search

nwojke / cosine_metric_learning

Deep Cosine Metric Learning for Person Re-identification
GNU General Public License v3.0
585 stars 201 forks source link

Which loss to consider when monitoring training with cosine-softmax loss mode? #91

Open samuelsciberras opened 3 years ago

samuelsciberras commented 3 years ago

I am training a custom dataset using cosine-softmax, but all triplet loss, magnet loss, and cross entropy loss are being shown on tensorboard.

Which one should I be looking at when using cosine-softmax loss mode? I understand that cosine-softmax minimizes cross entropy loss, so it should be that graph, but what I cannot understand is why the other graphs are also being adjusted.

Also, what is the recommended amount of data for training?

Thanks

sherlockchou86 commented 3 years ago

https://github.com/nwojke/cosine_metric_learning/blob/master/train_app.py#L651

maybe you can set monitor_xx as False when using cosine-softmax loss mode. the code can also monitor other loss value when using cosine-softmax mode.