Skip to content

Latest commit

 

History

History
27 lines (21 loc) · 1.41 KB

Save_Load_Models.md

File metadata and controls

27 lines (21 loc) · 1.41 KB
Raw

Saving and loading of models

Models trained with the BiLSTM-CNN-CRF architecture and be stored on disk and later be loaded to tag new data.

Saving

The path to store the models is found in the modelSavePath variable (for example Train_POS.py):

model.modelSavePath = "models/[ModelName]_[DevScore]_[TestScore]_[Epoch].h5"

This variable specifies that the model should be stored in the models-folder. There exists some variables which are replaced during the runtime with specific values:

  • [ModelName]: replaced by the name of the dataset, for example, unidep_pos for Train_POS.py
  • [DevScore]: Score (accuracy or F1-score) on the development set
  • [TestScore]: Score (accuracy or F1-score) on the test set
  • [Epoch]: Trainings epoch

A new model is only stored if the performance on the development set increases.

Loading

Models can be loaded by calling the static method BiLSTM.loadModel(). See RunModel.py for an example:

lstmModel = BiLSTM.loadModel(modelPath)
#...
tags = lstmModel.tagSentences(dataMatrix)

Input data (in the right format) can be tagged using the tagSentences() method. Note that the passed dataMatrix must have all input values that were used during the training. Per-default, tokens, casing and maybe characters are used as input features. You might need to extend your input data by this features. Only passing tokens to the model is usually not sufficient.