Yijunmaverick
Train the model only with the reconstruction loss
Hi, I train the model only with the reconstruction loss, than using the caffemodel to do the next train. However, I cannot obtain the expect result. So, dose it means that I shouldn't consider the local D, global D and face parsing when I train only with reconstruction loss. And how many iterations do you train in celebA?
@swordcheng Thanks for your interests.
You can take a look at Section 3.5 of the paper where I present the three-stage training strategy. There might be some failures as the GAN training is originally unstable. I just monitor the loss and stop the training when converged.
I have read the Section 3.5 of the paper. But, in your open source code, only two-stage were mentioned. In readme you say:
We provide an initial model that is only trained with the reconstruction loss, as a good start point for the subsequent GAN training. Please download it and put it under ./matlab/FaceCompletion_training/model/ folder.
Run ./matlab/FaceCompletion_training/demo_GFC_training.m for training.
the demo_GFC_training.m contains local_D loss, global_D loss and face parsing loss. It looks like combine stage2 and stage3.
You can adjust the loss weight (0 or not) to separate the training into different stages.
@swordcheng 您好 我在用matlab运行demo training的时候出现了很多问题，看你好像成功运行成功了，你能够提供一个联系方式我好好请教一下你如何配置的事吗
@swordcheng 你好,我在运行中遇到了点错误,就行MATLAB 会崩溃,能加你的联系方式吗?我的微信xcj809078075,QQ:809078075.谢了
I have diff GFC_Gen_training_G.m GFC_Gen_training_D.m ,they are similar，in fact ,this is the different
so what diffenent between them? Will it affect training?
Can someone explain it? Thank you !
@swordcheng 楼主你好！！求联系方式，我的QQ是470823627！！我在跑测试代码时matlab报错说prototxt文件错误，但我的matcaffe在其他模型上是没有问题的，哭唧唧T T
