Distinguish between Super Saiyan 1,2,3,4 and god using this deep learning model!
I used Google's Inception (2015) model. Inception is a deep convolutional neural network built for classifying real world images of thousand categories.
Input to the model: A picture file.
Output of the model: Probability distribution of each saiyan forms.
I retrained it by replacing the Inception's last layer with my Super Saiyan layer. Let's see whether or not features in real world are useful for classifying anime forms!
This model is about 83 MB big in file size. You can roughly interpret it as the brain of the classifier.
I trained it with 388 images of Super Saiyan Goku, Vegeta and their sons. That's only a few images, so don't expect good accuracy from this classifier.
You can take a look at the training images in the super_saiyan_forms folder.
You can use some images inside test_images to test this model's prediction. Or go find some images yourself.
Clone and run bash on this repository. (Ensure that you have TensorFlow installed.)
Then run python label_saiyan.py test_images/vegeta_saiyan4.jpg
Wait for the model's prediction.
Here's the result you should expect if things work correctly:
super saiyan 4 (score = 0.81662)
super saiyan 1 (score = 0.06831)
super saiyan 3 (score = 0.05949)
super saiyan god (score = 0.04132)
super saiyan 2 (score = 0.01425)Now you know that this Vegeta is surely a Super Saiyan 4 because the model predicted with 81% confidence!
Here's another prediction with our beloved Goku. I don't even know how to classify this myself because I haven't watched Dragonball Super yet!
Goku's Result:
super saiyan god (score = 0.67520)
super saiyan 3 (score = 0.18979)
super saiyan 2 (score = 0.06863)
super saiyan 4 (score = 0.04908)
super saiyan 1 (score = 0.01730)Not so surprising right? It's a Super Saiyan God Goku!
But The model isn't sure (like me). It says 67% confidence to be safe.
Try changing the image and see the new result yourself.
Super Saiyan 4 is easy to distinguish but Super Saiyan 1 and 2 is somewhat troublesome!
Also try to fool the model ! Learn something from it.