# Arguments for training

parser.add_argument("--n_epoch", default=10, type=int, help="Max n epoch during dual training.")

parser.add_argument("--learning_rate", default=0.00001, type=float, help="Learning rate")

parser.add_argument("--change_per_step", default=10, type=int, help="Change dual training direction per n steps")

parser.add_argument("--batch_size", default=64, type=int,

help="Batch size of dual training (You can enlarge the batch_size if you have enough memory).")

# Arguments for evaluation and model saving

parser.add_argument("--save_per_step", default=1000, type=int, help="Save model per n steps")

parser.add_argument("--eval_step", default=100, type=int, help="Evaluate model.")

# Arguments for calculating reward

parser.add_argument("--use_baseline", type=ast.literal_eval, default=True,

help="Use baseline in reward calculation, input should be either 'True' or 'False'.")

parser.add_argument("--normalize_reward", type=ast.literal_eval, default=False, # important !!

help="normalize reward or not, input should be either 'True' or 'False'.")

# Arguments for anneal teacher-forcing (MLE)

parser.add_argument("--teacher_forcing", nargs='+', default=["back_trans"],

help="Corpus used in teacher forcing (MLE), must in [`pseudo`, `back_trans`")

parser.add_argument("--MLE_anneal", action='store_true', help="Anneal the use of pseudo data via MLE")

parser.add_argument("--anneal_rate", default=1.1, type=float, help="The decay rate.")

parser.add_argument("--anneal_steps", default=1000, type=int, help="Increase gap")

parser.add_argument("--anneal_initial_gap", default=1, type=int, help="Initial gap value.")

parser = argparse.ArgumentParser(formatter_class=argparse.ArgumentDefaultsHelpFormatter)

add_dual_arguments(parser)

add_common_arguments(parser)

add_optimizer_arguments(parser)

parser = parser.parse_args()