Permalink
Switch branches/tags
Nothing to show
Find file Copy path
8cbc660 Sep 11, 2018
2 contributors

Users who have contributed to this file

@YuJiang01 @meetshah1995
73 lines (55 sloc) 2.25 KB

VQA v2.0

Download dataset

cd ../
mkdir -p orig_data/vqa_v2.0
cd orig_data/vqa_v2.0
./../../data_prep/vqa_v2.0/download_vqa_2.0.sh

Preprocess dataset

cd ../../VQA_suite 
mkdir data

export PYTHONPATH=.

python data_prep/vqa_v2.0/extract_vocabulary.py \
--input_files ../orig_data/vqa_v2.0/v2_OpenEnded_mscoco_train2014_questions.json \
 ../orig_data/vqa_v2.0/v2_OpenEnded_mscoco_val2014_questions.json \
 ../orig_data/vqa_v2.0/v2_OpenEnded_mscoco_test2015_questions.json \
--out_dir data/

python data_prep/vqa_v2.0/process_answers.py \
--annotation_file ../orig_data/vqa_v2.0/v2_mscoco_train2014_annotations.json \
--val_annotation_file ../orig_data/vqa_v2.0/v2_mscoco_val2014_annotations.json  \
--out_dir data/ --min_freq 9

python data_prep/vqa_v2.0/extract_word_glove_embedding.py  \
--vocabulary_file data/vocabulary_vqa.txt  \
--glove_file ../orig_data/vqa_v2.0/glove/glove.6B.300d.txt \
--out_dir data/

python data_prep/vqa_v2.0/build_vqa_2.0_imdb.py --data_dir ../orig_data/vqa_v2.0/ --out_dir data/

Download image features

cd data/
wget https://s3-us-west-1.amazonaws.com/pythia-vqa/data/rcnn_10_100.tar.gz
wget https://s3-us-west-1.amazonaws.com/pythia-vqa/data/detectron.tar.gz
gunzip rcnn_10_100.tar.gz 
tar -xvf rcnn_10_100.tar
rm -f rcnn_10_100.tar

gunzip detectron.tar.gz
tar -xvf detectron.tar
rm -f detectron.tar

Extract Image Features

We use detectron to extract image features. Setting up detectron and copy tools/extract_features.py to detectron for extracting features

Feature extraction works best with commit #3a38b7b of detectron and #0dd3284 of caffe2

download the pretrained detectron model

wget https://s3-us-west-1.amazonaws.com/pythia-vqa/detectron/FAST_RCNN_MLP_DIM2048_FPN_DIM512.pkl
wget https://s3-us-west-1.amazonaws.com/pythia-vqa/detectron/e2e_faster_rcnn_X-101-64x4d-FPN_1x_MLP_2048_FPN_512.yaml

python extract_features.py --cfg e2e_faster_rcnn_X-101-64x4d-FPN_1x_MLP_2048_FPN_512.yaml \
--wts FAST_RCNN_MLP_DIM2048_FPN_DIM512.pkl \
--min_bboxes 100 --max_bboxes 100 \
--feat_name gpu_0/fc6 \
--out_dir ~/temp_out $INPUT_DIR