layog / capsnets-tensorflow

Implementation of Capsule networks (With Dynamic Routing Algorithm) based on Aurélien's implementation

capsnets-tensorflow

Implementation of Capsule networks (With Dynamic Routing Algorithm) based on Aurélien's implementation

Original notebooks and videos

• Jupyter Notebook
• Capsule Network Explanation
• Implementation Explanation

Improvements

• The Dynamic Routing Algorithm is implemented in the original notebook without a loop and by creating new operations in the graphs. This limits the use of the routing algorithm as (I've changed it to use tf.while_loop)
  1. We cannot dynamically increase the number of routing loops
  2. Takes up much more space in the GPU
• In the Aurelien's notebook, many times tile function is used, which is not very efficient and hence I've replaced it with einsum

---

Installation

(Run the following in a python=2.7 environment, although should work on python3 as well, but I have not tested it yet)

• pip install -r requirements.txt
• (Optional) python -m ipykernel install --user --name <desired-kernel-name>

Evaluating the notebooks

• You can download the final tensorflow saver files from here. Keep the files in the root directory of the project and skip the training block in the notebook (for evaluation).
• Or, just run the complete notebook which will train the network as well

If you are using the latest tensorflow (1.5.0), you will encounter some warnings (regarding some argument deprecation). They won't break anything for now.