MultiClass label support #3326
MultiClass label support #3326
Conversation
5a2ae3e to
036350a
Compare
squashed commits
036350a to
1172164
Compare
|
@mtamburrano I also found a definition of multiclass, multilabel and multi-task classification here. According to this reference, softmax is inherently a multiclass classifier. I think your modification is more related to multi-task classification, which can be implemented by setting Basically, multilabel classification can also be done in this way, but one needs to duplicate the |
This is an uncomplete PR to support multiclass loss in caffe.
I doubt this will be ever merged but I wanted to share some modifications I made to memoryDataLayer to accept multiple labels and to softmax to handle mutliclass problems.
A good discussion about multilabel, multiclass and multitask problems in caffe is here
The main news this PR introduce are:
MemoryDataLayercalledaddMatVectorMultilabelthat accepts avector < vector < int > >as label for each data Mat. So a Mat can be easily associated with multiple classes. The resulting label blob will have the label values on the channel dimension (or more exactly in the dimension1since the N-dimension support).SoftmaxLossLayerto let it to compute the loss on different labels and on slices ofbottom[0]. The number of slices are specified with the prototxt parametersliceand must matchbottom[1]->shape(softmax_axis_)SoftmaxLayerto let it to compute the softmax independently on slices ofbottom[0], so theSoftmaxLossLayercan compute the loss for each slice comparing it with each corresponding label.The idea is to feed the network with
MemoryDataLayerandaddMatVectorMultilabel, then attaching aSoftmaxLossLayerspecifying the parametersliceto be equal to the size of the vector of the labels added. It is important to set the output ofbottom[0]of theSoftmaxLossLayerto be a multiple of the number of classes and of the slices.For example let's say we want to add a vector of 3 labels, and each labels can have 4 different classes, the input for
addMatVectorMultilabelcould be a vector like '{{0,0,0,1}, {1,0,0,0}, {0,1,0,0}}.Now after some conv layer we attach an
Inner_productlayer withnum_output: 12(3*4 = 12). Then we attach to theInnerProductLayerlayer aSoftmaxLossLayerspecifyingslice: 3. So the softmax will be applied independently to each slice of 4 elements of the output of theInnerProductLayerand the loss calculated for each softmax.Actually the cpu version of
SoftmaxLayerandSoftmaxLossLayeris working, but the gpu is not, I wrote some modification on the gpu part of theSoftmaxLayerbut some index is wrong ( the test doesn't pass whenbatch_size>2.I don't know when I will have time to fix that, if someone wants to contribute, any help is appreciated :)