nmt-attention-tf2

Effective Approaches to Attention-based Neural Machine Translation implemented as Tensorflow 2.0

Requirements

Tensorflow 2.0

Data

WMT'14 English-German data: https://nlp.stanford.edu/projects/nmt/

Download the datasets using the following script:

./download.sh

Usage

usage: main.py [-h] [--mode MODE] [--config-path DIR] [--init-checkpoint FILE]
               [--batch-size INT] [--epoch INT] [--embedding-dim INT]
               [--max-len INT] [--units INT] [--dev-split REAL]
               [--optimizer STRING] [--learning-rate REAL] [--dropout REAL]
               [--method STRING] 

train model from data

optional arguments:
  -h, --help            show this help message and exit
  --mode MODE           train or test
  --config-path DIR     config json path
  --init-checkpoint FILE
                        checkpoint file
  --batch-size INT      batch size <default: 32>
  --epoch INT           epoch number <default: 10>
  --embedding-dim INT   embedding dimension <default: 256>
  --max-len INT         max length of a sentence <default: 90>
  --units INT           units <default: 512>
  --dev-split REAL      <default: 0.1>
  --optimizer STRING    optimizer <default: adam>
  --learning-rate REAL  learning rate <default: 0.001>
  --dropout REAL        dropout probability <default: 0>
  --method STRING       content-based function <default: concat>

Train command example

python main.py --max-len 50 --embedding-dim 100 --batch-size 60 --method concat

Test command example

python main.py --mode test --config-path training_checkpoints/{TRAINING_CHECKPOINT}/config.json

Demo

I think this demo is poor performance because I don't have a large resource. So, The paper proposed embedding dimension sets 1000. But this demo's embedding dimension is 50. And this is trained only for 4 epochs.

If you don't have training_checkpoints directory, make training_checkpoints directory and proceed with the next step.

mkdir training_checkpoints
cd training_checkpoints

You can download here. And you put DEMO directory in training_checkpoints directory.

python main.py --mode test --config-path training_checkpoints/DEMO/config.json

Example

Input Sentence or If you want to quit, type Enter Key : What is your name?
Early stopping
<s> was ist ihr name ? </s> 
<s> what is your name ? </s>

Results

	Train Set BLEU	Test Set BLEU
Model	--	--

Reference

Effective Approaches to Attention-based Neural Machine Translation