FigureQA-Baseline

This repository contains the TensorFlow implementations of the CNN-LSTM, Relation Network and text-only baselines for our paper FigureQA: An Annotated Figure Dataset for Visual Reasoning [Project Page].

If you use code from this repository for your scientific work, please cite

Kahou, S. E., Michalski, V., Atkinson, A., Kadar, A., Trischler, A., & Bengio, Y. (2017). Figureqa: An annotated figure dataset for visual reasoning. arXiv preprint arXiv:1710.07300.

If you use the Relation Network implementation, please also cite

Santoro, A., Raposo, D., Barrett, D. G., Malinowski, M., Pascanu, R., Battaglia, P., & Lillicrap, T. (2017). A simple neural network module for relational reasoning. In Advances in neural information processing systems (pp. 4974-4983).

Getting started

The setup was tested with python 3, tensorflow 1.4 and 1.6.0-rc1. We recommend using the Anaconda Python Distribution.

Create a virtual machine, e.g. via

conda create -p ~/venvs/figureqa python=3

Activate the environment:

source activate ~/venvs/figureqa

Install dependencies:

conda install numpy tqdm six matplotlib pandas
pip install tensorflow-gpu

Download the FigureQA data set tar.gz archives (unextracted) into a directory named FigureQA.
Clone the baseline repository somewhere locally (here we're using $HOME/workspace)

mkdir -p ~/workspace
cd ~/workspace
git clone git@github.com:vmichals/FigureQA-Baseline.git

Training and Evaluation

Training a model

Run the training script for the model. It takes the following required arguments:

--data-path: the directory, in which you placed the tar.gz archives of FigureQA, referred to as DATAPATH in the following.
--tmp-path: a temporary directory, in which the script will extract the data (preferably on fast storage, such as an SSD or a RAM disk), from now on referred to as TMPPATH
--model: the model you want to train (rn, cnn or text), from now on referred to as MODEL
--num-gpus: the number of GPUs to use (in the same machine), from now on referred to as NUMGPU
--val-set: the validation set to use for early-stopping (validation1 or validation2), from now on referred to as VALSET
(additional configuration options can be found in the *.json files in the cfg subfolder)

cd ~/workspace/FigureQA-baseline
python -u train_model_on_figureqa.py --data-path DATAPATH --tmp-path TMPPATH \
    --model MODEL --num-gpus NUMGPU --val-set VALSET

Resuming interrupted training

To resume interrupted training, run the resume script, which takes the following required arguments:

--data-path: same as for the training script
--tmp-path: same as for the training script
--train-dir: the training directory created by the training script (a subfolder of the train_dir), from now on referred to as TRAINDIR
--num-gpus: same as for the training script
--resume-step: the time-step from which to resume (check training directory for the model-TIMESTEP.meta file with the largest TIMESTEP), from now on referred to as RESUMESTEP

cd ~/workspace/FigureQA-baseline
python -u resume_train_model_on_figureqa.py --train-dir TRAINDIR --resume-step RESUMESTEP \
    --num-gpus NUMGPU --data-path DATAPATH --tmp-path TMPPATH

Testing

To evaluate a trained model, run the eval script, which takes the following required arguments:

--train-dir: same as for the resume script
--meta-file: the meta file of the trained model, e.g. "model_val_best.meta" from the
--data-path: same as for the resume and training script
--tmp-path: same as for the resume and training script

Example:

cd ~/workspace/FigureQA-baseline
python -u ./eval_model_on_figureqa.py  --train-dir TRAINDIR --meta-file METAFILE \
    --partition test2 --data-path DATAPATH --tmp-path TMPPATH