Torch Models

学習済みモデル

Available models trained using OpenNMT.

Tutorials and Recipes

We provide tutorials for training these and other models in our forum.

Additonally, we plan on posting the scripts for each of our training recipes.

Benchmarks

This page benchmarks training results of open-source NMT systems with generated models of OpenNMT and other systems. If you have a competitive model - please contact us at info@opennmt.net with necessary information to reproduce, and we will register your system in this “hall of fame”.

English->German

Who/When Corpus Prep Training Tool Training Parameters Server Details Training Time/Memory Scores Model
2016/20/12
Baseline
WMT15 - Translation Task
+ Raw Europarl v7
+ Common Crawl
+ News Commentary v10
OpenNMT aggressive tokenization
OpenNMT preprocess.lua default option (50k vocab, 50 max sent, shuffle)
OpenNMT 111f16a default options:
2 layers, RNN 500, WE 500, input feed
13 epochs
Intel(R) Core(TM) i7-6800K CPU @ 3.40GHz, 256Gb Mem, trained on 1 GPU TITAN X (Pascal) 355 min/epoch, 2.5Gb GPU usage valid newstest2013:
PPL: 7.19
newstest2014 (cleaned):
NIST=5.5376
BLEU=0.1702
692M here

WMT15 training and validation data also available here.

German->English

Who/When Corpus Prep Training Tool Training Parameters Server Details Training Time/Memory Scores Model
2016/20/12
Baseline
WMT15 - Translation Task
+ Raw Europarl v7
+ Common Crawl
+ News Commentary v10
OpenNMT aggressive tokenization
OpenNMT preprocess.lua default option (50k vocab, 50 max sent, shuffle)
OpenNMT 111f16a default options:
2 layers, RNN 500, WE 500, input feed
13 epochs
Intel(R) Core(TM) i7-6800K CPU @ 3.40GHz, 256Gb Mem, trained on 1 GPU TITAN X (Pascal) 346 min/epoch, 2.5Gb GPU usage valid newstest2013:
PPL: 8.98
newstest2014 (cleaned):
NIST=6.4531
BLEU=0.2067
692M here

WMT15 training and validation data also available here.

Multi-way - FR,ES,PT,IT,RO<>FR,ES,PT,IT,RO

Following Toward Multilingual Neural Machine Translation with Universal Encoder and Decoder (Thanh-Le Ha et al, 2016) and Google’s Multilingual Neural Machine Translation System: Enabling Zero-Shot Translation (Johnson et al, 2016) we trained a multi-way engine between French, Spanish, Portuguese, Italian and Romanian.

The corpus used is completely parallel - and there are only 200,000 sentences per language - the tokenized corpus with test, valid is here.

Who/When Corpus Prep Training Tool Training Parameters Server Details Training Time/Memory Scores Model
2017/07/01
Baseline
OpenNMT aggressive tokenization with BPE 32k OpenNMT 481b784 4 layers, RNN 1000, WE 600, input feed, brnn
13 epochs
Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz, 96Gb Mem, trained on 1 GPU 1080 GeForce (Pascal) 887 min/epoch, 6Gb GPU usage (described in forum) 2.9G (GPU) here

English Summarization

Who/When Corpus Prep Training Tool Training Parameters Server Details Training Time/Memory Scores Model
2016/21/12
Baseline
Gigaword Standard OpenNMT 111f16a default options:
2 layers, RNN 500, WE 500, input feed
11 epochs
Trained on 1 GPU TITAN X   Gigaword F-Score R1: 33.13 R2: 16.09 RL: 31.00 572M here or cpu release