Pythia is a modular framework for vision and language multimodal research. Built on top of PyTorch, it features:

• Model Zoo: Reference implementations for state-of-the-art vision and language model including LoRRA (SoTA on VQA and TextVQA), Pythia model (VQA 2018 challenge winner) , BAN and BUTD.
• Multi-Tasking: Support for multi-tasking which allows training on multiple dataset together.
• Datasets: Includes support for various datasets built-in including VQA, VizWiz, TextVQA, VisualDialog and COCO Captioning.
• Modules: Provides implementations for many commonly used layers in vision and language domain
• Distributed: Support for distributed training based on DataParallel as well as DistributedDataParallel.
• Unopinionated: Unopinionated about the dataset and model implementations built on top of it.
• Customization: Custom losses, metrics, scheduling, optimizers, tensorboard; suits all your custom needs.

You can use Pythia to bootstrap for your next vision and language multimodal research project.

Pythia can also act as starter codebase for challenges around vision and language datasets (TextVQA challenge, VQA challenge)

If you use Pythia in your work, please cite:

@inproceedings{singh2019TowardsVM,
  title={Towards VQA Models That Can Read},
  author={Singh, Amanpreet and Natarajan, Vivek and Shah, Meet and Jiang, Yu and Chen, Xinlei and Batra, Dhruv and Parikh, Devi and Rohrbach, Marcus},
  booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
  year={2019}
}

and

@inproceedings{singh2018pythia,
  title={Pythia-a platform for vision \& language research},
  author={Singh, Amanpreet and Natarajan, Vivek and Jiang, Yu and Chen, Xinlei and Shah, Meet and Rohrbach, Marcus and Batra, Dhruv and Parikh, Devi},
  booktitle={SysML Workshop, NeurIPS},
  volume={2018},
  year={2018}
}

Learn more and read tutorials about Pythia here.

1. Pythia VQA.
2. BUTD Captioning.

First install the repo using

git clone https://github.com/facebookresearch/pythia ~/pythia

# You can also create your own conda environment and then enter this step
cd ~/pythia
python setup.py develop

Now, Pythia should be ready to use. Follow steps in specific sections to start training your own models using Pythia.

Default configuration assume that all of the data is present in the data folder inside pythia folder.

Depending on which dataset you are planning to use download the feature and imdb (image database) data for that particular dataset using the links in the table (right click -> copy link address). Feature data has been extracted out from Detectron and are used in the reference models. Example below shows the sample commands to be run, once you have the feature (feature_link) and imdb (imdb_link) data links.

cd ~/pythia
mkdir -p data && cd data
wget https://dl.fbaipublicfiles.com/pythia/data/vocab.tar.gz

# The following command should result in a 'vocabs' folder in your data dir
tar xf vocab.tar.gz

# Download detectron weights
wget https://dl.fbaipublicfiles.com/pythia/data/detectron_weights.tar.gz
tar xf detectron_weights.tar.gz

# Now download the features required, feature link is taken from the table below
# These two commands below can take time
wget feature_link

# [features].tar.gz is the file you just downloaded, replace that with your file's name
tar xf [features].tar.gz

# Make imdb folder and download required imdb
mkdir -p imdb && cd imdb
wget imdb_link

# [imdb].tar.gz is the file you just downloaded, replace that with your file's name
tar xf [imdb].tar.gz

After downloading the features, verify the download by checking the md5sum using

echo "<checksum>  <dataset_name>.tar.gz" | md5sum -c -

Once we have the data downloaded and in place, we just need to select a model, an appropriate task and dataset as well related config file. Default configurations can be found inside configs folder in repository's root folder. Configs are divided for models in format of [task]/[dataset_key]/[model_key].yml where dataset_key can be retrieved from the table above. For example, for pythia model, configuration for VQA 2.0 dataset can be found at configs/vqa/vqa2/pythia.yml. Following table shows the keys and the datasets supported by the models in Pythia's model zoo.

For running LoRRA on TextVQA, run the following command from root directory of your pythia clone:

cd ~/pythia
python tools/run.py --datasets textvqa --model lorra --config configs/vqa/textvqa/lorra.yml 

Note for BUTD model : for training BUTD model use the config butd.yml. Training uses greedy decoding for validation. Currently we do not have support to train the model using beam search decoding validation. We will add that support soon. For inference only use butd_beam_search.yml config that supports beam search decoding.

We are including some of the pretrained models as described in the table above. For e.g. to run the inference using LoRRA for TextVQA for EvalAI use following commands:

# Download the model first
cd ~/pythia/data
mkdir -p models && cd models;
# Get link from the table above and extract if needed
wget https://dl.fbaipublicfiles.com/pythia/pretrained_models/textvqa/lorra_best.pth

cd ../..
# Replace datasets and model with corresponding key for other pretrained models
python tools/run.py --datasets textvqa --model lorra --config configs/vqa/textvqa/lorra.yml \
--run_type inference --evalai_inference 1 --resume_file data/models/lorra_best.pth

The table below shows inference metrics for various pretrained models:

Note that, for simplicity, our current released model does not incorporate extensive data augmentations (e.g. visual genome, visual dialogue) during training, which was used in our challenge winner entries for VQA and VizWiz 2018. As a result, there can be some performance gap to models reported and released previously. If you are looking for reproducing those results, please checkout the v0.1 release.

1. If setup.py causes any issues, please install fastText first directly from the source and then run python setup.py develop. To install fastText run following commands:

git clone https://github.com/facebookresearch/fastText.git
cd fastText
pip install -e .

Pythia is licensed under BSD license available in LICENSE file