Capsule Networks
Implemented using TensorFlow
Based on the paper "Dynamic Routing Between Capsules"
Requirements
- python3
- TensorFlow
- NumPy
- Matplotlib
Dataset
- MNIST datasets You can download at download link and locate them in the './MNIST_data' directory.
How to run Training & Testing
Locate the MNIST datasets in the './MNIST_data' directory and just run capsNet.ipynb
Results
Accuracy
- The best error
| % | Training | Validation | Test |
|---|---|---|---|
| Implemented | 0.0273 | 0.641 | 0.6611 |
| Paper | - | - | 0.25 |
The performance measurement was done with a CapsNet with 3 routing iterations and reconstruction.
Sample MNIST Test Reconstruction
Reconstructions with less training(1 epoch)
Reconstructions with more trained(50 epoch)
- Sample MNIST test reconstructions of a CapsNet with 3 routing iterations.
- The images on the first row are the input images, and that on the second row are the reconstructions.
- We can see that the reconstructions preserves many of the details of the input while smoothing the noise.
- We can get more detailed reconstruction with more traind model.
What the individual dimensions of a capsule represent
- We can see what the individual dimensions represent by feeding a perturbed version of the activity vector to the decoder network and see how the perturbation affects the reconstruction.
- By the experiment, the representations of each dimension can be translated as follows.
- The dimension of the DigitCaps representation is tweaked by intervals of 0.05 in the range [−0.25, 0.25].
- Localized skew
- Scale of thickness
- Width
