Hand Bone Age Assessment Using Stacked Capsule Autoencoders

This project is forked from torch-scae: PyTorch implementation of Stacked Capsule Auto-Encoders.

A joint work with Nader Karayanni.

Installation

# clone project   
git clone https://github.com/bdsaglam/roymatza/torch-SCAE-handbone-age-assesment
 
# install project   
cd torch-SCAE-handbone-age-assesment
pip install -e .

Reproducing the Handbone experiments:

1. Download the needed utils from SCAE-utils repo.

# clone project   
git clone https://github.com/roymatza/SCAE-utils

2. Activate the attached conda env (using environment.yml file), and run the following commands to support visualizing the results and progress in the notebook:

jupyter nbextension enable --py widgetsnbextension
jupyter labextension install @jupyter-widgets/jupyterlab-manager

• You can acquire the data needed for the experiments from : https://www.kaggle.com/kmader/rsna-bone-age

• The torch_scae folder contains the implementation of the SCAE architecture using the lighting module interface. This folder contains all the logic described in the original SCAE paper (AR Kosiorek & al., 2019). The torch_scae_experiments contains specific models configurations for experiments, for example for our experiments we created a folder named 'boneage' which contains all the relevant info specifically defined for the corresponding task (Hyper params,images input sizes , data manipulation files etc...)

3. Process the data:

• Using the provided script (process-data.py, edit the script with your folders for the data and run it to get a new folder with the processed data.
• Create the folders needed to get the path "\src\torch_scae\data\boneage\train" and inside it place the processed data.
• Create the path \src\torch_scae\data\boneage\test\test0" and move some arbitrary image to that folder (this is required for the code to have a test data even though it is not used in the notebook).
• Run arrange_folders_by_age.py script with the corresponding train directory with the processed data. The data should be split into labels folders within the same folder.

4. After that you can simply run the train.ipynb jupyter notebook from "src" and the experiments are done within the notebook's cells.

Additional Notes:

• Make sure to fill the needed fields in the notebook accordingly such as the paths (marked with TBD in the notebook).
• Checkpoints and tensorboards are saved in "/src/lightning_logs".
• To get the optimal run that we know of, it is needed to manually change a boolean variable is_reg_reset to True inside stacked_capsule_autoencoder.py after several epochs, and then rerun the experiment from the last checkpoint.

Train with MNIST (original torch-scae repo)

PyTorch Lightning is used for training.

python -m torch_scae_experiments.mnist.train --batch_size 32 --learning_rate 1e-4

Results

Further information can be found in the project final report.