curioustauseef / deep-learning-with-cats

Deep learning with cats (^._.^)

Deep-learning-with-cats

This repository is a "toy" project so I can gain experience building deep neural networks. My first goal is learning to generate pictures of cats with Generative Adversarial Networks (^._.^).

Updates

• 17/07/2017 : Added CycleGAN for style transfer, so far doesn't work well with the CAT dataset and it's so slow (20-24h to fully train it) that optimizing it is very difficult. I'm leaving it here for those interested but I probably will try fast neural style instead.
• 15/07/2017 : Many people were concerned about whether the generated images are really novel or rather just copies of the training dataset so I made a python function that output the 5 most similar training (real) images to the given generated (fake) images. See: https://github.com/AlexiaJM/Generative-model-most-similar-images.

Objectives (so far)

• Generate images of cats using various types of Generative Adversarial Networks (GAN)
  • use DCGAN (Done)
  • use WGAN (Done)
  • use WGAN-GP (Done)
  • use LSGAN (Done)
  • use BEGAN
• Transform real cats into art pieces
  • use CycleGAN (Code done)
  • use Fast neural style
• Various/Others
  • Preprocess cat images so we get aligned cat faces for much better GAN convergence (Done)
  • Separate cats by size to be better able to generate cats of certain sizes (Done)
  • Fix DCGAN models so that they can adapt to different input image sizes (Done)
  • Keeping log for TensorBoard (Done)
  • Automatic folder setup (Done)
  • Add multi-gpu and non-CUDA option (Done)
  • Option to load previous models (Done)
  • Add log to output (Done)
  • L2 Regularization (Done)
  • Try making higher resolutions pictures
    • 128 x 128 (Done)
    • 256 x 256 (Need more higher res data to do this)
  • Try Self-Normalizing Neural Networks (SELU) as per https://arxiv.org/abs/1706.02515 (Done)
  • Add option to generate more cat pictures (Done)
  • Add option to see the 5 nearest real cats images and thus be able to tell if the generated cats are really unique (Done)
  • Dropouts in G
  • Try adding Frechet Inception Distance (FID) as per https://arxiv.org/pdf/1706.08500.pdf
  • Try soft and noisy labels as per https://github.com/soumith/ganhacks
  • Try adding decaying noise to input as per https://github.com/soumith/ganhacks

Needed

• Python 3.6, PyTorch, Tensorflow (for TensorBoard)
• Cat Dataset (https://web.archive.org/web/20150703060412/http://137.189.35.203/WebUI/CatDatabase/catData.html)
• TensorBoard logger (https://github.com/TeamHG-Memex/tensorboard_logger)

To run

$ # Download dataset and preprocess cat pictures 
$ # Create two folders, one for cats bigger than 64x64 and one for cats bigger than 128x128
$ sh setting_up_script.sh
$ # Move to your favorite place
$ mv cats_bigger_than_64x64 "your_input_folder_64x64"
$ mv cats_bigger_than_128x128 "your_input_folder_128x128"
$ # Generate 64x64 cats using DCGAN
$ python Meow_DCGAN.py --input_folder "your_input_folder_64x64" --output_folder "your_output_folder"
$ # Generate 128x128 cats using DCGAN
$ python Meow_DCGAN.py --input_folder="your_input_folder_128x128" --image_size 128 --G_h_size 64 --D_h_size 64 --SELU True
$ # Generate 64x64 cats using WGAN
$ python Meow_WGAN.py --input_folder "your_input_folder_64x64" --output_folder "your_output_folder"
$ # Generate 64x64 cats using WGAN-GP
$ python Meow_WGAN-GP.py --input_folder "your_input_folder_64x64" --output_folder "your_output_folder" --SELU True
$ # Generate 64x64 cats using LSGAN (Least Squares GAN)
$ python Meow_LSGAN.py --input_folder "your_input_folder_64x64" --output_folder "your_output_folder"

To see TensorBoard plots of the losses

$ tensorboard --logdir "your_input_folder"

Results

Discussion of the results at https://ajolicoeur.wordpress.com/cats.

DCGAN 64x64

DCGAN 128x128 with SELU

WGAN 64x64

WGAN-GP 64x64 with SELU