Skip to content

Latest commit

 

History

History

ChineseNMT

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
TED_data
TED_data
 
 
AttentionNMT.py
AttentionNMT.py
 
 
DataPreprocessing.py
DataPreprocessing.py
 
 
README.md
README.md
 
 
Seq2SeqNMT.py
Seq2SeqNMT.py
 
 
test.py
test.py
 
 
train.py
train.py
 
 

README.md

Chinese Neural Machine Translation

Dataset

TED Talk (for translation corpus) - readme

Usage

Data proprocessing

# Data Preprocessing
mkdir output_vocab train_data
python3 DataPreprocessing.py

Training

python3 train.py

Test

python3 test.py

Result

English: This is a test . <eos>
Chinese (seq2seq): <sos>這是一個測試。<eos>
Chinese (attention): <sos>這是一個測試。<eos>

English: Please give me a hundred on my final score . For the sake of how hard I paid on this course . <eos>
Chinese (seq2seq): <sos>請給我一分鐘。我的下一個問題是，我花了很長的時間才這樣做。<eos>
Chinese (attention): <sos>我最後一個人在這兒為我付出的錢。我的工作。<eos>

English: I just can't understand why I am so handsom . <eos>
Chinese (seq2seq): <sos>我只是想要解決我的問題。<eos>
Chinese (attention): <sos>我明白了為什麼我只是在理解這個理論，我為什麼只能在這個理論上解釋我的理解。<eos>

Training time and loss

Seq2seq

########################################################################
#      Date:           Tue Dec 25 17:34:15 PST 2018
#    Job ID:           8897.c009
#      User:           u22711
# Resources:           neednodes=1:ppn=2,nodes=1:ppn=2,walltime=24:00:00
########################################################################

In iteration: 1
After 0 steps, per token cost is 8.295

...

After 9000 steps, per token cost is 2.397
After 9010 steps, per token cost is 2.525

########################################################################
# End of output for job 8897.c009
# Date: Wed Dec 26 00:23:54 PST 2018
########################################################################

Attention

########################################################################
#      Date:           Tue Dec 25 17:36:09 PST 2018
#    Job ID:           8900.c009
#      User:           u22711
# Resources:           neednodes=1:ppn=2,nodes=1:ppn=2,walltime=24:00:00
########################################################################

In iteration: 1
After 0 steps, per token cost is 8.292

...

After 9000 steps, per token cost is 2.582
After 9010 steps, per token cost is 2.571

########################################################################
# End of output for job 8900.c009
# Date: Wed Dec 26 06:01:17 PST 2018
########################################################################

Training attention model is much slower than seq2seq model (about two times)

TODO

training

Current version can only translate one sentence per run. Make it a better way to reuse the model.

ValueError: Variable XXX already exists, disallowed. Did you mean to set reuse=True or reuse=tf.AUTO_REUSE in VarScope?

testing

Multiple graph will conflict now.

TensorFlow Graphs and Sessions - Programming with multiple graphs

Links