Accompanying code for "A Simple Loss Function for Improving the Convergence and Accuracy of Visual Question Answering Models" CVPR 2017 VQA workshop paper.
The repo contains code for reproducing the paper's experiments and efficient GPU implementation of the proposed loss function for torch, pytorch, and caffe.
Requirements
To run the experiments you would first need to install torch from https://github.com/torch/distro/. We used torch version from commit 5c1d3cfda8101123628a45e70435d545ae1bc771 but later versions probably would work too.
After installing torch you will need to install the following useful lua libraries:
C data structures for torch https://github.com/torch/tds, so we can allocate data in C memory space instead of lua's and thus avoid lua's memory limit and garbage collection.
luarocks install tds
RNN lib for torch https://github.com/Element-Research/rnn for mask zero lookuptable and other useful modules.
luarocks install rnn
threads for lua https://github.com/torch/threads for multi-threaded code.
luarocks install threads
The following libraries are required but you can modify the code and still run the experiments. However we recommend installing them anyway.
fb-debugger a source-level debugger for lua
Follow the install instructions at https://github.com/facebook/fblualib/blob/master/fblualib/debugger/README.md.
OptNet - Reducing memory usage in torch neural nets https://github.com/fmassa/optimize-net.
luarocks install optnet
Visdom for visualization https://github.com/facebookresearch/visdom.
pip install visdom
luarocks install visdom
Installation
We provide GPU implementation of the loss function for torch, pytorch, and caffe.
cd to loss_implementations to read further instructions on how to add the loss function to your framework's installation.
Experiments
First, head over the image_preprocess folder and follow the instructions there to extract feature tensors for MS COCO images.
After obtaining image feature tensors, head over experiments/pool or experiments/avg to run the experiments reported in the paper.
To run the models with the proposed loss function run:
gpu=0 ./run_soft_cross_entropy.sh
For standard cross entropy run:
gpu=0 ./run_cross_entropy.sh
gpu=0 specifies the ID of the GPU to be used i.e. it's an alias for CUDA_VISIBLE_DEVICES=0.