ESConv

Alignment of ESConv (Blender-Vanilla & Blender-Joint)

Our implementation is based on ESConv.

Data Download

Download ESConv.json and strategy.json and put them in the folder DATA.

Dara Preprocessing

Enter DATA and run python process.py.

To preprocess the training data (for Blender-Vanilla and Blender-Joint, respectively), run:

python prepare.py --config_name vanilla --inputter_name vanilla --train_input_file DATA/train.txt --max_input_length 160 --max_decoder_input_length 40

python prepare.py  --config_name strat --inputter_name strat --train_input_file DATA/train.txt --max_input_length 160  --max_decoder_input_length 40

Base Model Training

Run:

. RUN/train_vanilla.sh {gpu_id}

. RUN/train_strat.sh {gpu_id}

Aligned Model Training

Change the value of --preference_model_dir in the RUN/align_vanilla.sh and RUN/align_strat files. The--preference_model_dir is the path of the preference model checkpoint folder. Run:

. RUN/align_vanilla.sh {gpu_id} {model_path}

. RUN/align_strat.sh {gpu_id} {model_path}

The {model_path} is the path of the base model checkpoint folder.

Model Inference

Run:

. RUN/infer_model.sh {gpu_id} {model_name} {model_path}

{model_name} can be either vanilla or strat.