Images Caffe ImageNet 32X32

So my problem is that I cannot train the image with smaller images (32X32), when I resize them to 256X256 everything starts well. So I know the problem is with my settings.

I tried to set my own settings for:

deploy.prototxt: I set the last 2 input_data to 32

solver.prototxt: i set solver_mode: CPU (only everything else left)

train_val.prototxt: I set crop_size: 31 for both settings

all paths are correct as the training works fine with images resized to 256X256

but with the i installation described above, I get this error:

I1114 11:53:38.484948 4566 net.cpp:96] Setting up fc6
F1114 11:53:38.484967 4566 blob.cpp:72] Check failed: data_
*** Check failure stack trace: ***
@ 0x7f82ed8a0daa (unknown)
@ 0x7f82ed8a0ce4 (unknown)
@ 0x7f82ed8a06e6 (unknown)
@ 0x7f82ed8a3687 (unknown)
@ 0x45814e caffe::Blob<>::mutable_cpu_data()
@ 0x4c12fa caffe::GaussianFiller<>::Fill()
@ 0x4e5719 caffe::InnerProductLayer<>::LayerSetUp()
@ 0x4705e1 caffe::Net<>::Init()
@ 0x471eee caffe::Net<>::Net()
@ 0x452560 caffe::Solver<>::InitTrainNet()
@ 0x4537c6 caffe::Solver<>::Init()
@ 0x453926 caffe::Solver<>::Solver()
@ 0x415650 caffe::GetSolver<>()
@ 0x411b44 train()
@ 0x40ef21 main
@ 0x7f82ea913ec5 (unknown)
@ 0x410697 (unknown)
@ (nil) (unknown)
train_caffenet_letter.sh: line 4: 4566 Aborted (core dumped) ./build/tools/caffe train --solver=models/bvlc_reference_caffenet/solver_letter.prototxt

      

        
        
        
      

    

If anyone can explain what other settings should I change in these files, or am I doing it completely wrong?

Ive tried to follow this tutorial but it didn't do it: http://drubiano.github.io/2014/06/18/caffe-custom-data.html