Selective Amnesia (SA) for VAEs

Requirements

Install the requirements using a conda environment:

conda create --name sa-vae python=3.8
conda activate sa-vae
pip install -r requirements.txt

First train a conditional VAE on all 10 MNIST classes.
```
CUDA_VISIBLE_DEVICES="0" python train_cvae.py --config mnist.yaml --data_path ./dataset
```
where you should pass the GPU ID into CUDA_VISIBLE_DEVICES for GPU training, or pass an empty string for CPU training. The checkpoint should be saved in a folder that looks like results/yyyy_mm_dd_hhmmss, where yyyy_mm_dd_hhmmss corresponds to the date and time that your training run was started.

Calculate the FIM.

CUDA_VISIBLE_DEVICES="0" python calculate_fim.py --ckpt_folder results/yyyy_mm_dd_hhmmss

The FIM should be saved as fisher_dict.pkl in the same folder.

Forgetting training with SA
```
CUDA_VISIBLE_DEVICES="0" python train_forget.py --ckpt_folder results/yyyy_mm_dd_hhmmss --label_to_drop 0 --lmbda 100
```
Another folder results/yyyy_mm_dd_hhmmss will be created with the forgetting VAE checkpoint and logging samples stored.

First train a classifier on the MNIST dataset
```
CUDA_VISIBLE_DEVICES="0" python train_classifier.py --data_path ./dataset
```
The classifier checkpoint should be saved in the folder classifier_ckpts.
Generate samples of the forgetting class on the VAE checkpoint trained with SA.
```
CUDA_VISIBLE_DEVICES="0" python generate_samples.py --ckpt_folder results/yyyy_mm_dd_hhmmss --label_to_generate 0 --n_samples 1000
```
The samples should be stored in results/yyyy_mm_dd_hhmmss/0_samples.

Evaluate with the classifier

CUDA_VISIBLE_DEVICES="0" python evaluate_with_classifier.py --sample_path results/yyyy_mm_dd_hhmmss --label_of_dropped_class 0

The average probability and average entropy should be printed to your screen as follows

Average entropy: 2.194569170475006  
Average prob of forgotten class: 0.058080864138901234