why label sholud be [0 255]

[zack6514]

I have 4 classes and assign 0 0 1 1 to it ， and in train.prototxt label data not have the param scale , I think this is ok? why should be [0 255]?

[sukritshankar]

We are using data layer for labels. Data layer in Caffe automatically scales down by a factor of 255. Thats why labels should be in the scale of [0, 255], so that after scaling, you should have proper values between 0 and 1 !!

[vishalghor]

Hi @sukritshankar ,

Can you give more information on creating .mat file labels.
How to create the .mat file?
It would be helpful if you could help me for the same as i am new to caffe framework.

Thanks and Regards

[aniket03]

@sukritshankar I believe its the code in prototxt

Caffe-LMDBCreation-MultiLabel/train_vgg_11_sigmoid_cross_entropy_loss.prototxt

Line 43 in f4a92fc

scale: 0.00390625

due to which caffe will scale down the labels by 255. Thus if we eliminate this line of code should labels of the form [0, 1] not be directly useful. Without having to scale them up by 255.

[aniket03]

Hi, @sukritshankar any updated on above?

[iiwindii]

@aniket03 Agree! If your labels are 0 ,1, then it should work to eliminate this line.