ML_LEE_2022_hw6
Task introduction
- GAN:project some random variables into specific space
- 必須實作:DCGAN、WGAN、WGAN-GP
- 產生1000張動漫人臉
- <number>.jpg
Dataset
- 從Crypko爬下來的
- 71,314張pictures
faces
|
|_0.jpg
|_1.jpg
|
…
Evaluation metrics
- FID score
- AFD rate,表示多少張圖片能被辯認為動漫頭像
Baseline
Simple:SAMPLE CODE(DCGAN) , train 1 hr
Medium:DCGAN with more epochs , train 11.5 hr3 hr
Strong:WGAN / WGAN-GP , train 2
Boss:StyleGAN , train < 5 hr
Report Question
- 列出WGAN跟GAN的兩大差異
- 畫出WGAN跟WGAN-GP的gradient norm結果