Some questions about multi label image classification in MXNet.

[wangg12]

Hi,
I want to perform multi label image classification for my own dataset with mxnet. Here are some questions bother me.

• How to prepare the recordIO database or the index file?
  According to Extension: Mutliple Labels for a Single Image , the index file should have a fixed width of label field. What if my labels are not always the same width which is the common occasion. And will the labels of a single image be treated as a vector like (1,0,0,...,1,...0,...,1)? Or how can I prepare a binary vector label for the image data?
• What loss function should I assign to the training phase? Will the SoftmaxOutput deal with multi-label? Or should I implement a sum of squares like loss function?
• What evaluation criteria should I choose for the test or val phase? The accuracy metric is for single label.

Thanks!

[piiswrong]

you can refer to this example #940

[yuconglin]

@wangg12 did you figure out? Thanks a lot in advance.