Project 5 Generative Adversarial Network Zhifei Zhang Generative

  • Slides: 14
Download presentation
Project 5: Generative Adversarial Network Zhifei Zhang

Project 5: Generative Adversarial Network Zhifei Zhang

Generative Adversarial Network (GAN) https: //github. com/hindupuravinash/the-gan-zoo

Generative Adversarial Network (GAN) https: //github. com/hindupuravinash/the-gan-zoo

Generative Adversarial Network (GAN)

Generative Adversarial Network (GAN)

Generative Adversarial Network (GAN) New components: Transposed convolution, Batch Normalization Binary Classifier: Conv, Leaky

Generative Adversarial Network (GAN) New components: Transposed convolution, Batch Normalization Binary Classifier: Conv, Leaky Re. LU, FC, Sigmoid https: //github. com/Pramod. Shenoy/GANerations

Transposed Convolution (Deconvolution) Output Stride=1 Pad=Valid Stride=2 Pad=Same Conv Input Output Deconv Input https:

Transposed Convolution (Deconvolution) Output Stride=1 Pad=Valid Stride=2 Pad=Same Conv Input Output Deconv Input https: //github. com/vdumoulin/conv_arithmetic

Re. LU vs. Leaky Re. LU (LRe. LU) Used in Generator Used in Discriminator

Re. LU vs. Leaky Re. LU (LRe. LU) Used in Generator Used in Discriminator

Batch Normalization (BN) Saturation area Zero gradient https: //towardsdatascience. com/batch-normalization-in-neural-networks-

Batch Normalization (BN) Saturation area Zero gradient https: //towardsdatascience. com/batch-normalization-in-neural-networks-

Generative Adversarial Network (GAN) FC, BN, Re. LU Reshape Deconv, tanh/sigmoid Conv (kernel 5

Generative Adversarial Network (GAN) FC, BN, Re. LU Reshape Deconv, tanh/sigmoid Conv (kernel 5 x 5, stride 2) BN, LRe. LU (slope 0. 2) Reshape, FC, sigmoid Deconv (kernel 5 x 5, stride 2) BN, Re. LU https: //github. com/Pramod. Shenoy/GANerations

GAN on MNIST FC, BN, Reshape 100 Deconv BN, Re. LU 7 x 7

GAN on MNIST FC, BN, Reshape 100 Deconv BN, Re. LU 7 x 7 x 16 Deconv Tanh/Sigmoid 14 x 8 Generator Conv, BN, LRe. LU 28 x 1 Conv, BN, LRe. LU 14 x 8 7 x 7 x 16 Discriminator Reshape, FC, BN, LRe. LU 256 FC, Sigmoid 1

Conditional GAN on MNIST One-hot label FC, BN, Reshape 100 Deconv BN, Re. LU

Conditional GAN on MNIST One-hot label FC, BN, Reshape 100 Deconv BN, Re. LU 7 x 7 x 16 Deconv Tanh/Sigmoid 14 x 8 Generator Conv, BN, Re. LU 28 x 1 Conv, BN, Re. LU 14 x 8 7 x 7 x 16 Discriminator Reshape, FC, BN, Re. LU 256 FC, Sigmoid 1

Tensor. Flow Implementation of GAN Run the demo code: (Linux or Mac. OS) $

Tensor. Flow Implementation of GAN Run the demo code: (Linux or Mac. OS) $ git clone https: //github. com/carpedm 20/DCGAN-tensorflow $ cd DCGAN-tensorflow $ pip install tqdm (if you do not have this package) $ python download. py mnist $ python main. py --dataset mnist --input_height=28 --output_height=28 --train

Tensor. Flow Implementation of GAN • Input: z, image, (label) Training: • Network: D,

Tensor. Flow Implementation of GAN • Input: z, image, (label) Training: • Network: D, G for epoch • Loss: D, G • Optimizer: D, G for batch Update D Update G

https: //devblogs. nvidia. com/photo-editing-generative-adversarial-networks-1/

https: //devblogs. nvidia. com/photo-editing-generative-adversarial-networks-1/

Tensor. Flow Implementation of GAN Testing: • Random Generation • Conditional Generation • Interpolation

Tensor. Flow Implementation of GAN Testing: • Random Generation • Conditional Generation • Interpolation