hdjang / visual_question_answering_relation-networks_film Goto Github PK

This repository implements and reproduces two VQA models on Sort-of-CLEVR:

• Relation Networks (RNs)
• FiLM

To use this repo, please download the dataset (Sort-of-CLEVR) below "/data" directory

Train

• To train RNs

./scripts/train_RNs.sh

• To train FiLM

./scripts/train_FiLM.sh

Eval

For evaluation, trained model-weights should be located at "./models/here".

• To evaluate RNs

./scripts/eval_RNs.sh

• To evaluate FiLM

./scripts/eval_FiLM.sh

Below is benchmark result. For fair comparison on relational reasoning ability, I used same CNN visual feature extractor for both models.

• Ho-Deok Jang
• Email: [email protected]
• Homepage: https://sites.google.com/view/hdjangcv