CartoonGan-tensorflow

mnicnc404/CartoonGan-tensorflow is ranked 12,524 among these 8 peers; eaplatanios/tensorflow_scala is ranked 12,521.

This repo is no longer being maintained. To reproduce the result, please 100% follow the instructions including environment settings and training commands. CartoonGAN-TensorFlow2 Generate your own cartoon-style images with CartoonGAN (CVPR 2018), powered by TensorFlow 2. This repo demonstrates how to: Train your own CartoonGAN Generate anime using trained CartoonGAN Train your own CartoonGAN In this section, we will explain how to train a CartoonGAN using the script we provide. Setup Environment First clone this repo: To run

code in this repo properly, you will need: Python 3. 0 Alpha tqdm imageio tb-nightly For environment management, we recommend Conda. You can get by installing Anaconda or Miniconda. You can install all the packages by running the following commands: If Anaconda is not available, you can also run: You will also need TensorFlow version of keras-contrib for some custom Keras layers used in our CartoonGAN implementation: If all above complete successfully, you're good to go. Prepare Dataset

You also need to prepare your own dataset and arrange the images under folder as below: and folders contain real-world images, while contain images with desired cartoon style. Notice that 8 images in folder will be evaluated after each epoch, so they should not appear in . In order to generate , you can run : smooth. py credit to taki0112 Start training Although you may have to tune hyperparameters to generate best result for your own datasets, train following settings that

we found effective can be your starting point. If you get more than 16GB memory in your GPU, you can try these settings (Note that indicates that we are training GAN with a light-weight generator): Note that is for (source). If your GPU does not have 16GB memory, you can use a smaller and use lower learning rates accordingly. For example, for , you can try: Detailed log messages, model architecture and progress bar are all provided. This enable you

to gain a better understanding of what is happening when training a CartoonGAN. Choose model architecture Notice that we specified in our previous example: When specified, train. py will initialize a light-weight generator for training a CartoonGAN. When we design the light-weight generator, ShuffleNet V2 is taken as our reference. This generator is designed to minimalize inference time while achieving similar effect. We will make some minor adjustments to discriminator as well when is specified. Generator proposed by

the original CartoonGAN authors To train a CartoonGAN with the original generator/discriminator architecture proposed by the CartoonGAN authors, simply remove option: Monitor your training progress In our repo, TensorBoard is integrated perfectly so you can monitor model's performance easily by: After training for awhile, you should be able to see something like this: In addition to metrics and loss functions, it is good practice to keep an eye on the images generated by GAN during training as well. Using

our script, monitoring generated images on TensorBoard is a no-brainer: For further details about training, we recommend reading train.