Layers

The set of layers, layer settings and layer factories included with this package are listed below:

• Convolution (1D):

  • DeepConvolution1dNetworkSettings: Configurable repeated series of 1-dimension convolution, pooling, batch norm and activation layers.

  • DeepConvolution1d: Configurable repeated series of 1-dimension convolution, pooling, batch norm and activation layers.

• Word Embedding Layer:

  • EmbeddingNetworkSettings: A utility container settings class for models that use an embedding input layer that inherit from EmbeddingNetworkModule.

  • EmbeddingNetworkModule: An module that uses an embedding as the input layer.

• Conditional Random Field:

  • EmbeddedRecurrentCRFSettings: A utility container settings class for convulsion network models.

  • EmbeddedRecurrentCRF: A recurrent neural network composed of an embedding input, an recurrent network, and a linear conditional random field output layer.

• Transformers:

  • TransformerEmbeddingLayer: A transformer (i.e. BERT) embedding layer that allows for direct access to output embeddings. Sentence and token classification is supported out of the box and selected with the TransformerEmbedding.output attribute. The TransformerResource.model_id attribute gives a list of tested models.

  • TransformerSequence: A sequence based model for token classification use HuggingFace transformers. Named entity recognition (NER) is one use case of this model.

  • See transformer.conf for examples of how to configure transformer layers.