opinion

[nadongjin]

I ask for your opinion.
If the dataset is too large, there is a low probability that a similar person in another class will enter the batch.
So, all images for each class are vector transformed using a pre-trained engine.
Then, only the most similar people in different classes are constantly grouped into batches to learn.
Personally, I think this can reduce learning time and give more distinction to similar people.
how do you think?

[leondgarse]

I think that's what Triplet doing, the offline triplet mining. Technically, it's digging positive and negative pairs in a large data batch, or even the whole dataset, then feed the hardest ones for model training. I'm testing offline triplet mining + triplet loss with arcface margin for performance and if any improve, as current triplet loss digging only in the batch_size data is not working well.