Star Summit

I’ve recently bought into the hype over Generative Adversarial Networks and decided to create a GAN that can learn to generate sprite art. For this project I’m specifically training the network with sprites drawn from a JRPG style so I’m calling the project JRPG GAN (simple right?). Training this network will be particularly interesting for me since I remember the team I worked with for most of my college career seemed to be talented in many things, but art was not one of them. It unfortunately became a recurring problem since we liked video game based projects. I’m not expecting this project to create state of the art (*ahem*) art, but it would be cool to serve as an aid.

Currently, I have collected 4000+ sprites for training the network and that took a while to say the least. I’m testing out DCGAN and WGAN architectures for the training and it took quite a bit of work to get the training to stabilize for both, but so far I like where the training is headed. Once I figure out the architecture and hyperparameters I want to use, I will probably look into moving the training to AWS Sagemaker. My laptop can handle the training, but it’s also brand new and I don’t want to wear it down after just buying it.

Hopefully I’ll will be able to integrate this project into this site soon, but for now I will just give a sample from a training.

Epoch 0

Epoch 35

Before I talk about the new project, I gotta talk about the Reddit Upvote Ratio Predictor first! I have a functional model pipeline in place for the predictor and can now make predictions after giving info such as the title and body text of the submission. I see many places where I can improve the model, but my main goal is still to implement a MLOps pipeline so I’m working towards that first. I’ve chosen Heroku to host the FastAPI application I’ve made to serve predictions and this is where I plan to deploy to as part of the pipeline I mentioned. Hopefully I will be able to add place on this site to send requests to the app in not too long!

As for the new project I’ve picked up since:

See the next post, this is long enough already.

After nearly 4 years, I’ve returned to my abandoned website and with new project ideas that would actually be nice to host on this site, or at least better than my old p5.js college assignments I have on my projects tab.

I’ve decided to sharpen my knowledge in Machine Learning and MLOps by creating a ML application that can predict the success of a submission on the Paper Mario subreddit. Specifically, it would train a regression model that predicts the upvote ratio of a post given (but not limited to) data such as the title and body of the submission. For the model itself I plan to use a XGBRegressor and create features from the text fields using spaCy. The main focus for this project will be to enable the model to serve predictions for requests sent from this site and implement a full CI/CD pipeline for development on the model.

With that said, I’m gonna have to also brush up on my web development skills to host it on this site because like I said it’s been nearly 4 years…

My first semester of Senior year has ended and with my newfound time I figured that I should start making use of this website, otherwise I would just be wasting money. With that in mind, I think I’ll dedicate myself to updating this website more whether it be with new projects or changes to the site itself.