search

vlfeat / matconvnet

MatConvNet: CNNs for MATLAB
Other
1.4k stars 753 forks source link

Can we use an image data without converting to idx3-ubyte format in cnn_train? #720

Closed dhruvmullick closed 7 years ago

dhruvmullick commented 8 years ago

Will the results change if we don't use idx-ubyte format?

albanie commented 7 years ago

Hi @dhruvmullick, sorry for the slow response. All vl_nn* matconvnet operations work with the single data type (and from January 2016 onwards, also with thedouble data type). If you are referring to the minst example, the downloaded data is converted to the single data format before it is used (you can see this conversion here).

dhruvmullick commented 7 years ago

Hi @albanie, actually I was referring to this code. The labels are in the format of idx-ubyte. I don't understand what to do in case I don't have the lables in that format. Can you please help me out?

albanie commented 7 years ago

Ah, that's actually just the name of the file (it's a little confusing). In the next line you can see that the actual labels are read as uint8. You can have your labels as uint8, but other formats will also work - it won't affect the loss layers.

dhruvmullick commented 7 years ago

Oh. Makes sense. I was wondering if ubyte was used to refer to uint8. Thank you!