Ok, so I’m trying to train a model to eventually build out an app. I’m compiling my own training/validation/test sets. The app is going to be accessible via mobile, so the images will certainly be much, much bigger than needed.
The domain is a card/board game where photos will be processed to assess the game’s status in a web API. There are 81 classes, and it’s kind of a mix of “planet” and “dog breeds”. I made a training set of a few photos of each of the 81 classes isolated, but the final version of the app will have to classify 3-15 classes inside one photo. I haven’t finished compiling/labelling the test set yet. Once the classification is done, the rest of the app will just be basic computer science processing.
First of all, (I know, dumb question), do I need to re-size my images before training? I’m having trouble recalling the best practice involving choosing a size or what video Jeremy talks about that at any length. It always seems that the photos in the training/test sets we use in class and kaggle competitions are already sized accordingly, but mine aren’t. So I’m starting with photos that are 1920 pixels square (with really no reason to assume they won’t be much, much bigger in production). I know that images coming into my API will need to be resized, and I can do that once I know the best size. Is this just a factor to play with and see how it best performs, or is it better to do something fancy and make some special transforms for my situation?
I’m going to have an online meeting tomorrow about my project, but I’d rather not post too many details here as I’m planning to sell it eventually. I’ll post a link to a Zoom meeting here for tomorrow at 9am CST.
Any comments/suggestions are welcome here and all are welcome at the meeting tomorrow.