Tensorflow NMT benchmark setup

tf_nmt_bench.sh

#Git clone NMT
git clone https://github.com/tensorflow/nmt.git

#Download Dataset
cd nmt
nmt/scripts/download_iwslt15.sh /home/ubuntu/nmt_data

#Run training for few steps to get a model, so that we can use it for inference benchmarking.
python -m nmt.nmt \
    --src=vi --tgt=en \
    --vocab_prefix=/home/ubuntu/nmt_data/vocab  \
    --train_prefix=/home/ubuntu/nmt_data/train \
    --dev_prefix=/home/ubuntu/nmt_data/tst2012  \
    --test_prefix=/home/ubuntu/nmt_data/tst2013 \
    --out_dir=/home/ubuntu/nmt_model \
    --num_train_steps=100 \
    --steps_per_stats=100 \
    --num_layers=2 \
    --num_units=128 \
    --dropout=0.2 \
    --metrics=bleu \
    --batch_size=1024 \
    --infer_batch_size=1

#Combine the test files to generate one big test file, so we can benchmark on this one merged file
cat /home/ubuntu/nmt_data/tst2012.vi /home/ubuntu/nmt_data/tst2013.vi > /home/ubuntu/my_infer_file.vi

#Test Inference. Note, infer_batch_size parameter is not working. Need to modify in model/hparams, ex:In /home/ubuntu/nmt_model/hparams
python -m nmt.nmt \
    --out_dir=/home/ubuntu/nmt_model \
    --inference_input_file=/home/ubuntu/my_infer_file.vi \
    --inference_output_file=/home/ubuntu/nmt_model/output_infer \
    --infer_batch_size=1024

#To run inference on multiple batch sizes, git clone a script
git clone https://gist.github.com/ravi9/cf62c27d852a261981441b37bf3d2e4f

#Install prereqs: sudo apt install -y moreutils jq
#Edit file paths accordingly. 
./nmt_infer_batchsize.sh