Optimal batch size for triplet loss (hard batch mining)

OML-Team / open-metric-learning

Metric learning and retrieval pipelines, models and zoo.

https://open-metric-learning.readthedocs.io/en/latest/index.html

Apache License 2.0

890 stars 61 forks source link

Optimal batch size for triplet loss (hard batch mining) #626

Closed sujit-dn closed 1 month ago

sujit-dn commented 2 months ago

I am currently training triplet loss with n=2, l=2 with margin 0.5 on custom dataset. Training model with ViTs16 and linear classification layer as projector. Model is able to converge as can be seen from classification loss. However during test time inference no. of matches produced are not very high. Model still gets confused and produces correct pair match with less confidence. Does increasing batch size by training with l=16/32, n=4/8 will help here? vits16_graph_0 5_hard2

AlekseySh commented 2 months ago

hey, @sujit-dn and thank you for your interest in OML!

I would say, bs=4 (n=2, l=2) is a kinda small. Do you have more memory on your GPU?

I like setting batch size as this:

set the number of instances per class in the batch equal to the median size of your class (it works well when it's about 4 or 5)
set the number of classes in the batch as big as your GPU let you to set

AlekseySh commented 1 month ago

due to inactivity

sujit-dn commented 1 month ago

Hey @AlekseySh, We did trained model with semi hard triplet loss num labels =16, num instances per label=2. Found it did indeed improve overall performance of results (in terms of mAP and CMC score) as well as accuracy.

AlekseySh commented 1 month ago

@sujit-dn I'm glad to hear it. Have a nice day!