kFactorVAE: Self-Supervised Regularization for Better A.I. Disentanglement

https://scholarworks.wm.edu/honorstheses/2040/

This repo. contains all work conducted for my honors thesis project at William & Mary. It contains a few subdirectories for different VAE-based models I investigated adding my regularization term into, notably kFactorVAE in the kFactorVAE folder, Beta-VAE from the Disentangling folder, and Beta-TCVAE from the beta-tcvae folder.

It borrows from Professor Shao's ControlVAE GitHub repository. Note though that I do not use the ControlVAE model itself in kFactorVAE, although it is an open avenue!

Dependencies

I used a conda environment and the YML file for its CUDA dependencies can be found here and for its non-CUDA dependencies here.

Reproducing Thesis Paper Results

Hardware Setup

If you are affiliated with William & Mary, I recommend using the lab machines in McGlothlin-Street (McG) Hall. To remotely log onto these machines, if you do not possess a CS account already and/or you are new to using the machines, follow the guidance provided by Professor Timothy A. Davis here.

Otherwise, I recommend using a computer that has a NVIDIA GPU and is compatible with NVIDIA CUDA. The NVIDIA GPU models I used were:

1. NVIDIA RTX A4000 (fastest at around 70 training iterations/second)
2. NVIDIA A40 (middle at around 50-70 training iterations/second)
3. NVIDIA RTX A5000 (slowest at around 30-40 training iterations/second)

I made no typos. Surprisingly, yes, the A4000 is faster than the A5000 even though the specifications tell me otherwise.

And of course, all these models are available on the lab machines. GPU specifications for each machine may be found here. As years pass by, these computers will probably have significant upgrades in their GPUs. However, you can probably expect these GPUs to stick around if you are reading this in the year 2023 or a few years afterwards. But maybe I will be proven wrong.

Software Instructions

Visdom Server Initialization

You will need to perform this step if you want to see the training metric graphs, reconstructions, and latent traversals when reproducing experimental results on a convenient website interface. All these results are also stored in directories.

On a (Linux) shell, run:

❯ chmod +x run_visdom_server.sh

❯ run_visdom_server.sh [port number] [optional relative path to Visdom log file to replay]

You may replay as many log files as you want. Keep in mind you will get a .out file every time run run_visdom_server.sh.

kFactorVAE Scripts Setup

You may head over to the kFactorVAE README.

Technical Support for McG Hall Computers

If you have any issues/requests, please reach out to Joseph Hause in the W&M Computer Science Slack.

Useful Tidbits Learned

This is just for personal memory's sake or for your own benefit too, in case you were curious as to the minor techniques I have acquired.

(1) How to make a forked repository private

• This was initially a challenge since by default, GitHub does not allow you to make a forked repo. private.
• Source

# 1. Clone without any remote branches
git clone --bare [folder to repo. fork]

# 2. Create a new private repo. on the GitHub website

# 3. Push repo. contents + all branches 
cd [local repo. fork directory]
git push --mirror [copied link to private. repo you just created] 

# 4. Delete your local repo. fork
rm -rf .git
cd ..
rm -rf [folder to repo. fork]

# 5. Now just clone the private repo., and voila!

I needed to do this because I wanted to make a private clone of Professor Shao's ControlVAE and work with that private clone and finish the thesis before making it public.

(2) How the transpose operator works for higher dimensions

• It's sill switching "rows" and "columns." aka $a_{ij} = a_{ji}$
• But each element $a$ of the "matrix" you are transposing might something other than a singular number. An element could now be an entire matrix, an actual row/column, or a tensor.

Why I needed to learn this was to understand how Prof. Shao implemented the latent variable traversal (viz_traverse) and displayed the results to the visdom server.

(3) How to Search Within a Highlighted Text Section within Vim

1. Shift + V and select all that you want
2. You can yank or just press esc twice
3. /\%V[your_search_string]

(4) Shell scripting

• conditionals to check directories
• operations on directories
• checking if a port is being used by a specific process
• for loops
• variables
• realizing that ChatGPT is solid in shell scripting :)

(5) Literature Review

• When bogged down by the technical details, making a slides presentation, where each paper gets only one slide, with abstracts, conclusions, advantages, and disadvantages helped me maintain engagement with the literature. The literature review and processing a new idea in light of it was by far the most challenging part.

(6) Logging data for visualizations

• Visdom, CSV filewriting

(7) LaTeX

• Side-by-side plots, images, adjacent text to images to label factors of variation
• \input command to break down a LaTeX file into subfiles
• inserting PDFs