Re: I have a question about Nematus

[Rico Sennrich]

Hi Safae,

What is the content of en-de/input-test.en? You're saying that you're running Nematus on Windows - this isn't heavily tested and I'm not sure if there's some compatibility problem in Nematus itself, or with the way you invoke translate.py.

"python train.py -h" will show you some documentation of parameters. Here's the ones you inquire about:

validation parameters:
  --valid_freq INT      validation frequency (default: 10000)

display parameters:
  --disp_freq INT       display loss after INT updates (default: 1000)
  --sample_freq INT     display some samples after INT updates (default:
                        10000)

valid_freq is useful to keep an eye on training progress, and for early stopping. disp_freq and sample_freq don't affect training, and are purely a way to get some regular feedback about training progress (you can disable sample_freq, but I recommend using disp_freq, which will also show you training speed, and gives regular feedback about how many updates the model has made).

best wishes,
Rico

On 19/09/2019 11:27, safae berrichi wrote:

Dear Rico Sennrich,

I just started to test your toolkit "Nematus" in the windows system. But unfortunately, I have some difficulties during the execution of this toolkit.

First, after obtaining the training model using RNN on English-German, I run the script ./translate.sh on the test sentences and I received NFO: Translated 1 Sent into 3.1948089599609375e-05 sec, but the output file (reference.de) obtained is empty. Is there any explanation in case that no translation is provided through the test sentences? please guide me where i am going wrong. My translation scrpit is as follows:

THEANO_FLAGS=mode=FAST_RUN,floatX=float32,device=$device,on_unused_input=warn python nematus/translate.py -m models-en-de/model.npz  -i en-de/input-test.en -o en-de/reference.de

Secondly, I would like to have if possible, an explanation about the parameters used in the training script. So what is the role and distinction between -sampleFreq 100 --validFreq=100 --dispFreq=100.

Thank you so much.

Best Regards,

Safae.

--

=========================================

 Safae BERRICHI

PhD Candidate in 

Computer Science 

Laboratory 

Department of Mathematics and Computer Science

Faculty of Science, 

Mohammed First 

University

Oujda, Morocco.

T

el: (+212)6 50 36 08 79

=========================================

-- 
Rico Sennrich
School of Informatics
University of Edinburgh
10 Crichton Street
Edinburgh, EH8 9AB, United Kingdom
http://homepages.inf.ed.ac.uk/rsennric/