I am using Cyclegan TensorFlow with python to transform from human to cartoon. I am using CelebA dataset and Cartoonset10k. I used the implementation of cyclegan example by TensorFlow https://www.tensorflow.org/tutorials/generative/cyclegan. they are using UNET generator. I cropped both datasets to get the exact frontal face and trained for 100 epochs but I faced a model collapse when no matter what input to the generator it produces the same output. Next, I implemented my own generator with RESNET as describe in paper unpaired image to image translation but same discriminator as in example, trained for 100 epochs with 10000 examples from Human and 5000 examples from cartoonset. when tested with test human examples, the generator produced these outputs in the pictures. I don't know if that is also a model collapse or something else is wrong. I also Attached the losses graphs.
