Recently, Efficient-CapsNet was released. I sure like capsule-net based neural networks, but toy datasets as MNIST are simply too easy. Therefore, I tried it on CMNIST which is a tiny bit more challenging.

All results were obtained used the MNIST settings provided in the Efficient-CapsNet repo. The only thing modified was the data loader.

efficient capsnet results after 2,75 and 300 epochs on cmnist-x datasets

Given that the model is only of 6 MB, the results look quite good - especially given the speed-up compared to earlier CapsNet versions.