ALLGANs

머신러닝 프레임워크를 활용한 비교사(Unsupervised) 학습 모델 구현 프로젝트

Unsupervised Learning Model Implementation Project Using Machine Learning Framework

Authors

• 김유준(demetoir, Yu Jun Kim)
  • e-mail: wnsqlehlswk@naver.com
• 배경민(WKBae, William Kyungmin Bae)
  • e-mail: williambae1@gmail.com
• 박성규(StarG, Soung Gyu Park)
  • e-mail: psk7142@naver.com

Index

• ALLGANs
  • Training dataset
  • Results
    • GAN
    • LSGAN
    • WGAN
  • Observation
  • Getting Started
  • Folder structure
  • Dependency
  • Authors
  • Reference

Training Dataset

• MNIST
• Fashion-MNIST
• CIFAR 10
• LLD(Large Logo Dataset)

Results

• GAN, LSGAN, WGAN are have same generator and discriminator CNN structure.
• No data augmentation.

GAN

dataset	MNIST(20 epoch)	fashion-mnist(20 epoch)
result image
loss D
loss D_real
loss D_gen
loss G

dataset	CIFAR10(20 epoch)	LLD(20 epoch)
result image
loss D
loss D_real
loss D_gen
loss G

LSGAN

dataset	MNIST(20 epoch)	fashion-mnist(20 epoch)
result image
loss D
loss D_real
loss D_gen
loss G

dataset	CIFAR10(20 epoch)	LLD(20 epoch)
result image
loss D
loss D_real
loss D_gen
loss G

WGAN

dataset	MNIST(20 epoch)	fashion-mnist(20 epoch)
result image
loss D
loss D_real
loss D_gen
loss G

dataset	CIFAR10(20 epoch)	LLD(20 epoch)
result image
loss D
loss D_real
loss D_gen
loss G

WGAN result - after training LLD dataset for 5 epoch

Observation

1. MNIST dataset is easy for GAN, LSGAN, WGAN.
2. Training with Fashion-MNIST dataset, GAN is slightly better others.
3. Training with CIFAR 10 dataset, WGAN are better than others. but clearly does not generate looks pretty image like original image.
4. Training with LLD dataset, three GANs generate similar image. Compare with original image, three GANs generate ugly image.

Compare other GANs, WGAN generate best image in less epoch, but after 5 epoch generate worse than others.

Above loss of generator and discriminator, generator overpowered discriminator.

GANs trained outline of all original dataset(fashion-mnist, CIFAR10, LLD), but did not train detail.

Getting Started

1. python ./setup.py install and follow console (If you already installed dependencies, follow step 2.)
2. Run bench code python ./main.py
3. If you want to run another instance, modify workbench/bench_code.py.

Folder structure

├─data          # default dataset
├─data_handler
├─dict_keys
├─instance      # default model instance
├─model
├─unit_test
├─util
├─visualizer
└─workbench     # bench code, DatasetHelper

Dependency

• Python 3.5+
  • matplotlib (2.1.2)
  • numpy (1.14.0)
  • pandas (0.22.0)
  • Pillow (5.0.0)
  • scikit-image (0.13.1)
  • scikit-learn (0.19.1)
  • scipy (1.0.0)
  • tensorflow (1.4.1)
  • tensorflow-gpu (1.4.1)
  • tensorflow-tensorboard (0.4.0)
  • opencv-python (3.4.0.12)
  • requests (2.18.4)

Reference

• Generative Adversarial Networks
• Least Squares Generative Adversarial Networks
• Wasserstein GAN
• Improved Training of Wasserstein GANs
• Going Deeper with Convolutions
• Rethinking the Inception Architecture for Computer Vision
• Inception-v4, Inception-ResNet and the Impact of Residual Connections on Learning