$ python ./data/build_corpus.py -h
usage: build_corpus.py [-h] --input INPUT --lang LANG --output OUTPUT
                       [--valid_ratio VALID_RATIO] [--test_ratio TEST_RATIO]
                       [--no_shuffle]

$ ls ./data
corpus.en  corpus.ko
$ python ./data/build_corpus.py --input ./data/corpus --lang enko --output ./data/corpus
total src lines: 384105
total tgt lines: 384105
write 7682 lines to ./data/corpus.valid.en
write 7682 lines to ./data/corpus.valid.ko
write 376423 lines to ./data/corpus.train.en
write 376423 lines to ./data/corpus.train.ko

$ python train.py -h
usage: train.py [-h] --model MODEL --train TRAIN --valid VALID --lang LANG
                [--gpu_id GPU_ID] [--batch_size BATCH_SIZE]
                [--n_epochs N_EPOCHS] [--print_every PRINT_EVERY]
                [--early_stop EARLY_STOP] [--max_length MAX_LENGTH]
                [--dropout DROPOUT] [--word_vec_dim WORD_VEC_DIM]
                [--hidden_size HIDDEN_SIZE] [--n_layers N_LAYERS]
                [--max_grad_norm MAX_GRAD_NORM] [--adam] [--lr LR]
                [--min_lr MIN_LR] [--lr_decay_start_at LR_DECAY_START_AT]
                [--lr_slow_decay] [--lr_decay_rate LR_DECAY_RATE]
                [--rl_lr RL_LR] [--n_samples N_SAMPLES]
                [--rl_n_epochs RL_N_EPOCHS] [--rl_n_gram RL_N_GRAM]

optional arguments:
  -h, --help            show this help message and exit
  --model MODEL         Model file name to save. Additional information would
                        be annotated to the file name.
  --train TRAIN         Training set file name except the extention. (ex:
                        train.en --> train)
  --valid VALID         Validation set file name except the extention. (ex:
                        valid.en --> valid)
  --lang LANG           Set of extention represents language pair. (ex: en +
                        ko --> enko)
  --gpu_id GPU_ID       GPU ID to train. Currently, GPU parallel is not
                        supported. -1 for CPU. Default=-1
  --batch_size BATCH_SIZE
                        Mini batch size for gradient descent. Default=32
  --n_epochs N_EPOCHS   Number of epochs to train. Default=18
  --print_every PRINT_EVERY
                        Number of gradient descent steps to skip printing the
                        training status. Default=1000
  --early_stop EARLY_STOP
                        The training will be stopped if there is no
                        improvement this number of epochs. Default=-1
  --max_length MAX_LENGTH
                        Maximum length of the training sequence. Default=80
  --dropout DROPOUT     Dropout rate. Default=0.2
  --word_vec_dim WORD_VEC_DIM
                        Word embedding vector dimension. Default=512
  --hidden_size HIDDEN_SIZE
                        Hidden size of LSTM. Default=768
  --n_layers N_LAYERS   Number of layers in LSTM. Default=4
  --max_grad_norm MAX_GRAD_NORM
                        Threshold for gradient clipping. Default=5.0
  --adam                Use Adam instead of using SGD.
  --lr LR               Initial learning rate. Default=1.0
  --min_lr MIN_LR       Minimum learning rate. Default=.000001
  --lr_decay_start_at LR_DECAY_START_AT
                        Start learning rate decay from this epoch.
  --lr_slow_decay       Decay learning rate only if there is no improvement on
                        last epoch.
  --lr_decay_rate LR_DECAY_RATE
                        Learning rate decay rate. Default=0.5
  --rl_lr RL_LR         Learning rate for reinforcement learning. Default=.01
  --n_samples N_SAMPLES
                        Number of samples to get baseline. Default=1
  --rl_n_epochs RL_N_EPOCHS
                        Number of epochs for reinforcement learning.
                        Default=10
  --rl_n_gram RL_N_GRAM
                        Maximum number of tokens to calculate BLEU for
                        reinforcement learning. Default=6

$ python train.py --model ./models/enko.pth --train ./data/corpus.train --valid ./data/corpus.valid --lang enko --gpu_id 0 --word_vec_dim 256 --hidden_size 512 --batch_size 32 --n_epochs 15 --rl_n_epochs 10 --early_stop -1

$ python translate.py -h
usage: translate.py [-h] --model MODEL [--gpu_id GPU_ID]
                    [--batch_size BATCH_SIZE] [--max_length MAX_LENGTH]
                    [--n_best N_BEST] [--beam_size BEAM_SIZE]

optional arguments:
  -h, --help            show this help message and exit
  --model MODEL         Model file name to use
  --gpu_id GPU_ID       GPU ID to use. -1 for CPU. Default=-1
  --batch_size BATCH_SIZE
                        Mini batch size for parallel inference. Default=128
  --max_length MAX_LENGTH
                        Maximum sequence length for inference. Default=255
  --n_best N_BEST       Number of best inference result per sample. Default=1
  --beam_size BEAM_SIZE
                        Beam size for beam search. Default=5

$ python translate.py --model ./model/enko.12.1.18-3.24.1.37-3.92.pth --gpu_id 0 --batch_size 128 --beam_size 5