search

PaddlePaddle / book

Deep Learning 101 with PaddlePaddle （『飞桨』深度学习框架入门教程）
http://www.paddlepaddle.org/documentation/docs/zh/1.2/beginners_guide/quick_start/index.html
2.74k stars 1.32k forks source link

Network structure not consistent with proposed DCGAN #851

Closed a411919924 closed 4 years ago

a411919924 commented 4 years ago

The original DCGAN proposed to replace all fully connected layers with convolutional layers. But in the DCGAN implementation in PaddlePaddle, the developer used a fully connected layer as the last layer in both the generator and discriminator. This may raise an inconsistent result.

ceci3 commented 4 years ago

Thanks for your attention. It is inconsistent with origin paper. DCGAN in PaddlePaddle developed based on https://github.com/carpedm20/DCGAN-tensorflow, it uses linear in the network, and the result seems not bad than origin paper, so I guess fully-connected layer or convolution layer is not that important for the last layer.

a411919924 commented 4 years ago

Thanks for the quick response and good explanation.