SARN: Sequential Attention Relational Network

This is Pytorch implementation of our work "SARN: Relational Reasoning through Sequential Attention ". We tried to solve exploding relations(quadratic) problem of Relational Network module by paying attention sequentially.

Requirements

• Pytorch 0.4

Benchmarks

Baseline

python main.py --model baseline

Relational Network

python main.py --model rn

FILM

python main.py --model film

SARN

python main.py --model sarn

SARN att(Work in progress)

python main.py --model sarn_att

Results

Results of benchmarks with varying size of image and objects on modified Sort-Of-Clevr dataset.\

model	image - object	64 - 4	64 - 5	75 - 5	128 - 4	128 - 5	128 - 8	128 - 16
baseline	non-rel	0.9941	0.9933	0.9961	0.9883	0.9978	0.9883	-
baseline	rel	0.812	0.8163	0.8205	0.7227	0.8532	0.7227	-
baseline	total	0.8803	0.8827	0.8864	0.8223	0.9074	0.8223	-
rn	non-rel	0.9944	0.9964	0.9961	0.9922	0.9931	0.9922	0.9719
rn	rel	0.8415	0.843	0.8517	0.826	0.7719	0.826	0.7708
rn	total	0.8989	0.9005	0.9058	0.8883	0.8555	0.8883	0.8462
sarn	non-rel	0.997	0.9948	0.9986	0.9992	0.9988	0.9992	0.9819
sarn	rel	0.8949	0.837	0.9	0.931	0.8669	0.931	0.8533
sarn	total	0.9345	0.897	0.9347	0.9566	0.9163	0.9566	0.9016

Reference

• MHDPA: https://github.com/jadore801120/attention-is-all-you-need-pytorch