search

lisa-groundhog / GroundHog

Library for implementing RNNs with Theano
BSD 3-Clause "New" or "Revised" License
598 stars 229 forks source link

sample.py gives unreasonable translation #38

Open sjtufs opened 8 years ago

sjtufs commented 8 years ago

Hi, I had some problem when translating with trained model. I trained on the dataset given here which is used in NMT with alignment paper on one GPU for approx.200 hours.

At training step, the results given by sampling are very promising. But when I tried sampling with beam size of 12 / 60, the translations are basically all 'UNK', even on the training set. I tried using a smaller dataset and still got this problem.

My testing script is like: python sample.py --source=test.fr --beam-search --beam-size 12 --trans test.trans --state search_state.pkl, search_model.npz 2>>log.txt

Has anyone seen this problem before?

I know you guys have shifted to Blocks but maybe you can tell me how to correctly test a trained model. It would be a lot of help. Thanks!

xpertasks commented 8 years ago

Hi @sjtufs Did you resolve the issue? I'm facing the same problem.