THUMT Documentation

November 2, 2020 ยท View on GitHub

THUMT is an open-source toolkit for neural machine translation developed by the Tsinghua Natural Language Processing Group.

Benchmarks

We benchmark THUMT on the following datasets:

The testsets for WMT14 En-De and WMT18 Zh-En are newstest2014 and newstest2017 respectively.

DatasetModelSizeStepsGPUsBatch/GPUModeBLEU
WMT14 En-DeTransformerBase100k42*4096FP1626.85
WMT14 En-DeTransformerBase100k42*4096FP3226.91
WMT14 En-DeTransformerBase100k84096FP3226.95
WMT14 En-DeTransformerBase86k82*4096FP3227.21
WMT14 En-DeTransformerBig300k84096FP1628.71
WMT14 En-DeTransformerBig20k168*4096DistFP1628.68
WMT18 Zh-EnTransformerBig300k82*4096FP1624.07

return to index