While I studied Generative Adversial Network models, I wondered if I can make a GAN model which draws a animal which resembles me. It's just pure curiosity. And I'd like to create this. I built a paired human face and animal face dataset with searching on Google. Due to the characteristics of GAN, it is hard to train on personal computer. My computer's GPU RAM is just 4GB and I used Google Colab for training.
input | output |
---|---|
CycleGAN is one of the most popular methods for image-to-image translation. I make two pairs of Generator and Discriminator. I trained this model on Google Colab for 500 epochs for an 2 hours with premium GPU.
Test | |
Human2Animal | |
Animal2Human |
The result was great for dataset input. But as you can see in the Test GIF, this model creates like a input not an animal face when new input comes in..