Pytorch implementation of " A simple neural network module for relational reasoning" paper aka Relational networks for visual reasoning. https://arxiv.org/abs/1706.01427

This implementation includes only the visual pipeline for CLEVR dataset. Best validation accuracy acheived with this implementation is 72% compared to 96.8% reported in the paper. This result was acheived by applying a learning rate schedule that doubles the learning rate every 20 epochs (motivated by warmup in https://arxiv.org/abs/1706.02677). The paper itself does not discuss any schedules used, running with schedules gets 65% at best.

Pull requests and suggestions are welcome to reproduce the results from the paper.

• Pytorch
• Visdom
• H5py
• opencv-python

python3 runtime

Arguments

• lr : Learning rate. default: 2.5e-4
• batch_size: default : 64
• warmup: A flag to turn on doubling the learning rate every 20 epochs. default: False
• save_path: path to checkpoints. Checkpoints are saved for every new best validation accuracy.
• vis_screen: Visdom env name. default: RelNet

• https://github.com/rosinality/relation-networks-pytorch
• https://github.com/mesnico/RelationNetworks-CLEVR