• Add Pipeline data structure to sticker-utils that allows loading of several models, that are then applied in sequence.
• Update sticker-tag and sticker-server to support annotation pipelines. These commands now accept multiple model configuration files as arguments and will apply these models in sequence when tagging.
• Add the --input and --output options to sticker-tag. Since we cannot disambiguate between model configuration arguments and input/output corpora, the input and output files are now specified through these two flags. If these flags are not used, stdin/stdout will be used as before.
• The address/port to bind to is now specified using the addr option of sticker-server. By default, sticker-server will bind to localhost:4000.

Update the CoNLL-X dependency to version 0.12. Unfortunately, this requires bumping the version, since this is an API-breaking change.

This release contains one large change: the loading of quantized models is speeded up by computing the unknown word embedding as an avarage of the subquantizers, rather than an average of all in-vocab word embeddings.

This release adds support for word representations using byte RNNs. These word representations can be enabled by simply adding

[input]
subwords = true

For consistency, the embeddings.word and embeddings.tag options have also been moved to the input section. Hopefully, this is one of the last changes to the configuration file format.

Various options have been added to the graph writing scripts to control the hyper parameters of byte RNNs.

Features

• Add support for writing TensorBoard summaries.
• Split out pretraining support into a separate sticker-pretrain utility.
• Add the --residual option for RNN residual connections.
• Add the --maxlen option to sticker-{pretrain,train} for discarding long sentences.
• Write the model hyperparameters to the graph and stdout. Provide the sticker-graph-metadata utility to print the hyperparameters.
• Add the gpu_allow_growth option to allocate video RAM on a by-need basis.

Maintenance/bugfixes

• Restore the hidden_size option for dilated convolutions.
• Remove the DefaultConfig class from sticker-graph.
• Support Tensorflow 1.14.
• Fix CRF support.
• Factor out tagger construction boilerplate to TaggerWrapper.

Thanks to @twuebi for the many contributions to this release!