIR + 'evaluation.csv')
NITER = 140000
# PITER = 20000
PITER = 10000
SITER = 10000
parser = local_dsprites_parser()
args = parser.parse_args() # parameter required for model
ROOT_DIRS = {
'dsprites': '/mnt/hdd/repo_results/DisentanglementICML19/Results_dsprites/',
'3dshapes': '/mnt/hdd/repo_results/DisentanglementICML19/Results_3dshapes/'
}
ROOT = ROOT_DIRS[args.dataset]
RESULT_DIR = ROOT+'{}/'.format(KEY)
fim = FileIdManager(ID_STRUCTURE)
np.random.seed(args.rseed)
FILE_ID = fim.get_id_from_args(args)
SAVE_DIR, LOG_DIR, ASSET_DIR = subdirs5resultdir(RESULT_DIR, True)
SAVE_SUBDIR, ASSET_SUBDIR = muldir2mulsubdir([SAVE_DIR, ASSET_DIR], FILE_ID, True)
if args.dataset == 'dsprites':
dm = dsprites_manager()
else:
dm = shapes_3d_manager()
dm.print_shape()
model = Model(dm, LOG_DIR+FILE_ID+'.log', args)
model.set_up_train()
model.initialize()