Introduce triplet loss layer to caffe.
Concretely, we use cosine matric to constrain the distance between samples among same label/different labels.
1st, you need create a ordered file list for training.
This file list control the exactly data read-in order during training phase. Suppose in each mini-batch, you have data from 4 labels and 2 samples in each label, then the content of the file list should be like this:
img_path_1_from_label_1 label_1
img_path_2_from_label_1 label_1
img_path_1_from_label_2 label_2
img_path_2_from_label_2 label_2
...
img_path_1_from_label_4 label_4
img_path_2_from_label_4 label_4
2nd, define network structure in your train_val.prototxt.
Setup SampleTripletLayer to sample triplets in each mini-batch. Currently, triplets are made up by all anchor-positive pairs in the sample label and one hardest negative sample from other labels.
layer {
name: "sample_triplet"
type: "SampleTriplet"
bottom: "fully_connected_feature"
top: "triplet"
sample_triplet_param {
label_num: 4
sample_num: 2
}
}
Setup TripletLossLayer to calculate loss.
layer {
name: "triplet_loss"
type: "TripletLoss"
bottom: "fully_connected_feature"
bottom: "triplet"
top: "triplet_loss"
triplet_loss_param {
margin: 0.1
}
loss_weight: 1
}