Kenneth Heafield
Kenneth Heafield
multi-bleu.perl has been deprecated for years now because it encourages people to use non-standard tokenization. This repository contains another non-standard BLEU implementation that a user might not notice they are...
Some text is left untranslated "Welches Format hat Ihre Sendung?"~~ 0. Install extension 1.1.1buildid20220506.201912 with en UI laguage 1. Visit https://www.deutschepost.de/de/p/portoberater.html#/ 2. Click extensions translation button 3. Click letters 4....
https://github.com/huggingface/nlp/blob/7d1526dfeeb29248d832f1073192dbf03ad642da/metrics/bleu/bleu.py#L76 assumes the inputs are tokenized by the user. This is bad practice because the user's tokenizer is usually not the same as the one used by `mteval-v13a.pl`, the closest...
This script is harmful because it propagates a non-standard way to compute BLEU that is not reflective of the WMT 2014 task. Entirely too many papers are submitted with BLEU...
https://github.com/oneapi-src/oneDNN/ aka MKLDNN aka DNNL now has better performance for MT-size matrices: https://github.com/apache/incubator-mxnet/issues/17980 . And it's open source. The same teams write the GEMM for MKL and oneDNN. Would be...
Unfortunately Marian redefines Release to include `-g`. cmake has Release and RelWithDebInfo: Release does not have debugging info. RelWithDebInfo has debugging info. If I wanted debug symbols, I can compile...
[2021-05-18 14:57:19] [SentencePiece] Creating temporary file /tmp/marian.a4Gi0k^@ Where ^@ is just my editor's representation of a null byte. Steps to reproduce: http://data.statmt.org/heafield/train-isl-eng/
The existing implementation of Select for CPU is very slow. Reimplemented the `Select` function for CPU using `std::copy` for the case when stuff to copy after the axis is contiguous,...
FP16 support
Should add pervasive FP16 support, not just calling tensor cores. We've known about this for a while, but Facebook produced a number: 2.9x from FP16 https://arxiv.org/pdf/1806.00187.pdf
Both @alvations and Intel have noticed SentencePiece is taking ~10% of inference time and wonder if this can be optimized. First thing to try would be updating the submodule from...