---

Focal Adversarial Autoencoder (FAAE) Code Documentation

Introduction

The Focal Adversarial Autoencoder (FAAE) is a cutting-edge deep learning model tailored for normative modeling in the medical field. It specializes in establishing a norm based on a healthy control group and evaluates individual deviations to assess patient conditions. FAAE's training exclusively involves data from a healthy group and assesses deviations among both healthy and patient groups during testing.

This model encompasses six methods: AE, AAE, VAE, CVAE, ACVAE, and FAAE.

How to Run the Code

Conda Environment Setup

1. Dependencies Installation:

   • Conda:

     conda env create -f environment.yml

   • Pip:

     pip install -r requirements.txt

   Note: The environment.yml might contain redundant dependencies. Feel free to create your custom environment with major library versions.

Data Preparation

1. Data Format: Feature vector size in ADNI is 1x100.

2. Data Storage:

   • Training data: PROJECT_ROOT/data/ADNI_TRAIN
   • Test data: PROJECT_ROOT/data/ADNI

3. File Details:

   • Features and indices are in freesurferData.csv.
   • Labels and conditions are in participants.tsv.

4. Index Files:

   • Training data's indices: PROJECT_ROOT/outputs/cleaned_ids.csv
   • Test data's indices: PROJECT_ROOT/outputs/ADNI_homogeneous_ids.csv

Training and Testing

1. Scripts Execution:

   • Grant execution permissions:

     chmod +x *

   • Run training, testing, and analysis scripts sequentially for each model:

   # Example for AE
   ./bootstrap_train_ae_supervised.py
   ./bootstrap_test_ae_supervised.py
   ./bootstrap_ae_group_analysis_1x1.py
   
   # Replace AE with AAE, VAE, CVAE, or ACVAE as needed

FAAE Specific Scripts

• Hyperparameter Fine-Tuning:
  • commands_list1.sh: Tunes alpha, gamma, r values.
• Baseline and FAAE Model Training, Testing, and Analysis:
  • commands_list2.sh

Code for ACVAE and FAAE

• The ACVAE and FAAE share the same codebase, differentiated by parameter settings.
• For ACVAE: Set parameters -A 0 -G 1 -R 0.
• For FAAE: Vary alpha (-A), with alpha > 0 and < 1, and gamma (-G) > 1.

---