Pytorch Backend/Wrapper

NOTE: in order to use this, the Python environment needs to get set up first

The PytorchWrapper is currently one of two wrappers for using “Deep Learning” / Neural Networks with the Learning Framework. See Using Neural Networks for an overview.

The PytorchWrapper provides two algorithms:

• PytorchWrapper_CL_DR: a classification algorithm that processes individual instances. This creates a JSON training data file where each line represents the feature vector for an instance and the corresponding class label
• PytorchWrapper_SEQ_DR: a sequence tagging algorithm that processes whole sequences of instances (where each instances has to get assigned a label). This creates a JSON training data file where each line represents a sequence of feature vectors for each instance in the sequence and the corresponding sequence of labels for each instance in the sequence.

In both cases the default action taken by the wrapper for training is to try to dynamically create a neural network that can process the features as specified in the feature specification file to predict the class labels. Heuristics are used for how the network architecture is created and how hyperparameters like number of embedding dimensions, number of hidden units etc are chosen.

The generated architecture and hyperparameters are shown to the user and logged. This can be used to more easily adapt or implement the Pytorch neural network module specifically to the problem at hand. The pytorch wrapper library also comes with a number of pre-defined special-purpose modules which can be used as starting points for specific tasks (e.g. NER or POS tagging).

Here is an overview of the PytorchWrapper documentation:

• Training - all the details about how to train a model, specify parameters, change the neural network architecture and hyperparameters, use and adapt a predefined special-purpose module from the library etc.
• Application - how to use a trained model and apply it to new documents
• Modules - documentation of predefined network architectures which can be used in place of the automatically generated generic architectures