问题
I was wondering if it's possible to predict a 1-by-n feature associated to an input image using caffe. In this post there is a solution to make caffe predict a binary vector such as [1 0 1 0].
Is this solution also suitable if I have a 1-by-n float vector as label (such as [0.2, 0.1, 0.3, 0.4] ? I want to predict such a vector, not a binary vector label.
回答1:
You can also think about this MultiTaskData Layer. It can parse float typed label vector as you mentioned in your question.
回答2:
Yes; I believe that if you simply follow the instructions in the post you referenced, but use float labels in place of the binary ones, that you'll get useful results. Caffe is happy to support you. :-)
来源：https://stackoverflow.com/questions/38095280/how-to-predict-float-vector-labels-with-caffe