RNN prediction is broken in bleeding-edge Theano

elliottd / GroundedTranslation

Multilingual image description

https://staff.fnwi.uva.nl/d.elliott/GroundedTranslation/

BSD 3-Clause "New" or "Revised" License

46 stars 25 forks source link

RNN prediction is broken in bleeding-edge Theano #8

Closed elliottd closed 8 years ago

elliottd commented 9 years ago

If I take a model over which I know the BLEU score (in this case, 11.11) and run it on bleeding-edge Theano code (pip install git+git://github.com/Theano/Theano.git) then something is seriously broken:

BLEU = 0.00, 40.2/0.0/0.0/0.0 (BP=0.000, ratio=0.050, hyp_len=1255, ref_len=24880)

If anyone else experiences this problem, please reply to this issue. It may be caused by an unexpected change to how recurrence() works in bleeding-edge Theano, or something even more sinister.

elliottd commented 8 years ago

Closing because it seems to work fine in Theano 0.71a1 (https://github.com/Theano/Theano/releases/tag/rel-0.7.1a1)