RUN.txt

/// Training Usage

--> Usage prescribed by original NSLT(Camgoz, CVPR2018 paper)
python -m nslt \
        --src=sign \
        --tgt=de \
        --train_prefix=modeldata/phoenix2014T.train \
        --dev_prefix=modeldata/phoenix2014T.dev \
        --test_prefix=modeldata/phoenix2014T.test \
        --out_dir=./nslt_model \
        --vocab_prefix=modeldata/phoenix2014T.vocab \
        --source_reverse=True \
        --num_units=1000 \
        --num_layers=4 \
        --num_train_steps=150000 \
        --steps_per_stats=100 \
        --residual=True \
        --attention=luong \
        --base_gpu=0 \
        --metrics=bleu \
        --unit_type=gru

--> Experiments and Tests performed using following low resource usage command
python -m nslt \
        --src=sign \
        --tgt=de \
        --train_prefix=modeldata/phoenix2014T.train \
        --dev_prefix=modeldata/phoenix2014T.dev \
        --test_prefix=modeldata/phoenix2014T.test \
        --out_dir=./nslt_model \
        --vocab_prefix=modeldata/phoenix2014T.vocab \
        --source_reverse=True \
        --num_units=32 \
        --num_layers=2 \
        --learning_rate=0.001 \
        --src_max_len=100 \
        --tgt_max_len=50 \
        --src_max_len_infer=150 \
        --num_train_steps=150 \
        --steps_per_stats=2 \
        --residual=True \
        --attention=luong \
        --base_gpu=0 \
        --metrics=bleu \
        --unit_type=gru

Inference Usage

python -m nslt \
        --out_dir=./nslt_model \
        --inference_input_file=modeldata/phoenix2014T.test.sign \
        --inference_output_file=./phoenix2014T.test.inf_out.de \
        --inference_ref_file=modeldata/phoenix2014T.test.de \
        --base_gpu=0

View Logging in Tensorboard

tensorboard --port 22222 --logdir ./nslt_model