Generate cartoon faces with variational auto-encoders. Built with TensorFlow 2.0-beta1, trained 4 epochs with GTX1060 3GB and took 2 hours. Also in this software, you shall see a great use of TensorBoard. Here are results generated with random bottlenecks;
You can download pre-trained models from the directory named models
You can download the dataset that i used from here
Bottom, you can see files, classes in files, functions in classes and what those functions do. My class names are from TV Show named 'How I Met Your Mother', it is like this because i kinda like use the names of show characters on my code :)
- load_image:
loads image, no label - read_all_data:
gets all the paths and labels, saves it as npy
Summary: Create tensorflow dataset object with map
- encoder_model:
create encoder model - decoder_model:
create decoder model - encode:
gets bottleneck and split into two(mean, logvar) - decode:
gets reparameterized input and put it into decoder - reparameterize:
gets mean and logvar and reparameterize them - generate_sample:
creates random input for decoder and generate image by through decoder - log_normal_pdf:
sub loss function - compute_loss:
do everything and calculate loss - save:
save encoder and decoder - train_step:
combine those things, and optimize model by through self.optimizer
Summary: Create model structure, loss, optimizer and simple train step
- save_images_to_tensorboard:
gets regenereted and real images, generate some from random and save to tensorboard - train_model:
train model by through barney and marshall, use 'save_images_to_tensorboard' to test it
Summary: Train model
[WARNING] This project still in developing faze which means that i will add more algorithms, data structures and stuff like that.
Please ask if you have questions :)