Why my next topic I will be learning is ML-ops

While I'm still developing my neural transfer project. A skill that I'm lacking very dearly is publishing models and projects. I have released other projects but they are not ML projects. So I want to make a project in the near future. Where I can collect data from the user can improve the product. A lot of skills regarding implementing deep learning models. Are not advertised that much. They some course that I'm thinking to learn from. Full-stack deep learning. And a YouTube course on Papers with code. But learning some continuous integration and continuous deployment skills will be nice. While notebooks are great. They are only accessible to fellow nerds. If I want to share my work with the wider public then sharing them via a website or app may be better.

ML-ops is a new field. Which probably explains why it hasn’t been getting much attention. Only until recently. I guess there is a critical mass of people that can make decent models. But starting to learn that implementing those models in real life is a bit of a pain. SO they want to learn how to deploy those models more effectively. For me, I'm still learning how to create a good model consistently. While I don’t think I'm bad. I'm not sure if I can hold it on my own. I don’t know it may be some type of impostor syndrome. As when I make models I was using a template from somewhere. Like example code from a Pytorch tutorial.  While I mostly know what’s going on. Sometimes I'm a bit lost. I think the answer like Jeremy Howard says is to train more models.

While this issue should go away soon. I haven’t been spending more time on my projects. Meaning that is less time to iterate and learn. I need to increase the iteration process. Of my learning. Meaning I want to be creating more models. Right now its probably one personal project a month.  Again a lot of university work. Has slowed me down. But major deadlines have passed. So I should have more time for my projects.

For ML-ops, I guess I will be using things like Github-actions. Which is a tool I tend to see a lot. When people showing a few screenshots on Twitter. I guess they are other tools that I don’t know about. I think deep learning education online is still weighted towards the research side. Which is fine if you want to do research. And a lot of research is very interesting. But they are less focus on implementing and deploying those models. I think I wrote a blog post about this a while back. Where I want to focus more on deploying products. Which I did not do for the green tea and oolong tea classifier. But should do with future projects.

ML-ops may be an important skill for me to learn. If I want to start making software I want to sell in the near future. It is unlikely to be a SasS product. Due to the fact, sass products take a lot of time and energy. And most importantly I have no experience selling products online. So it is better to get that experience on my belt first before I do anything crazy. This product could be a small ML model that people can pay a small fee to use. I don’t know. Right now I'm just thinking out loud.

While learning deep learning is cool. If I do want to show my wares. Then they should be accessible for non-technical people. A lot of the projects that go viral tend to be highly visual like toonify. And/or an easy way to play with the model. Like a website or app. This is not a real person.com was a website that produced fake people from GANs. Or the user had to do was refresh the webpage. Then a new fake person appeared. If the knowledge was stored in a notebook only a few people can access it. Even worst a research paper. Where a very small amount of people can understand and compared what's going on.

A lot of problems when it comes to me learning deep learning. Is simply increasing my output. I just need to pump out for stuff. Pytorch has allowed me to finish projects easier. But I still need to do more work.

ML-Ops is a topic I don’t think I can just read up. So it's likely I will need to apply them to many of the projects, that I will be working on in the future. Can't imagine me using some MNIST dataset and deploying that to Github. ML-Ops frankly is used for real deep learning projects. I don’t know how long learning the topic will take. But I will be happy to add that skill into my arsenal. As that means I can deploy ML models to the web. In good condition. With users even adding feedback to the model. That can be the backbone of many great products in the future.

Or maybe. I just get distracted again and start learning about a new topic. To be honest, AR is looking very interesting right now. So I may work on that pretty soon.