Triplet loss finetuning?

[Davidc2525]

Can you leave an example of how to do finetuning with triplet loss?

[ihsingh2]

Hey,

Given a list of triplets, you need to define a dataset for them. You can refer to these two examples to get an idea of how to define custom datasets. Once that is done, you can setup a dataloader for this dataset, and use it in your training loop. You can use torch.nn.TripletMarginLoss for the loss function.

However, if you are using training.pass_epoch, do note that it expects to unpack a batch of single images at a time (see this). You will have to modify it to unpack a batch of tuple of three images and evaluate each image on the model individually, before passing on to the loss function.

Hope that helps!

[fn-hide]

Any updates?