/students

Efficient teacher-student models and scripts to make them

Primary LanguageHandlebarsOtherNOASSERTION

NMT models for Bergamot

Teacher and student NMT models are described in subdirectories. We have prepared models for the following language pairs:

  • Estonian-English
  • English-Estonian
  • Spanish-English
  • English-Spanish
  • English-German

Instructions

  1. Download and compile the latest marian-dev in the upper directory:

     bash install.sh

    The compilation requires MKL and gperftools, see the Marian documentation to find which packages are needed on Ubuntu.

  2. Go to a subdirectory, download data and model files, e.g.:

     cd eten
 bash download-data.sh
 bash download-models.sh

    See README file for more details.

  3. Go to a model directory and run the evaluation script, e.g.:

     cd eten.student.tiny11
 bash eval.sh -d 0       # -d 0 means use first GPU

    or benchmarking scripts:

     bash speed.gpu.sh -d 0
 bash speed.cpu.sh
 bash speed.cpu.packed8.sh   # requires CPU with avx2

    See individual scripts for details.