Image multi-label training with Caffe backend

[beniz]

Multi-label + accuracy with SigmoidCrossEntropy.

This PR is by @fantes and myself, and brings the following:

• Support for ImageDataLayer that takes a train and test lists of images and read them directly at training / testing time without requiring the building of an LMDB
• Support for images with multiple labels: 1, 0 or -1 for ignoring the label in the accuracy and loss itself
• API keywords multi_label and root_folder.

Example data file:

/path/to/img1.jpg 0 0 -1 1 1

At prediction, positive indices indicates predicted labels, negative values correspond to labels to be ignored.

Example:

• service creation

curl -X PUT http://localhost:8080/services/miml -d '{"mllib":"caffe","description":"multilabel test","type":"supervised","model":{"repository":"path/to/model_dir/","templates":"../../templates/caffe/"},"parameters":{"input":{"db":false,"connector":"image","multi_label":true},"mllib":{"template":"resnet_50","nclasses":5,"finetuning":true},"output":{}}}'

• training

curl -X POST http://localhost:8080/train -d '{"service":"miml","async":false,"parameters":{"input":{"db":false,"width":224,"height":224},"mllib":{"gpu":true,"solver":{"iterations":10000,"test_interval":500,"base_lr":0.001,"iter_size":4},"net":{"batch_size":8}},"output":{"measure":["acc"]}},"data":["/path/to/train.txt","/path/to/test.txt"]}'