Cannot train lenet in matcaffe.

[deartonym]

Update2: 05/14/2016

I found that the Caffe version of last year does not have this problem. Thus this bug might be a new one brought with the newest Caffe.

Right now I am using the matcaffe version of July, 2015.

To people who are experiencing the same problem: Please use an older matcaffe and have a try.

Best,

%-------------------------------------------------------------------------------------------------

Update:

I found the problem might be caused by failure to load lmdb data file. However, the lenet lmdb files are prepared.

%------------------------------------------------------------------------------------------------- As the title states, I meet with a very strange situation:

1. My compiling succeeded.
2. I can run all .sh examples, and matcaffe was successfully built, and it passed the mattest.
3. I can run the classification_demo.m.
4. I can load any .caffemodel in to matlab, and I can do forwarding.
5. I can load solver.prototxt using caffe.Solver(), e.g., solver = caffe.Solver('./examples/mnist/lenet_solver.prototxt'); and I can see two Caffe network in the solver.

However, I just cannot train this solver. That is, when I run: solver.step(1000); or solver.solve(); Matlab just leads the program to Solver.m, and gets stuck at: caffe_('solver_solve', self.hSolver_self);. Btw, there is no error and warning during this process.

I'm not sure whether any of you met with this problem before. It seems a very weird one.

Thanks for your time.

Best, Shu

[encore0117]

Same problem here. Any help will be appreciated!