Zhedong Zheng

你好 Mac上面你有配置cuda么? 我在ubuntu14.04上配合titanX 跑的。 不过应该也可以cpu跑的。 你直接用 vl_compilenn来编译。 你要尝试把model都import到cpu 模式（数据类型从gpuArray 改到 single）, 你可以参考的函数 gather() Zhedong Zheng University of Technology Sydney https://zdzheng.xyz > On 14 Oct 2017, at 8:10 pm, NinaLiJiachen wrote:...

Have you `mex zzd.c`?

You can use the public model produced by Oxford VGG although it is not the best model on LFW. http://www.vlfeat.org/matconvnet/pretrained/#face-recognition But I still think it is a good public model...

I did not use VGG16. I just deepen the 48 network. For more detail, you can refer to my thesis https://github.com/layumi/Fudan_BS_PaperCopy/blob/master/%E6%AF%95%E4%B8%9A%E8%AE%BA%E6%96%87_123-7130083_%E9%83%91%E5%93%B2%E4%B8%9C_%E5%9F%BA%E4%BA%8E%E5%8D%B7%E7%A7%AF%E7%A5%9E%E7%BB%8F%E7%BD%91%E7%BB%9C%E7%9A%84%E4%BA%BA%E8%84%B8%E6%A3%80%E6%B5%8B.pdf

@Seanlinx 1.Because there is a stride 2 function in the net, it will skip some image and get a 1/4 small heatmap. So I make up this loss by running...

@caiqi It's easy to solve. The matconvnet use two different name. weights{1} weights{2} and filters bias You may set a break point and see what to replace them

@JHUKUAN I think you can try to change 'parfor' to 'for'. 'parfor' is a multi-thread function for matlab. OR I have test my code on the Ubuntu14.04 as well as...

@gogolgrind 1. The net which is deeper has been proved to a better performance(Like res-net). And dropout can help net to be more robust and perform well on validation. 2....

@gogolgrind 1.Actually I wrote softmax layer in my test code. You can see something like "T=exp(X-max(X));Y=X/T;" in my test code. 2.You can refer to the model file and find its...

@windpls 12net train-error about 5% 24net train-error about 2% 48net train-error about 1% 12netc train-error about 50% 24netc train-error about 30% 48netc train-error about 20% I am sorry. I didn't...