Closed bastings closed 6 years ago
Example broken output on IWSLT Vi-En:
How How How How How How How How How How do do do do do do do do do do I I I I I I I I I I tell tell tell tell tell tell tell tell tell tell ...
@bastings Thanks, there is a bug in the model.decode when we reshape the final output. I will update with a fix soon.
It looks like beam search fails when time_major is set to False.
I've tested on IWSLT15 vi->en with beam_width 10.
With time_major true, I get >0.0 BLEU scores after 1 epoch. With time_major false, BLEU scores remain 0.0 after many epochs (while perplexity goes down.)
I'm not yet sure what causes this.
Tested on tf-1.2 branch, but probably applies to master as well.