Most of the day (yesterday) was spent playing around with PyTorch in order to create a simple CNN (shoutout to Kamal from Data Science!!). We had a data set of images of different rooms and we wanted to classify them into six different categories, such as bedroom, living room, kitchen, and etc. With the state-of-the-art model that he worked on, it had about 85% accuracy so we aimed for roughly 50% (with the baseline being around 16.7% with just random selection) as the model we would make would not be as complex.
We first created a feed-forward network but eventually added some convolutional layers so I’ll mostly talk about that, especially as we had a problem with the weights not updating thanks to the great documentation by the developers of PyTorch. There’s a very important fact that they left out in the documentation that took up about two hours of our time to solve… maybe I’ll talk about that when I get to it.
However I think running into bugs is usually an ideal situation to learn a lot of new things (like how I should email the devs and tell them their documentation is slightly wrong or at least missing something). We got a 47% accuracy with two convolutional layers, which was actually a bit more decent than we thought, especially as it peaked at about 4 epochs. So, goal semi-achieved. I was actually going to write about this in more detail, however I didn’t have enough time yesterday and there’s a lot I want to get down on to paper today.
After some discussions yesterday and further discussions today, I began thinking about a probably masterpiece that I could do alongside the data science team. As I said, there’s not too much time for me to think about what I want to do so I’ll be planning it the next two weeks so I can start off the next phase starting on the project immediately rather than losing a week or two on planning.
Let’s see… We have a ton of images (many different kinds) and we have textual data such as reviews, which can contain positive and negative aspects about a certain hotel. After having a meeting with one of the UX leads this morning, I began to think about the business side of things and what I can do with these data that would help the end-users have a better experience on our website.
The software engineer in me wants to gather requirements before collecting ideas. I mean, that’s always the first step in the creation of something. It makes life easier if we know what the end-users want and look for. It would help me/the data team tune an idea in a way that is doable for us as well as beneficial to the company.
Anyway, I’m still at a point where I’m collecting some information and letting it all marinate in my head.
Just me rambling about things that are in my head. Somewhat the usual.
Mein Weihnachtsbaum wurde heute morgen geliefert!