Memory, Attention and Composition (MAC) Network for CLEVR/GQA from Compositional Attention Networks for Machine Reasoning (https://arxiv.org/abs/1803.03067) implemented in PyTorch

Requirements:

• Python 3.6
• PyTorch 1.0.1
• torch-vision
• Pillow
• nltk
• tqdm
• block.bootstrap.pytorch murel.bootstrap.pytorch

To train:

1. Download and extract either
   CLEVR v1.0 dataset from http://cs.stanford.edu/people/jcjohns/clevr/ or
   GQA dataset from https://cs.stanford.edu/people/dorarad/gqa/download.html

For GQA

cd data
mkdir gqa && cd gqa
wget https://nlp.stanford.edu/data/gqa/data1.2.zip
unzip data1.2.zip
mkdir questions
mv balanced_train_data.json questions/gqa_train_questions.json
mv balanced_val_data.json questions/gqa_val_questions.json
mv balanced_testdev_data.json questions/gqa_testdev_questions.json
cd ..
wget http://nlp.stanford.edu/data/glove.6B.zip
unzip glove.6B.zip
wget http://nlp.stanford.edu/data/gqa/objectFeatures.zip
unzip objectFeatures.zip
cd ..

2. Preprocessing question data and extracting image features using ResNet 101 (Not required for GQA)
   For CLEVR
   a. Extract image features

python image_feature.py data/CLEVR_v1.0

b. Preprocess questions

python preprocess.py CLEVR data/CLEVR_v1.0

For GQA
a. Merge object features (this may take some time)

python merge.py --name objects
mv data/gqa_objects.hdf5 data/gqa_features.hdf5

b. Preprocess questions

python preprocess.py gqa data/gqa

!CAUTION! the size of file created by image_feature.py is very large! You may use hdf5 compression, but it will slow down feature extraction.

3. Run train.py with dataset type as argument (gqa or CLEVR)

python train.py gqa

CLEVR -> This implementation produces 95.75% accuracy at epoch 10, 96.5% accuracy at epoch 20.

Parts of the code borrowed from https://github.com/rosinality/mac-network-pytorch and
https://github.com/stanfordnlp/mac-network.