-
Notifications
You must be signed in to change notification settings - Fork 2k
Cannot replicate English to Vietnamese - on python 2 and 3 #398
Comments
Hi there, I checked your python 3 logfile and hparams and the parameters that were used for training are quite a bit different from the original https://github.com/tensorflow/nmt/blob/master/nmt/standard_hparams/iwslt15.json Are you specifying the configuration file correctly? It looks like your run did not use attention and only used 32 units. |
@bastings Thanks. I have made the changes and running it now. Will update you in sometime. -Mohammed Ayub |
@bastings Thanks for you help. Mohammed Ayub |
Hi,
Using the below configuration:
Python 2.7.15
Tensorflow: 1.10.0
Training Query:
python -m nmt.nmt --src=en --tgt=vi --hparams_path=/home/ubuntu/nmt/standard_hparams/iwslt15.json --vocab_prefix=/home/ubuntu/en_vi_dataset/vocab --train_prefix=/home/ubuntu/en_vi_dataset/train --dev_prefix=/home/ubuntu/en_vi_dataset/tst2012 --test_prefix=/home/ubuntu/en_vi_dataset/tst2013 --out_dir=/home/ubuntu/en_vi_dataset/models/attention_model27
Best Bleu score comes out to be: 1.5698. Not sure whats going wrong, hoping its not encoding as I used that files directly from the download.
Below file contains the Params and the Log for Python 2 run
run1_py_2.7.15.zip
Python 3 Run:
Python 3.6.5
Tensorflow: 1.10.0
Training Query:
python -m nmt.nmt --src=en --tgt=vi --hparams_path=/home/ubuntu/nmt/standard_hparams/iwslt15.json --vocab_prefix=/home/ubuntu/en_vi_dataset/vocab --train_prefix=/home/ubuntu/en_vi_dataset/train --dev_prefix=/home/ubuntu/en_vi_dataset/tst2012 --test_prefix=/home/ubuntu/en_vi_dataset/tst2013 --out_dir=/home/ubuntu/en_vi_dataset/models/attention_model
Best Bleu score comes out to be: 2.18338
Below file contains the Params and the Log for Python 3 run:
run2_py_3.6.5.zip
Appreciate any help.
Lopping some folks who have answered similar questions before.
@bastings @oahziur
Thanks !
Mohammed Ayub
The text was updated successfully, but these errors were encountered: