def main():
args = utils.bin_config(get_arguments)
if args.multi_gpus: # use all GPUs in parallel
assert torch.cuda.is_available()
args.n_gpus = torch.cuda.device_count()
os.environ['MASTER_ADDR'] = args.master_addr
os.environ['MASTER_PORT'] = str(args.master_port)
elif args.gpu_id != -1:
args.n_gpus = 1
else:
args.n_gpus = 0
config = json.load(open(args.config_file))
assert config['config_target'] == 'naive_psychology'
# if args.mode == 'train':
prefix = 'semisupervised' if args.is_semisupervised else 'unsupervised'
train_inputs = get_dataset_inputs(prefix, args)
dev_inputs = get_dataset_inputs('dev', args)
if args.n_gpus > 1:
mp.spawn(train,
nprocs=args.n_gpus,
args=(train_inputs, dev_inputs, config, args))
else:
train(args.gpu_id, train_inputs, dev_inputs, config, args)
def load_dataset_inputs(prefix):
input_f = os.path.join(args.input_dir, '{}_inputs.h5'.format(prefix))
return input_f
def main():
config = json.load(open(args.config_file))
assert config['config_target'] == 'naive_psychology'
set_seed(
args.gpu_id, args.seed
) # in distributed training, this has to be same for all processes
if args.mode == 'train':
train_dinputs = load_dataset_inputs('train')
dev_dinputs = load_dataset_inputs('dev')
train(train_dinputs, dev_dinputs)
# test
test_dinputs = load_dataset_inputs('test')
test(test_dinputs, args.output_dir)
else: # test
test_dinputs = load_dataset_inputs('test')
test(test_dinputs, args.from_checkpoint)
if __name__ == "__main__":
args = utils.bin_config(get_arguments)
logger = utils.get_root_logger(args)
main()