MultiClass label support

[mtamburrano]

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:

• a new method for MemoryDataLayer called addMatVectorMultilabel that accepts a vector < 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 dimension 1since the N-dimension support).
• extended SoftmaxLossLayer to let it to compute the loss on different labels and on slices of bottom[0]. The number of slices are specified with the prototxt parameter slice and must match bottom[1]->shape(softmax_axis_)
• extended SoftmaxLayer to let it to compute the softmax independently on slices of bottom[0], so the SoftmaxLossLayer can compute the loss for each slice comparing it with each corresponding label.

The idea is to feed the network with MemoryDataLayer and addMatVectorMultilabel, then attaching a SoftmaxLossLayer specifying the parameter slice to be equal to the size of the vector of the labels added. It is important to set the output of bottom[0] of the SoftmaxLossLayer to 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 addMatVectorMultilabel could 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_product layer with num_output: 12 (3*4 = 12). Then we attach to the InnerProductLayer layer a SoftmaxLossLayer specifying slice: 3. So the softmax will be applied independently to each slice of 4 elements of the output of the InnerProductLayer and the loss calculated for each softmax.

Actually the cpu version of SoftmaxLayer and SoftmaxLossLayer is working, but the gpu is not, I wrote some modification on the gpu part of the SoftmaxLayer but some index is wrong ( the test doesn't pass when batch_size>2.
I don't know when I will have time to fix that, if someone wants to contribute, any help is appreciated :)

[xdshang]

@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 inner_num_ in caffe. In the above example, you can use ReshapeLayer to reshape (12,) into (4, 3) before forwarding into SoftmaxWithLossLayer, and set softmax_axis_ to the 1st axis. Then inner_num_ will be automatically set to 3 and multi-task classification is executed.

Basically, multilabel classification can also be done in this way, but one needs to duplicate the inner_product output (4-D in your example) number_of_labels (number_of_tasks) times. This needs write a new DuplicateLayer or simply use SplitLayer + ConcatLayer. However, either method will consume more memory space. So in PR #3268, I did some modification in SoftmaxWithLossLayer to support multilabel classification without more memory requirement.