A tiny implementation of the recently published Capsule Network by G.Hinton and colleague ( paper link )
- A novel concept of 'Capsules' which is basically vector-neurons as opposed to scalar neurons (the familier ones)
- A routing algorithm (refer paper)
- a not-so-deep network of capsules and normal convolutions
- state-of-the-art performance on MNIST
My implementation is still running on a 1050-TI GPU and hopefully is correct. I will add the results as soon as I can
Validation Accuracies
| Epoch-1 | Epoch-2 | Epoch-3 | Epoch-4 | Epcoh-5 | Epoch-6 | Epoch-7 | Epoch-8 |
|---|---|---|---|---|---|---|---|
| 96.2% | 98.27 | 98.82 | 98.89 | 99.01 | 99.07 | 99.15 | 99.27 |
Okay I am not running anymore before updating the code
A screenshot from tensorboard
Reconstructions: (I produced the reconstruction images from the end of the regularizing decoder and this is how they look like)
Although, my goal is not to achieve highest test-accuracy but to understand the dynamics of the network and the capsules
There is an amazing implementation (github.com/naturomics/CapsNet-Tensorflow)[https://github.com/naturomics/CapsNet-Tensorflow]. Go check it out. It helped me a little