DANLI

Code for EMNLP 2022 Paper DANLI: Deliberative Agent for Following Natural Language Instructions [paper] [arXiv]

Installation

Create a virtual environment with Python 3.8 such as using conda:

conda create --name danli python=3.8
conda activate danli

Install the dependencies:

pip install -r requirements.txt
pip install -e .

Install the fast-downward PDDL planner:

cd ..
git clone https://github.com/aibasel/downward.git fast_downward
cd fast_downward && ./build.py
cd ../DANLI

Download the TEACh dataset and the model weights

Download the raw dataset from the official teach repo into teach-dataset.

Download the pre-processed data:

sh download_data.sh

Download the model weights:

sh download_model.sh

Run DANLI

Set paths:

export DANLI_ROOT_DIR=$(pwd)
export DANLI_DATA_DIR=$DANLI_ROOT_DIR/teach-dataset
export DANLI_MODEL_DIR=$DANLI_ROOT_DIR/models
export DANLI_EVAL_DIR=$DANLI_ROOT_DIR/evals

# replace with your fast downward installation path
export FASTDOWNWARD_DIR=<YOUR_FAST_DOWNWARD_INSTALLATION_DIR>

Start an x-server (a prerequisite to launch the ai2thor environment) and set the DISPLAY variable:

sudo python3 start_x.py start 9
export DISPLAY=:9

Run the evaluation:

python run/run_neural_symbolic.py \
    --eval_name danli_eval
    --benchmark edh
    --split valid_seen
    --num_processes 2
    --num_gpus 2

Note that the above command runs DANLI for the valid_seen split on TEACh EDH benchmark by running 2 workers in parallel. The output will be stored under $DANLI_EVAL_DIR/danli_eval.

Compute the metrics

teach_eval \
    --data_dir $DANLI_DATA_DIR \
    --inference_output_dir $DANLI_EVAL_DIR/danli_eval/predictions \
    --split divided_val_seen \
    --benchmark edh\
    --metrics_file $DANLI_EVAL_DIR/danli_eval/metrics/metrics

After running the evaluation for both the valid_seen and valid_unseen splits, the metrics on divided_val_seen, divided_val_unseen, divided_test_seen and divided_test_unseen can be computed through running the above command with the corresponding split argument. See here for an explaination about the difference between the divided version of data split and the original ones.

Contact

Feel free to create an issue or send email to zhangyic@umich.edu