config

Config File Documentation

Please note that the suggested settings in the configuration file provided is not fine-tuned, one has to adjust the settings based on the data. The description of parameters in the configuration file:

Embedding settings

Parameter	Description	Note
data_path	The path of the code base for training the embedding model.	By default, the data for training the embedding model stores in data\
embedding_model_saved_path	The path of the trained embedding model. When the embedding training is completed, the trained embedding model i.e., the dictionary file will be placed in the path specified.	By default, the trained embedding files will be stored in embedding\
seed	The seed for replicating the result.	By default, the seed is 1.
n_workders	The number of threads for training.	By default, we use 4 threads. Using more threads to speed up training.
size/components	The dimensionality of the word vectors.	By default, we use 100. After embedding, each input sequence will be a tensor with the shape of (1000, 100).
window	The maximum distance between the current and predicted word within a sentence.	In this code, we use a value of 5.
epoch (GloVe & FastText)	The epcoh used for training.	In this code, we use the value of 40.
min_count (Word2vec & FastText)	Ignores all words with total frequency lower than this.	In this code, we set this value to 5.
algorithm (Word2vec & FastText)	Training algorithm: 1 for skip-gram; otherwise Continouse Bag-Of-Word (CBOW).	In this code, we use CBOW.
learning_rate (GloVe)	The learning rate used for training.	In this code, we set this value to 0.001.

Model settings

Parameter	Description	Note
model	The name of the neural network models for training.	Currently, the code supports the DNN, RNNs (i.e., LSTM and GRU ), BiRNN (i.e., bidirectional LSTM and bidirectional GRU), and textCNN)
optimizer	The optimizer used.	A user can choose different optimizers based on their tasks. In this code, we use the SGD with its default settings.
loss_function	The loss function to minimize.	We use the binary cross entropy.
handle_data_imbalance	Whether to handle the data imbalance issue.	The cost-sensitive learning will be applied if it is set to True
max_sequence_length	The length for each input code sequence.	In this code, we use 1,000 as the maximal input length.
use_dropout	Whether to use dropout	In this code, we use a dropout to prevent overfitting.
dropout_rate	The dropout value.	In this code, we use the value of 0.5.
dnn_size	The number of neurons used for DNN (the first layer)	In this code, we set this value to 128.
rnn_size	The number of neurons used for RNNs (the first layer)	In this code, we set this value to 128.
birnn_size	The number of neurons used for bidirectional RNNs (the first layer)	In this code, we set this value to 64.
embedding_trainable	Whether allows the trained embedding layer to be tuned.	By default, we set this value to False.

Training settings

Parameter	Description	Note
Test_set_ratio	If not using a separate test set, set the test set ratio.	In this code, we partition the dataset into training, validation and test sets with a ratio of 6:2:2. Users can use their own test set. If users use their own test set, they should set the 'using_separate_test_set' to True and ignore this value.
using_separate_test_set	Suggest whether use a separate test set.	If this is set to True, please specify the path of test set.
test_set_path	The path contains the test data.	If a user uses a separate test set, the user should specify a path leading to the test data.
Validation_set_ratio	Suggest the percentage of the total dataset that is used as the validation set.	We use part of the training set as the validation set.
batch_size	The size of mini batch	A relatively small batch size leads to better generalization. We use the value of 16.
epochs	The number of forward and backward pass of all the training examples.	In this code, we set this value to 150.
patcience	The number of epochs with no improvement after which training will be stopped.	In this code, we set this value to 35.
save_training_history	Whether to save the training history.	The default value is True.
plot_training_history	Whether to plot the training/validation curve.	The default value is True.
validation_metric	The quantity to be monitored.	In this code, we choose to monitor validation loss.
save_best_model	If save_best_only=True, the latest best model according to the quantity monitored will not be overwritten	In this code, we set this value to True.
period_of_saving	Interval (number of epochs) between checkpoints.	In this code, we set this value to 1.
log_path	The path where the log files are stored.	By default, the log files are stored in the logs/.
model_save_path	The path where the trained models are stored.	By default, the trained models are stored in the result/models/.
model_saved_name	The name of the trained model.	By default, the trained model is called 'test_model'.