def run(m_idx, run_idx, params=params, list_mode='all'):
dset = 0
print('model number %s' % m_idx)
print('model name %s' % gm.model_list[m_idx].__name__)
for dl_train, dl_test in zip(train, test):
print('dset %s;' % dset)
t0 = time.time()
path = os.path.join(save_path, 'run%s' % run_idx, 'model' + str(m_idx))
pathlib.Path(os.path.join(path, 'data')).mkdir(parents=True,
exist_ok=True)
pathlib.Path(os.path.join(path, 'models')).mkdir(parents=True,
exist_ok=True)
for seed in seeds:
np.random.seed(seed)
torch.manual_seed(seed)
model = gm.model_list[m_idx](*params)
opt = torch.optim.Adam(model.parameters(), lr=lr)
one_run(dset,
seed,
n_epochs,
model,
opt,
dl_train,
dl_test,
path,
cuda=False,
list_mode=list_mode)
t = time.time()
print('total running time for one ds %s seconds' % str(t - t0))
dset += 1
def run(m_idx, run_idx):
dset = 0
print('model number %s' % m_idx)
print('model name %s' % gm.model_list_double[m_idx].__name__)
for dpath_train, dpath_val in zip(train_5, val_5):
print('dset %s;' % dset)
t0 = time.time()
dl_train = load_dl(
os.path.join(prefix, dpath_train), double=True)
dl_val = load_dl(
os.path.join(prefix, dpath_val), double=True)
path = os.path.join(
args.directory, 'run%s' % run_idx, 'model' + str(m_idx))
pathlib.Path(
os.path.join(path, 'data')).mkdir(parents=True, exist_ok=True)
pathlib.Path(
os.path.join(path, 'models')).mkdir(parents=True, exist_ok=True)
for seed in seeds:
np.random.seed(seed)
torch.manual_seed(seed)
model = gm.model_list_double[m_idx](*params)
opt = torch.optim.Adam(model.parameters(), lr=lr)
one_run(
dset,
seed,
n_epochs,
model,
opt,
dl_train,
dl_val,
path,
cuda=False)
t = time.time()
print('total running time for one ds %s seconds' % str(t - t0))
dset += 1
def cur_run(m_idx, run_idx, params=params):
"""
For using a curriculum of dataloaders.
"""
dlist = [load_dl(os.path.join(cur_prefix, p), double=True) \
for p in train_cur]
dl_val = load_dl(os.path.join(cur_prefix, val), double=True)
dset = 0
path = os.path.join(
args.directory, 'run%s' % run_idx, 'model' + str(m_idx))
pathlib.Path(
os.path.join(path, 'data')).mkdir(parents=True, exist_ok=True)
pathlib.Path(
os.path.join(path, 'models')).mkdir(parents=True, exist_ok=True)
for seed in seeds:
t0 = time.time()
np.random.seed(seed)
torch.manual_seed(seed)
model = gm.model_list_double[m_idx](*params)
opt = torch.optim.Adam(model.parameters(), lr=lr)
# res = model(g1, g2)
one_run(
dset,
seed,
n_epochs,
model,
opt,
dlist,
dl_val,
path,
cuda=False)
t = time.time()
print('total running time for one seed %s seconds' % str(t - t0))
opt = torch.optim.Adam(model.parameters(), lr=hparams['lr'])
mpath = op.join(path, m_str)
pathlib.Path(op.join(mpath, 'data')).mkdir(parents=True,
exist_ok=True)
if not preload_model:
pathlib.Path(op.join(mpath, 'models')).mkdir(parents=True,
exist_ok=True)
else:
srcmpath = op.join(srcpath, m_str, 'models')
copytree(srcmpath, op.join(mpath, 'models'))
one_run(i,
seed,
hparams['n_epochs'],
model,
opt,
train_dls,
test_dls,
mpath,
cuda=cuda,
n_obj=n_obj,
preload=preload_model)
log(logfile,
'run completed, results saved in {}\n'.format(mpath))
log(logfile,
'training time for one seed %s\n' % (time.time() - t0))
# close log file
log(
logfile,
'finished experiment {} at {}'.format(expe_idx,
str(datetime.datetime.now())))
logfile.close()