ciffar conv_net example does not learn through data although a flat fullyconnected model does.

[Erogol]

I have a binary classification problem with 60.000 pos and 60.000 neg examples of 40x20 images. I tried two different models. One is simple one hidden layer fully connected model and the other is a CIFAR model provided in the examples. Fully connected model can reach 96% at the end but CIFAR model does not improve and its loss values is just waving around a constant value. I tried to give learning rate 0 to see something wrong about the data but loss value stays same expect some changes around ~0.00001. I also tried very small learning rate but the result is same and loss value does not decrease. I believe if there would be something wrong about the data fully connected model would also not toe converge. What would you suggest to investigate the problem more? 

I give the prototxt of my cifar model in this link for the sanity. https://gist.github.com/78fe47aaa31cd9c72f10.git

[Erogol]

I solved this by carrying HDF5 data to levedb format. I am not sure about the reason but levedb version is just fine

[echoaimaomao]

yes,I had the same problem with you. The hdf5 data type does not converge but the leveldb does. I could not figure it out neither.

在 2014年12月25日星期四UTC+8上午2时41分10秒，Erogol写道：