None convergence with SoftmaxLoss when num_output becoming like 800k,

[peyer]

I have comment a problem while training a model for face recognition that when increasing num_output to 800k, the prob of each category became 1.0/800k and never channged! I have tried the same network and optimization parameters with BVLC-CAFFE, it successfully became to convergence after 8k iterations. How should I do ?