Extending OpenNMTs TensorFlow Lite to Include Transformer Models
Abstract
Since its release in 2017 the OpenNMT project has provided open development tools for Neural Machine Translation (NMT) including machine-learning inference with artificial neural-network models on Android platforms. Rapid advances in OpenNMT methods were achieved using TensorFlow since 2018; however, most of these advances were not deployable for use on Android platforms pending completion of the TensorFlow Lite library. The US Army Combat Capabilities Development Command Army Research Laboratorys Shareable Components project team closely tracked progress on TensorFlow Lite and succeeded in implementing a new method for converting OpenNMT models from standard TensorFlow to the Lite variant. Deployable on Android devices, these converted models provide important gains in execution speed while occupying less space. This extension adds more features to OpenNMT-tf, which allows the export of better-performing Transformer models onto Android. Beam search and <unk> replacement have also been added to this extension, which will generally help increase the performance.
Document Details
- Document Type
- Technical Report
- Publication Date
- Jul 01, 2021
- Accession Number
- AD1144269
Entities
People
- Gerardo Cervantes
- Stephen Larocca
Organizations
- United States Army Research Laboratory