Santoro et al. A simple neural network module for relational reasoning. NIPS, 2017. (in PyTorch)

Perez et al. FiLM: Visual Reasoning with a General Conditioning Layer. AAAI, 2018 (in PyTorch)

Description

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

• Relation Networks (RNs)
• FiLM

Get Started

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

Train/Eval

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

Benchmark (Sort-of-CLEVR)

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

models	#param	Accuracy (relational qst)[%]	Accuracy (non-relational qst)[%]
RNs	1.38M	92	99
FiLM	1.72M	94	99

Contact

• Hodeok Jang
• Email: openhd34@gmail.com