Olivier Moindrot
Olivier Moindrot
As you say triplet loss is one technique to learn a good embedding. The whole field is called **representation learning**, and the goal is to learn a good representation (the...
One potential issue is that you create a batch on the full CIFAR dataset, which might be too big. Maybe try smaller batches and concatenate the resulting labels: ```python dataset...
You could for instance train a good embedding with triplet loss, then freeze this network and add one or more fully connected with a final softmax. ``` image --> frozen...
@justasbieliauskas : yes the margin parameter can be used as a threshold to classify two images as belonging to the same class or not.
Good point @bartoszpawlowicz
Hi @KarthikeyaKaushik, I think the easiest way to make sure that distances are normalized is to make sure that all the embeddings are in the L2 sphere. You can modify...
What are you measuring when you say "performance is not good"? If you are looking at embeddings on a separate test set, it is possible that the model overfits with...
I'm not sure if the mean embedding will work here. Maybe the person has very different poses, different ages or different hair colors across photos. So you might want to...
It looks like training is collapsing, so you may want to decrease your learning rate maybe? Or maybe have bigger batches to stabilize training? You can also monitor the average...
Maybe pre-training with a softmax loss could help.