Skip to content
This repository has been archived by the owner on Jul 7, 2023. It is now read-only.

Cannot reproduce "transformer" paper's result with BPE32k #309

Closed
skyw opened this issue Sep 14, 2017 · 5 comments
Closed

Cannot reproduce "transformer" paper's result with BPE32k #309

skyw opened this issue Sep 14, 2017 · 5 comments

Comments

@skyw
Copy link

skyw commented Sep 14, 2017

I can train translate_ende_wmt32k and get reasonable bleu with transformer_big model. But if I train translate_ende_wmt_bpe32k, it doesn't translate anything meaningful.

I've tried several times, all with default settings. Any suggestions?

@mehmedes
Copy link

mehmedes commented Sep 15, 2017

Have you made sure that you preprocessed the translatable text in the same way like the training data? You'll need to tokenize and apply bpe before translating.

@skyw
Copy link
Author

skyw commented Sep 18, 2017

I believe I tried both newstest2015.en, and newstest2015.tok.clean.bpe.32000.en.
let me confirm and report back.

@skyw
Copy link
Author

skyw commented Sep 25, 2017

Here is what I got. Most paper reports bleu scores 0~100, is this really 0.31 or 31.3?
translate_ende_wmt_bpe32k/approx_bleu_score = 0.313501, metrics-translate_ende_wmt_bpe32k/rouge_2_fscore = 0.398882, metrics-translate_ende_wmt_bpe32k/rouge_L_fscore = 0.593164

@martinpopel
Copy link
Contributor

The original BLEU was defined as 0-1, but most people report it as 0-100. Note that approx_bleu is not BLEU (usually it is about 1.5 times higher).

@skyw
Copy link
Author

skyw commented Sep 25, 2017

0.313501 / 1.5 is still a reasonable number, I suppose

Sign up for free to subscribe to this conversation on GitHub. Already have an account? Sign in.
Labels
None yet
Projects
None yet
Development

No branches or pull requests

4 participants