sample.yml

# WARNING:
# Do not use this file as is! It only lists and documents available options
# without caring about their consistency. Prefer using predefined yaml files
# in reporoot/default_configs/ as a base configuration and specialize it with
# additional configuration files.

# The directory where models and summaries will be saved. It is created if it does not exist.
# for all mode, model_dir should be provided, by default: models
model_dir: models

# training options
train:
  # for pseudo Multi-GPU training, each step goes through `update_cycle` * batch_(tokens_)size samples
  update_cycle: 1
  # batch size, the number of sentences passed into one step, by default: 80
  batch_size: 80
  # the number of words for each batch, by default: None
  # sentence pairs will be batched together by approximate sequence length
  batch_tokens_size:
  # Save a checkpoint every this many steps. by default: 1000
  save_checkpoint_steps: 1000
  # Train for this many steps. If None, training forever. by default: 10000000
  train_steps: 10000000
  # Evaluate and save summaries every this many steps and display. by default: 100
  eval_steps: 100
  # The maximum length of feature sequences during training. by default: None
  maximum_features_length: 50
  # The maximum length of label sequences during training. by default: None
  maximum_labels_length: 50
  # whether to reverse the target side to train the model. by defalt: False
  reverse_target: False
  # whether to shuffle data between epochs, if provided,
  # use it as postfix of the shuffled data, by default: None
  shuffle_every_epoch:

# training and evaluating data
data:
  # training features file, by default: None
  train_features_file: testdata/toy.zh
  # training labels file, by default: None
  train_labels_file: testdata/toy.en0
  # evaluating features file, by default: None
  eval_features_file: testdata/toy.zh
  # evaluating labels file, by default: None
  # if there are more than one reference, only provide the prefix filename, the program
  # will automatically recognize the multi references by "eval_labels_file"0, "eval_labels_file"1, ...
  eval_labels_file: testdata/test.en
  # source side vocabluary, by default: None
  source_words_vocabulary: testdata/vocab.zh
  # target side vocabulary, by default: None
  target_words_vocabulary: testdata/vocab.en
  # source side BPE codes, if provided, BPE with be applied to word tokens
  # in features_files. by default: None
  # Note that the source_words_vocabulary must be generated after applying BPE.
  source_bpecodes:
  # target side BPE codes, if provided, BPE with be applied to word tokens
  # in labels_files. by default: None
  # Note that the target_words_vocabulary must be generated after applying BPE.
  target_bpecodes:

# auxiliary training hooks, by default: empty list
# Now, there is no training hook implemented yet.
hooks: []

# list of evaluation metrics on eval data (eval_features_file and eval_labels_file)
#  Now we only provide LossMetricSpec and BleuMetricSpec
metrics:
  - class: LossMetricSpec  # evaluate using loss function on "eval_labels_file" or "eval_labels_file"0
    params:
      start_at: 0 # start to evaluate from this step, by default: 0
      eval_steps: 100 # evalute every this steps, by default: 100
      batch_size: 128  # evaluation batch size, by default: 128
      do_summary: true  # whether make summaries for tensorboard, by default: true

  - class: BleuMetricSpec # evaluate using multi-bleu.perl (do beam search inference)
    params:
      start_at: 10000  # start to evaluate from this step, by default: 0
      eval_steps: 1000  # evaluate every this steps, by default: 1000
      batch_size: 32  # inference beam size, by default: None
      beam_size: 4 # inference beam size, if None, inherit from training model parameters, by default: None
      maximum_labels_length: 150 # maximum generation length, if None, inherit from training model parameters, by default: None
      length_penalty: -1.0 # inference length penalty, if None, inherit from training model parameters, by default: None
      delimiter: " "  # output delimiter, by default: " "(space)
      char_level: false  # whether output in char-level, by default: false
      maximum_keep_models: 5  # maximum keeping checkpoints in tar.gz format, by default: 5
      early_step: true  # whether to use BLEU to do early stop, by default: true
      estop_patience: 30  # the maximum patience for early stop
      do_summary: true  # whether make summaries for tensorboard, by default: true

# optimizer parameters
optimizer_params:
  optimizer.name: Adam  # optimizer name, by default: Adam
  optimizer.params: # optimizer-specific parameters
    epsilon: 0.0000008
  optimizer.learning_rate: 0.0005  # learning rate
  optimizer.clip_gradients: 1.0 # clip gradient norm
  # parameters for learning rate decaying scheme
  optimizer.lr_decay:
    # decaying function name, can be *_decay functions in tf.train, or loss_decay / noam_decay
    #   by default: None
    decay_type:
    # minimum learning learning rate: 1.0e-5
    min_learning_rate: 0.00005
    # see tf.train.*_decay for more details for the rest parameters
    decay_steps: 100  # by default: 100, but for noam_decay, this means the "warmup steps"
    decay_rate: 0.99  # by default: 0.99
    start_decay_at: 0   # by default: 0
    stop_decay_at: 1000000000 # by default: sys.maxsize
    staircase: false  # by default: false
    # parameter for noam_decay, by default: None
    dmodel: 512
    # parameter for noam_decay, by default: 10.0
    scale: 2.0
    # parameter for loss_decay, by default: None
    patience: 30

# options for inference or mutual
infer:
  # source side vocabluary, by default: None
  source_words_vocabulary: testdata/vocab.zh
  # target side vocabulary, by default: None
  target_words_vocabulary: testdata/vocab.en
  # source side BPE codes, if provided, BPE with be applied to word tokens
  # in features_files. by default: None
  # Note that the source_words_vocabulary must be generated after applying BPE.
  source_bpecodes:
  # target side BPE codes, if provided, BPE with be applied to word tokens
  # in labels_files. by default: None
  # Note that the target_words_vocabulary must be generated after applying BPE.
  target_bpecodes:
  # inference batch size, by default: 32
  batch_size: 32
  # inference beam size, by default: 10
  beam_size: 10
  # The maximum length of label sequences for inference. by default: 150
  maximum_labels_length: 150
  # length penalty, by default: -1.0
  length_penalty: -1.0
  # inference output delimiter, by default: " " (one space)
  delimiter: " "
  # output in charactor level, for inference only, by default: false
  char_level: false

# testdata for inference
# list of testsets
infer_data:
    # source features file, line by line, by default: None
  - features_file: source_file1
    # target labels file, line by line according to features file,
    # if provided, use multibleu_script to compute BLEU score, by default: None
    labels_file: reference1
    # output file for translations, line by line, by default: None
    output_file: translation_output1
     # whether to output attention score in a specific format for seq2seq/tools/plot_heatmap.py, by default: false
    output_attention: false
  - features_file: source_file2
    labels_file: reference2
    output_file: translation_output2
    output_attention: true


# network parameters
# for more details, see the example yaml configurations