README.md

Dialog Generation Model

Installation

export INSTALL_DIR=$PWD

# install apex
cd $INSTALL_DIR
git clone https://github.com/NVIDIA/apex.git
cd apex
python setup.py install --cuda_ext --cpp_ext

# install oscar
cd coco_caption
./get_stanford_models.sh
cd ..
python setup.py build develop

# install requirements
pip install -r requirements.txt

unset INSTALL_DIR

Training

Script to train dialog generation model:

python oscar/run_dialog_generate.py \
    --model_name_or_path unilm_checkpoints \
    --do_train \
    --do_lower_case \
    --evaluate_during_training \
    --add_od_labels \
    --learning_rate 0.00003 \
    --per_gpu_train_batch_size 64 \
    --num_train_epochs 30 \
    --save_steps 5000 \
    --output_dir output/

Inference

Script to inference on Reddit with beam search:

python oscar/run_dialog_generate.py \
    --do_test \
    --do_eval \
    --test_yaml test.yaml \
    --per_gpu_eval_batch_size 64 \
    --num_beams 5 \
    --max_gen_length 20 \
    --eval_model_dir your_model_for_evaluation 

Script to inference on Reddit with top_k/top_p sampling:

python oscar/run_dialog_generate.py \
    --do_test \
    --do_sample \
    --do_eval \
    --test_yaml test.yaml \
    --per_gpu_eval_batch_size 64 \
    --top_k k_val \
    --top_p p_val \
    --max_gen_length 20 \
    --eval_model_dir your_model_for_evaluation 

Reference

The code is implemented as a fork from OSCAR, more details please refer to their repo, many thanks for their work.