Skip to content

Latest commit

 

History

History
57 lines (44 loc) · 3.91 KB

README.md

File metadata and controls

57 lines (44 loc) · 3.91 KB
Raw

piglet

PIGLeT: Language Grounding Through Neuro-Symbolic Interaction in a 3D World [ACL 2021] This repo contains code and data for PIGLeT. If you like this paper, please cite us:

@inproceedings{zellers2021piglet,
    title={PIGLeT: Language Grounding Through Neuro-Symbolic Interaction in a 3D World},
    author={Zellers, Rowan and Holtzman, Ari and Peters, Matthew and Mottaghi, Roozbeh and Kembhavi, Aniruddha and Farhadi, Ali and Choi, Yejin},
    booktitle ={Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics},
    year={2021}
}

See more at https://rowanzellers.com/piglet

What this repo contains

Physical dynamics model

  • You can get data yourself by sampling trajectories in sampler/ and then converting them to tfrecord (which is the format I used) in tfrecord/. I also have the exact tfrecords I used at gs://piglet-data/physical-interaction-tfrecords/ -- they're big files so I turned on 'requester pays' for them.

    • These TFrecords contain the world state that we trained on for this paper, and additionally, image frames 😄.
    • You can download the tfrecords using something like gsutil -u $myusername gs://piglet/physical-interaction-tfrecords/*.tfrecord .
    • Confusingly, the training files are numbered like 0XXXof0256, but there is no fold 0018. I only realized this after making these tfrecords 😅 so in other words, this was used for my internal experiments too. Basically you hopefully shouldn't have to worry about this at all, just don't panic if you don't see a file like train-0018of0256.tfrecord!

  • You can pretrain the model and evaluate it in model/interact/train.py and model/interact/intrinsic_eval.py

  • Alteratively feel free to use my checkpoint: gs://piglet/checkpoints/physical_dynamics_model/model.ckpt-5420

Language model

  • You can process data (also in tfrecord format) using data/zeroshot_lm_setup/prepare_zslm_tfrecord.py, or download at gs://piglet-data/text-data/. I have both 'zero-shot' tfrecord data, basically a version of BookCorpus and Wikipedia where certain concepts are filtered out, as well as non-zero shot (regularly processed). This was used to evaluate generalization to new concepts.
  • Train the model using model/lm/train.py
  • Alternatively, feel free to just use my checkpoint: gs://piglet/checkpoints/language_model/model.ckpt-20000

Tying it all together

  • Everything you need for this is in model/predict_statechange/ building on both the physical dynamics model and language model pretrained.
  • I have annotations in data/annotations.jsonl for training and evaluating both tasks -- PIGPeN-NLU and PIGPeN-NLG.
  • Alternatively you can download my checkpoints at gs://piglet/checkpoints/pigpen-nlu-model/ for NLU (predicting state change given english text) or gs://piglet/checkpoints/pigpen-nlg-model/ for NLG.

That's it!

Getting the environment set up

I used TPUs for this project so those are the only things I support right now, sorry!

I used tensorflow 1.15.5 and TPUs for this project. My recommendation is to use ctpu to start up a VM with access to a v3-8 TPU. Then, use the following command to install dependencies:

curl -o ~/miniconda.sh -O  https://repo.continuum.io/miniconda/Miniconda3-4.5.4-Linux-x86_64.sh  && \
     chmod +x ~/miniconda.sh && \
     ~/miniconda.sh -b -p ~/conda && \
     rm ~/miniconda.sh && \
     ~/conda/bin/conda install -y python=3.7 tqdm numpy pyyaml scipy ipython mkl mkl-include cython typing h5py pandas && ~/conda/bin/conda clean -ya
     
echo 'export PATH=~/conda/bin:$PATH' >>~/.bashrc
source ~/.bashrc
pip install "tensorflow==1.15.5"
pip install --upgrade google-api-python-client oauth2client
pip install -r requirements.txt