This repository is cloned version of Efficient-CapsNet which contains code that we developed to train model for categorizing Japaneese characters Kuzushiji-MNIST dataset.
- Clone this repository
git clone https://github.com/bukson/kmnist-efcaps
- Make sure you have installed proper version of cuda on your system
cuda-11.8 cudnn-8.6.0.
- Install the required packages
pip3 install -r requirements.txt
Out purpose is to used modified network architecture and components from Efficient Capsule Networks to create model for Kuzushiji-MNIST dataset. We show that it is possible to make deep capsule architecture, train it on consumer grade pc setup in few gpu hours on dataset with high number of classes and achieve competitive results comparing to more complicated and expensive architecture (4th place for KMNIST and 6th place for Kuzushiji-49) Kuzushiji-MNIST benchmark.
We also like to popularize Efficient Capsule Networks as it is extremely well written piece of software, that can be easily modified by other scientist and repurpose or use to check others result. This is in our knowledge a rare case among other Capsule Network modifications.
In notebooks directory you can find notebooks that we used to train our models.
kmnist_train_test.ipynb contains code for training and testing model on KMNIST dataset.
k49_train_test.ipynb contains code for training and testing model on K49 dataset.