All packages required to run the code are contained in environment.yml.

Project 1: Interpretable and Explainable Classification for Medical Data

The code for this project is in the 'explainability' folder. The folder structure is as follows:

Part 1: Heart Disease Prediction Dataset

All code for this part is included in heart_disease_prediction.ipynb, the data and column descriptions can be found on Kaggle. The implementation of Neural Additive Models is contained in the folder /nam and taken from Google Research github.com/lemeln/nam.

Part 2: Pneumonia Prediction Dataset

Q1-Q3: Data Analysis, CNN Classifier and Integrated Gradients

All code for this part can be found in cnn_imaging.ipynb.

Q4: Grad-CAM

For Q4, use the cnn_grad_cam.ipynb file. Be sure to download our trained model from the polybox that is linked in the notebook. The ten images we used for visualisation can also be found in a polybox. The link is also at the beginning of the notebook. Make sure that you specify all paths correctly. The pictures on polybox are in a folder called 'img_for_saliency' and are grouped into subfolders based on their classes for the 'ImageFolder' to assign labels correctly. The "CNN_grad_cam" class and all other functionality are defined in the grad_cam.py file, and make sure the functions you are using in the notebook are imported from this file.

Q5: Data Randomization Test

For the permutation test for Grad-Cam in Q5, you need the grad_cam_permutation_test.ipynb file. This notebook builds on the permutated model and 10 test images. For all those two components, you can find the link to a polybox at the beginning of the notebook. Ensure the folder structure is correct so the models are correctly imported and the 'ImageFolder' assigns the labels correctly. Again, from the grad_cam.py file, you need the class "CNN_grad_cam" and the function display_datasets_heatmap to visualise the Grad-Cam saliency maps with the permuted model. In order the perform the data randomization test on Integrated Gradients, refer to cnn_permuted.ipynb. This mostly follows the structure of the Grad-Cam notebook using techniques and packages introduced in 2.3.

Project 2: Time Series and Representation Learning

Part 1: Supervised Learning on Time Series

All code for this part is included in `time_series" folder. We use the following datasets: ThePTBDiagnostic

● MIT-BIH Arrhythmia Database (https://physionet.org/physiobank/database/mitdb/)

● ECG Database (https://physionet.org/content/ptbdb/1.0.0/)

The folder structure is as follows:

Q1: Exploratory Data Analysis, Q2: Classical machine learning methods

All code for this part can be found in eda1.ipynb and classic_ml1.ipynb.

Q3: Recurrent Neural Networks

Code to train RNN with LSTM can be found in rnn.ipynb and rnn.py contains model class and training function. Code for the bidirectional LSTM can be found under ts_bilstm.ipynb. The model is under models/bilstm.pth.

Q4: Convolutional Neural Network

Code to train the two CNN models (vanilla and with residual blocks) can be found at ts_cnn.ipynb. The corresponding models are stored at models/vanillacnn.pth and models/residualcnn.pth.

Q5: Attention and Transformers

Code to train the Transformer and visualize the attention maps can be found in transformer.ipynb. The model classes and training function can be found in transformer.py.

Part 2: Transfer and Representation Learning

Q1: Supervised Model for Transfer

Transfer learning for the CNN encoder is covered in the first half of the notebook umap.ipynb. We create the model transfercnn.pth. The embeddings are stored in the polybox under https://polybox.ethz.ch/index.php/s/bu0w5P5x9DHU86G

Q2: Representation Learning Model

Our autoencoder model is defined in models/encoder.py. The training and evaluation is done in representation_learning.ipynb.

Q3: Visualising Learned Representations

Code for the UMAP representations for the CNN encoder can be found under umap.ipynb. Our code for the kulback-leibler divergence is found in models/kl.py The representations of the autoencoder are visualised in viualize_encoder.ipynb

Q4: Fine-tuning Strategies

The notebook fine_tuning_encoder_q1.ipynb contains code for classic ML and MLP strategies using encoder from Q1 and fine_tuning_encoder_q2.ipynb using encoder from Q2 (Autoencoder).