/caps-em-routing-cifar

The code implements Hinton's matrix capsule with em routing for Cifar-10 dataset

Primary LanguagePythonMIT LicenseMIT

Matrix Capsules with EM Routing

A PyTorch implementation of Matrix Capsules with EM Routing

Usage

  1. Install PyTorch

  2. Start training (default: CIFAR-10)

python train.py --batch-size 20 --test-batch-size 20

CIFAR-10 experiments

The experiments are conducted on Tesla K80 Specific setting is lr=0.01, batch_size=20, weight_decay=0, Adam optimizer, without data augmentation.

Following is the result after training:

Arch EM-Iters Coord Add Loss Epochs Test Accuracy
A=B=C=D=32 2 Y Spread 10 62.480
A=64 B=8 C=16 D=16 2 Y Spread 10 60.17000
A=32 B=24 C=32 D=24 2 Y Spread 35 70.1100
A=B=C=D=32 2 Y Spread 35 71.16000
A=64 B=8 C=16 D=16 2 Y Spread 35 71.5200

Reference