CapsNet-CNTK
March 13, 2018 ยท View on GitHub
A CNTK implementation of CapsNet based on the paper titled 'Dynamic Routing Between Capsules' by Sara Sabour, Nicholas Frost and Geoffrey Hinton.
Requeriments
- Python >=3.5
- CNTK 2.4
- Tensorboard (optional)
Architecture
Training
git clone https://github.com/southworkscom/CapsNet-CNTK.git
cd CapsNet-CNTK
python get_data.py
python main.py
Tensorboard
Visualize training progress with tensorboard:
tensorboard --logdir tensorboard
Then, Open your favorite browser and navigate to http://localhost:6006 (assuming you are running the training at localhost)
Results
Test Error
| Method | Routing | Reconstruction | MNIST (%) | Paper |
|---|---|---|---|---|
| Baseline | -- | -- | -- | 0.39 |
| CapsNet | 3 | no | 0.55 | 0.35 (0.036) |
| CapsNet | 3 | yes | WIP | 0.25 (0.005) |
Reconstruction layer
Dimension perturbations
|
|
|
|
|
|
|
|
TODO
- Improve performance on CPU
- Complete benchmarks
- Capsule visualization
- Improve the API
- Try other datasets