Using DCGAN model generate deer picture in cifar10

將訓練資料透過捲積神經網路經過訓練產生鹿的圖像
相較於傳統的深度學習網路，CNN把全連接層都換成了卷積層，並透過跨步卷積的方式
來取得局部特徵及縮減圖片大小，達到大幅度提高訓練速度的效果

Environment

• Jupyter Notebook
• Python 3.6
• Tensorflow 1.9-GPU

Archeitecture

Note:此架構之shape-size僅為示意圖，與本project不同

DCGAN's advantages compare to GAN

• In GAN model, both generator and discriminator use Fully Connected Layer, but DCGAN use Convolution Neural Network archeitecture.
• Both Genrator and Discriminator use Batch Normalization for speeding up calculation.
• Both Genrator and Discriminator different from traditional CNN ,they don't use pooling layer.
• Generator use tansposed convolution layer ,as known as deconvolution ,and Discrinator use convolution layer.

Note:Generator's final layer use tanh,and discriminator's is sigmoid.

Result

加載說明

若透過網頁載入此ipynb文件失敗，可直接透過此連結查閱

Reference:https://arxiv.org/abs/1511.06434