run_reorder.sh

#execute this script to train and validate model performance on dataset, comment out the train and validation section and uncomment the evaluate section 
#to test final model performance. Our codes depend on the tensor2tensor library. Please visit https://github.com/tensorflow/tensor2tensor for more detail. 

problem=ptb2016_shuffle #see registered problem at usr/t2t_usr/ptb2016.py for more detail
model=nalm #see registered model at usr/t2t_usr/nalm.py for more detail
hparams=transformer_base_multistep8 #adjusted based on the number of available gpus, current setting is for dual-gpu machine, 
                                    #see tensor2tensor/models/transformer.py for more detail
beam_size=1 #for non-autoregressive model, please set this as 1, for autoregressive model, adjust according to need.

train_dir=./t2t_train/reorder/$problem/$model'_'$hparams
trans_dir_base=$model'_'$hparams/$problem/$beam_size
translation_dir=test/$trans_dir_base

source_dir=./tmp/t2t_datagen/ptb2016/ptb2016.valid.txt #ground truth validation data
source_ref=./tmp/t2t_datagen/ptb2016/ptb2016.valid.fullyshuffled.txt #validation input
test_dir=./tmp/t2t_datagen/ptb2016/ptb2016.test.txt #ground truth test data
test_ref=./tmp/t2t_datagen/ptb2016/ptb2016.test.fullyshuffled.txt #test input

steps=100000
min_steps=9999
best_score=93644 #modify this based on validation results before running the evaluation.

##Training and validation
t2t-trainer --t2t_usr_dir=./usr/t2t_usr --data_dir=./t2t_data --problem=$problem --model=$model --hparams_set=$hparams --output_dir=$train_dir --train_steps=$steps --worker_gpu=2 --worker_gpu_memory_fraction=0.9 --save_checkpoints_secs=3600 --tmp_dir=./tmp/t2t_datagen --schedule='train' --generate_data

t2t-translate-all --t2t_usr_dir=./usr/t2t_usr --model=$model --hparams_set=$hparams --tmp_dir=./tmpt2t_datagen --source=$source_ref --translations_dir=$translation_dir --model_dir=$train_dir --problem=$problem --data_dir=./t2t_data --min_steps=$min_steps --beam_size=$beam_size

t2t-bleu --translations_dir=$translation_dir --reference=$source_dir --event_dir=event/$trans_dir_base

##Evaluate
#t2t-decoder --t2t_usr_dir=./usr/t2t_usr --output_dir=$train_dir --data_dir=./t2t_data --problem=$problem --model=$model --hparams_set=$hparams --checkpoint_path=$train_dir/model.ckpt-$best_score --decode_to_file=$translation_dir/best --eval_use_test_set=True --keep_timestamp --decode_hparams="batch_size=32, beam_size=$beam_size" --decode_from_file=$test_ref

#python compute_meteor.py $translation_dir/best $test_dir
#t2t-bleu --translation=$translation_dir/best --reference=$test_dir