Leveldb data format about caffe HOT 3 CLOSED

I have the same question. Since I am not so familiar with the leveldb format, I simply try to look into the code in e.g., convert_cifar_data.cpp and figure out how it works. To convert .mat file to work with caffe, I first convert it to binary format (use fwrite in matlab) as the cifar-10 dataset did according to this description (http://www.cs.toronto.edu/~kriz/cifar.html). After that, I use the modified version of convert_cifar_data.cpp to convert it to leveldb format. This approach works well when my .mat file is some image data, i.e., they are stored in uint8. However, I have trouble when it comes to int16 or float format. (It seems that caffe assumes uint8 data.)
It would be helpful if there is another example like convert_cifar_data.cpp but deal with data in other format, like float.

from caffe.

Caffe accepts single/float image data. If you are testing, the matlab wrapper should be suitable for your needs as long as you pass images in the expected dimension order. For efficiency or for parceling up data for training, conversion to leveldb via the convert.py utility in caffe/python/convert.py might be simplest.

It could be helpful to take a look at the matlab wrapper and the python convert examples as well as the io utility code (although you could have to do an intermediate step of exporting your .mat to images).

from caffe.

Thanx ChenlongChen

I just follow what you say, and now can transform cifar10 .mat format to binary format, then just follow the cifar10 example and get the correct training accuracy.
Matlab code

for i=1:5
name_load= sprintf('data_batch_%d',i);
name_bin = sprintf('data_batch_%d.bin',i);
load(name_load);
dlabels = [labels data]';
fileID = fopen(name_bin,'w');
fwrite(fileID,dlabels,'uint8');
fclose(fileID);
end

from caffe.