This code implements multi-layer Recurrent Neural Network (RNN, LSTM, and GRU) for training/sampling from character-level language models.
- Python 3.6
 - TensorFlow 1.4
 - hb-config
 
- Using Higher-APIs in TensorFlow
 - Korean SamhangSi (like acrostic poem)
 
example: check_tiny.yml
data:
  data_dir: 'data/tiny_lyricskor'
model:
  batch_size: 4
  input_keep_prob: 0.8
  log_dir: 'logs'
  num_layers: 1
  output_keep_prob: 0.8
  rnn_size: 64
  seq_length: 20
train:
  train_steps: 10000
  model_dir: 'tiny_checkpoints'
  save_every: 1
  learning_rate: 0.001
  loss_hook_n_iter: 100
  check_hook_n_iter: 1000
  min_eval_frequency: 100First, check if the model is valid.
python main.py --config check_tiny --mode trainThen, train the model
python main.py --config kor_ballad --mode train_and_evaluateAfter training, generate Korean Samhangsi.
python generator.py --config kor_ballad --word 삼행시- 삼행시
 
삼이야 그리움이 좇아 사랑은늘 도망가
행른 잊어버리고 그대 이 세상
시제 너의 곁을 떠나면 빗물에 꽃씨하나 흘러가듯
- 기계
 
기를 바라보네 두 손 잡고 고개 끄덕여 달라 하기에
계 울고 싶어 내 맘을 떠나가던 날
- 여름
 
여도 지금하럼 커피는 날개니
름다웠던 그대모습 다시 볼 수 없는것 알아요
- 커피
 
커나가 그래 돌아서 눈 감으면 잊을까
피고 내가 가고 싶지 아파 만날 날 기다려왔어

