Skip to content

Latest commit

 

History

History
57 lines (52 loc) · 3.58 KB

Configuration.md

File metadata and controls

57 lines (52 loc) · 3.58 KB
Raw

I/O

train_dir=xx    #string (necessary in training). Set training file directory.
dev_dir=xx    #string (necessary in training). Set dev file directory.
test_dir=xx    #string . Set test file directory.
model_dir=xx    #string (optional). Set saved model file directory.
word_emb_dir=xx    #string (optional). Set pretrained word embedding file directory.

raw_dir=xx    #string (optional). Set input raw file directory.
decode_dir=xx    #string (necessary in decoding). Set decoded file directory.
dset_dir=xx    #string (necessary). Set saved model file directory.
load_model_dir=xx    #string (necessary in decoding). Set loaded model file directory. (when decoding)
char_emb_dir=xx    #string (optional). Set pretrained character embedding file directory.

norm_word_emb=False    #boolen. If normalize the pretrained word embedding.
norm_char_emb=False    #boolen. If normalize the pretrained character embedding.
number_normalized=True    #boolen. If normalize the digit into `0` for input files.
seg=True    #boolen. If task is segmentation like, tasks with token accuracy evaluation (e.g. POS, CCG) is False; tasks with F-value evaluation(e.g. Word Segmentation, NER, Chunking) is True .
word_emb_dim=50    #int. Word embedding dimension, if model use pretrained word embedding, word_emb_dim will be reset as the same dimension as pretrained embedidng.
char_emb_dim=30    #int. Character embedding dimension, if model use pretrained character embedding, char_emb_dim will be reset as the same dimension as pretrained embedidng.

NetworkConfiguration

use_crf=True    #boolen (necessary in training). Flag of if using CRF layer. If it is set as False, then Softmax is used in inference layer.
use_char=True    #boolen (necessary in training). Flag of if using character sequence layer. 
word_seq_feature=XX    #boolen (necessary in training): CNN/LSTM/GRU. Neural structure selection for word sequence. 
char_seq_feature=CNN    #boolen (necessary in training): CNN/LSTM/GRU. Neural structure selection for character sequence, it only be used when use_char=True.
feature=[POS] emb_size=20 emb_dir=xx   #feature configuration. It includes the feature prefix [POS], pretrained feature embedding file and the embedding size. 
feature=[Cap] emb_size=20 emb_dir=xx    #feature configuration. Another feature [Cap].
nbest=1    #int (necessary in decoding). Set the nbest size during decoding.

TrainingSetting

status=train    #string: train or decode. Set the program running in training or decoding mode.
optimizer=SGD    #string: SGD/Adagrad/AdaDelta/RMSprop/Adam. optimizer selection.
iteration=1    #int. Set the iteration number of training.
batch_size=10    #int. Set the batch size of training or decoding.
ave_batch_loss=False    #boolen. Set average the batched loss during training.

Hyperparameters

cnn_layer=4    #int. CNN layer number for word sequence layer.
char_hidden_dim=50    #int. Character hidden vector dimension for character sequence layer.
hidden_dim=200    #int. Word hidden vector dimension for word sequence layer.
dropout=0.5    #float. Dropout probability.
lstm_layer=1    #int. LSTM layer number for word sequence layer.
bilstm=True    #boolen. If use bidirection lstm for word seuquence layer.
learning_rate=0.015    #float. Learning rate.
lr_decay=0.05    #float. Learning rate decay rate, only works when optimizer=SGD.
momentum=0    #float. Momentum 
l2=1e-8    #float. L2-regulization.
#gpu=True  #boolen. If use GPU, generally it depends on the hardward environment.
#clip=     #float. Clip the gradient which is larger than the setted number.