Should I train from a checkpoint if I have added new classes?

Community Help
1

I’ve added new annotation classes to my dataset, will training from a checkpoint hurt my models performance? Should I train from scratch?

Thanks!

2

Training from a checkpoint, assuming it has seen images from a similar domain, should only help. I wrote a blog post experimenting with this here: https://blog.roboflow.com/transfer-learning-similarity/

1 Like