Triplet loss for multi classification ?

[mahfujau]

Hi! Thank you so much for this work. I have a dataset which consists of 7 classes. I want to use your code to classify my data. How can I make datasets labeling?

For an example, you preprocess the images for binary classification:

XXX_train_lmdb_pos1: all images label are 0

XXX_train_lmdb_pos2: all images label are 1

XXX_train_lmdb_neg: all images label are 0

XXX_valrank_lmdb_pos: all images label are 0 except the first image which label is 1

XXX_valrank_lmdb_neg: all images label are 0 except the first image which label is 1

How can we label our images if the total classes are 7? Thank you so much.

[cwhgn]

The samples of three files, XXX_train_lmdb_pos1/pos2/neg, are one-to-one correspondding, which can be used as a triplet for our triplet loss. For example, the three 9th images in XXX_train_lmdb_pos1/pos2/neg form the 9th triplet sample. For the binary classification loss, the sample contains two images. The corresponding two images in XXX_train_lmdb_pos1/pos2 form a positive sample, and the corresponding two images in XXX_train_lmdb_pos1/neg form a negative sample. And the labels in XXX_train_lmdb_pos2/neg are used to be sent to binary classification loss as sample labels.

[mahfuj9346449]

Thank you so much for your reply. Is there any way to classify for multiple categories. For an example: we have different images from 7 categories such as: Pen Computer Book Paper Backpack Pencil Bottle

How we can classify these images using triplet loss?

[cwhgn]

Your problem is a classification issue, i suggest you use a softmax cross entropy loss to solve this issue. If you insist to try the triplet loss, you can build the sample with three images, called a triplet, (img_i_a,img_j_a,img_k_b), where i,j,k are the index of images in their categories, and a,b are the name of categories. img_i_a,img_j_a are two images from the same category, and img_k_b is another image from a different category.