This repository contains code used to produce the results in the following paper:
Ranjay Krishna, Michael Bernstein, Li Fei-Fei
IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019
If you are using this repository, please use the following citation:
@inproceedings{krishna2019information,
title={Information Maximizing Visual Question Generation},
author={Krishna, Ranjay and Bernstein, Michael and Fei-Fei, Li },
booktitle={IEEE Conference on Computer Vision and Pattern Recognition},
year={2019}
}
I have most likely introduced errors while making this public release. Over time, I will fix the errors.
You can clone the repository and install the requirements by running the following:
git clone https://github.com/ranjaykrishna/iq.git
cd iq
virtualenv -p python2.7 env
source env/bin/activate
pip install -r requirements.txt
git submodule init
git submodule update
mkdir -p data/processed
To download the dataset, visit our website.
Note that we only distribute the annotations for the answer categories. To download the images for the VQA dataset, please use the following links:
To train the models, you will need to (1) create a vocabulary object,
(2) create an hdf5 dataset with the images, questions and categories,
(3) and then run train and evaluate scripts:
# Create the vocabulary file.
python utils/vocab.py
# Create the hdf5 dataset.
python utils/store_dataset.py
python utils/store_dataset.py --output data/processed/iq_val_dataset.hdf5 --questions data/vqa/v2_OpenEnded_mscoco_val2014_questions.json --annotations data/vqa/v2_mscoco_val2014_annotations.json --image-dir data/vqa/val2014
# Train the model.
python train_iq.py
# Evaluate the model.
python evaluate_iq.py
This script will train the model and save the weights in the --model-dir
directory. It will also save the configuration parameters in a
args.json file and log events in train.log.
However, if you decide that you want more control over the training or evaluation scripts, check out the instructions below.
The vocabulary object you create contains , , , tokens and decides which objects to include in the vocabulary and which to consider as . You can customize the creation of this vocabulary object using the following options:
-h, --help Show this help message and exit.
--vocab-path Path for saving vocabulary wrapper.
--questions Path for train questions file.
--answer-types Path for the answer types.
--threshold Minimum word count threshold.
The dataset creation process can also be customized with the following options:
-h, --help Show this help message and exit.
--image-dir Directory for resized images.
--vocab-path Path for saving vocabulary wrapper.
--questions Path for train annotation file.
--annotations Path for train annotation file.
--ans2cat Path for the answer types.
--output Directory for resized images.
--im_size Size of images.
--max-q-length Maximum sequence length for questions.
--max-a-length Maximum sequence length for answers.
The model can be trained by calling python train.py with the following command
line arguments to modify your training:
-h, --help Show this help message and exit.
--model-type [ia2q | via2q | iat2q-type | via2q-type | iq | va2q-
type]
--model-path Path for saving trained models.
--crop-size Size for randomly cropping images.
--log-step Step size for prining log info.
--save-step Step size for saving trained models.
--eval-steps Number of eval steps to run.
--eval-every-n-steps Run eval after every N steps.
--num-epochs
--batch-size
--num-workers
--learning-rate
--info-learning-rate
--patience
--max-examples For debugging. Limit examples in database.
--lambda-gen coefficient to be added in front of the generation
loss.
--lambda-z coefficient to be added in front of the kl loss.
--lambda-t coefficient to be added with the type space loss.
--lambda-a coefficient to be added with the answer recon loss.
--lambda-i coefficient to be added with the image recon loss.
--lambda-z-t coefficient to be added with the t and z space loss.
--vocab-path Path for vocabulary wrapper.
--dataset Path for train annotation json file.
--val-dataset Path for train annotation json file.
--train-dataset-weights Location of sampling weights for training set.
--val-dataset-weights Location of sampling weights for training set.
--cat2name Location of mapping from category to type name.
--load-model Location of where the model weights are.
--rnn-cell Type of rnn cell (GRU, RNN or LSTM).
--hidden-size Dimension of lstm hidden states.
--num-layers Number of layers in lstm.
--max-length Maximum sequence length for outputs.
--encoder-max-len Maximum sequence length for inputs.
--bidirectional Boolean whether the RNN is bidirectional.
--use-glove Whether to use GloVe embeddings.
--embedding-name Name of the GloVe embedding to use.
--num-categories Number of answer types we use.
--dropout-p Dropout applied to the RNN model.
--input-dropout-p Dropout applied to inputs of the RNN.
--num-att-layers Number of attention layers.
--use-attention Whether the decoder uses attention.
--z-size Dimensions to use for hidden variational space.
--no-image-recon Does not try to reconstruct image.
--no-answer-recon Does not try to reconstruct answer.
--no-category-space Does not try to reconstruct answer.
The evaluations can be run using python evaluate.py with the following options:
-h, --help Show this help message and exit.
--model-path Path for loading trained models.
--results-path Path for saving results.
--preds-path Path for saving predictions.
--gts-path Path for saving ground truth.
--batch-size
--num-workers
--seed
--max-examples When set, only evalutes that many data points.
--num-show Number of predictions to print.
--from-answer When set, only evalutes iq model with answers;
otherwise it tests iq with answer types.
--dataset Path for train annotation json file.
We welcome everyone to contribute to this reporsitory. Send us a pull request. Feel free to contact me via email or over twitter (@ranjaykrishna).
The code is under the MIT license. Check LICENSE for details.