| Sanqui 2060 ![]() ![]() ![]() ![]() 💛🤍💜🖤🦉 Level: 87 Posts: 1859/2066 EXP: 6305067 For next: 87707 Since: 12-20-09 Pronouns: any ✨ From: Czechia | Estonia Since last post: 6 days Last activity: 1 day |
|
||
|
*Sanqui flops, wings involved. level 82 exp 5044033 exppct 23.7% numposts 1859
I have to get into machine learning (so I can pass my class) so I'm catching up on neural networks. One of the most basic tasks is building a binary image classifier. Most tutorials use cats and dogs, so I went ahead with something more interesting to me.
I built a dataset of 30.795 screenshots from Sonic 1 and Sonic 2 TASes and longplays and I'm currently training a neural network to predict whether a screenshot is from one or the other. Because real world data is fuzzy, I'm applying data augmentation, in particular, I mess with the brightness, contrast, and hues on some images, as well as random crops (up to 16 pixels in the borders). Here are some examples from my dataset. ![]() Any human being who's played the games can instantly recognize these postage stamps screenshots, so if I can't train an AI to do it I should just give up right here. ![]() So I started with a basic tutorial and spent a few hours debugging issues because of my different augmentation method, but I finally got it to start and I'm training. The process is looking pretty good so far, the validation accuracy is trending up with each epoch! Check out the training progress here. I'm not done yet, but I'm posting this so that I have to answer for myself even if I fail. So wait tight for the results! ____________________
In our brief lives, • we've managed to meet. • Treasure this gift, • this precious time that we have.
|
I have no idea how to read your training output, but what a fun idea for a project. It's silly an unnecessary but I love it!
Good luck with your class!
____________________
Good luck with your class!
____________________


























click to see larger
click me



