Marian (Neural Machine) Translation Service
Перейти к файлу
abhi-agg 733ab64cd2
Merge pull request #4 from mozilla/development-abhi
Hook mozilla repo for 'ssplit-cpp' submodule
2020-06-29 15:04:08 +02:00
3rd_party Hook mozilla repo for 'ssplit-cpp' submodule 2020-06-29 13:37:19 +02:00
cmake Get things to configure and compile. 2020-05-06 13:18:55 +01:00
docs Fix version in docs/source/config.py. 2020-05-13 12:02:49 +01:00
src Update qe.cpp: const diligence 2020-06-23 16:21:54 +01:00
.compute Compile the repo on ML cluster successfully 2020-06-23 11:54:09 +02:00
.gitignore Transferred Marian Translation Service code from Marian code base. 2020-05-06 12:00:56 +01:00
.gitmodules Hook mozilla repo for 'ssplit-cpp' submodule 2020-06-29 13:37:19 +02:00
CMakeLists.txt Get things to compile with gcc-7 2020-05-13 15:41:15 +01:00
LICENSE Create LICENSE 2020-05-12 15:12:13 +01:00
README.md Update README.md 2020-05-07 20:35:45 +01:00
VERSION Transferred Marian Translation Service code from Marian code base. 2020-05-06 12:00:56 +01:00

README.md

Marian Translation Service

The code in this repository implements a Marian-based translation service. Currently, it offers an REST http server. The lack of https is currently on purpose, as in the vast majority of the use cases that I envision, the service will run locally and not through a registered domain address, so only self-signed certificates would work. If you need https, go through a proxy and put the Marian REST server behind a firewall.

Currently, the server can serve only a single configuration (model or ensemble of models).

The REST API is described here.

Compilation

Local

Notice that you need all the dependencies and libraries for a full compilation of Marian.

git clone https://github.com/ugermann/mts /path/to/your/local/clone/of/mts
mkdir /path/to/where/you/want/to/build
cd /path/to/where/you/want/to/build/mts
cmake /path/to/your/local/clone/of/mts
make -j

Dockerized Version

(coming soon, based on http://github.com/ugermann/marian-docker ...)

Configuring a Model to serve

These instructions assume that you have the following files from a marian training process:

  • at least one model (more for ensemble decoding; do not try ensembed decoding unless your REST server has access to a GPU)
  • the respective vocabulary file(s)
  • a shortlist file for faster hypothesis generation (optional)

Preparation

  • Convert the model file(s) to binary format with

    marian-conv -f model.npz -t model.bin
    

    This is optional, but makes the model load faster.

  • Create a decoder.yml file, e.g. like this one:

    relative-paths: true
    models: [ model.bin ]
    vocabs: [ joint-vocab.spm, joint-vocab.spm ]
    alignment: true
    beam-size: 4
    normalize: 1
    word-penalty: 0
    mini-batch: 128
    maxi-batch: 100
    maxi-batch-sort: src
    
    # The following are specific to the marian REST server
    # source-language and target-language are used for the Demo
    # interface; the ssplit-prefix-file is from the Moses sentence splitter
    # and comes with the marian REST server image. Pick the right one
    # for your source language. SSPLIT_ROOT_DIR is set to the appropriate
    # value in the `mariannmt/marian-rest-server` image.
    source-language: German
    target-language: English
    ssplit-prefix-file: ${SSPLIT_ROOT_DIR}/nonbreaking_prefixes/nonbreaking_prefix.de
    
  • Copy the appropriate nonbreaking_prefix.* file for sentence splitting into an appropriate location, or

    export SSPLIT_ROOT_DIR=/path/to/your/local/clone/of/mts/3rd_party/ssplit-cpp
    

Running the Server

/path/to/your/build/directory/rest-server -c /path/to/decoder.yml -p <port of your choice>

Known bugs

  • Rest-server currently inherits the version info from the marian submodule, which is obviously incorrect.