classify.py stuck in defining input_scale and raw_scale

[Saman Sarraf]

Dear caffe users,

I'm kind of got stuck in defining classify.py parameters and I am getting arbitrary results / predictions from this interface. 

I trained LeNet (exactly the same network of MNIST data) for 57 classes (instead of 10) and got Test net output #0: accuracy = 0.928275. I assume this is the accuracy of testing samples. I am using classify.py against the same testing samples and I am getting very arbitrary predictions. 

I attached my solver and network model. The only thing I need to mention is the scaling factor "scale: 0.00390625" in the model. So my question is how to set raw_scale and input_scale ? Also image_size is cropping images or rescaling images? because my testing samples are 32*32 but I trained LeNet bu 28*28. and the last thing , as I am not subtracting the mean of images during training, I don't need to subtract it in prediction , correct?

Honestly, I have tried everything mentioned above and still get the bizarre predictions, just wanted to double check if they are correct or not ? there should be definitely something wrong that I don't get the correct predictions.   

Your prompt help is much appreciated,

Saman

[Saman Sarraf]

Hey folks, 

Thanks for your help. 

The problem was in doing normalization twice first in deploy.prototext and second in classsify.py command line.

Best.

Saman