def main():
# get config and arguments
args, config = get_downstream_args()
# Fix seed and make backends deterministic
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
if torch.cuda.is_available(): torch.cuda.manual_seed_all(args.seed)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = True
# mkdir
if args.expdir == '':
if args.name is None: args.name = 'exp_' + str(random.randint(0, 999))
expdir = os.path.join(
'result/result_' + args.upstream + '_' + args.task + '/',
args.name)
else:
expdir = args.expdir
if not os.path.exists(expdir):
os.makedirs(expdir)
copyfile(args.config, os.path.join(expdir, args.config.split('/')[-1]))
# get upstream model
upstream_model = get_upstream_model(
args) ######### plug in your upstream pre-trained model here #########
# get dataloaders
train_loader, dev_loader, test_loader = get_dataloader(
args, config['dataloader'])
# get downstream model
downstream_model = get_downstream_model(args, upstream_model.out_dim,
train_loader.dataset.class_num,
config)
# train
runner = Runner(args=args,
runner_config=config['runner'],
dataloader={
'train': train_loader,
'dev': dev_loader,
'test': test_loader
},
upstream=upstream_model,
downstream=downstream_model,
expdir=expdir)
runner.set_model()
runner.train()
runner.save_results(config['runner']['save_file'])
def main():
torch.multiprocessing.set_sharing_strategy('file_system')
torchaudio.set_audio_backend('sox_io')
hack_isinstance()
# get config and arguments
args, config, backup_files = get_downstream_args()
if args.cache_dir is not None:
torch.hub.set_dir(args.cache_dir)
# When torch.distributed.launch is used
if args.local_rank is not None:
torch.cuda.set_device(args.local_rank)
torch.distributed.init_process_group(args.backend)
if args.mode == 'train' and args.past_exp:
ckpt = torch.load(args.init_ckpt, map_location='cpu')
now_use_ddp = is_initialized()
original_use_ddp = ckpt['Args'].local_rank is not None
assert now_use_ddp == original_use_ddp, f'{now_use_ddp} != {original_use_ddp}'
if now_use_ddp:
now_world = get_world_size()
original_world = ckpt['WorldSize']
assert now_world == original_world, f'{now_world} != {original_world}'
# Save command
if is_leader_process():
with open(os.path.join(args.expdir, f'args_{get_time_tag()}.yaml'),
'w') as file:
yaml.dump(vars(args), file)
with open(os.path.join(args.expdir, f'config_{get_time_tag()}.yaml'),
'w') as file:
yaml.dump(config, file)
for file in backup_files:
backup(file, args.expdir)
# Fix seed and make backends deterministic
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
if torch.cuda.is_available(): torch.cuda.manual_seed_all(args.seed)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
runner = Runner(args, config)
eval(f'runner.{args.mode}')()
def main():
# get config and arguments
args, config = get_downstream_args()
# Fix seed and make backends deterministic
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
if torch.cuda.is_available(): torch.cuda.manual_seed_all(args.seed)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = True
runner = Runner(args, config)
eval(f'runner.{args.mode}')()
runner.logger.close()