Is there a way to retrain the detection head with way less classes? #11066
Unanswered
healthmatrice
asked this question in
Q&A
Replies: 1 comment
-
the purpose is to try to increase the recall of person and car. Right now on the low resolution image. The person and car are usually get assign to other classes. However, in our scene we barely has those items. 99% are just persons and cars, thus the recall is low. I wonder maybe reduce the classes can help. If anyone has opposite experience or reason showing this is not the way to go. It will also be helpful. If I do not modify the class number but refine the model with only person and car labeled data while apply some class_weights augmentation^1^, will it be more efficient than retrain the header layers? |
Beta Was this translation helpful? Give feedback.
0 replies
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
-
I want to reduce the 80 classes to 3: people, car, other stuff, background. I wonder is there away to reuse most of the weights in the pretrained model and speed up the training? Then I might just relabel the common training set and feed to it before I start to refine it with my own data.
Can we load only certain layers' weights and recreate some layers for example in this case I would have to recreate the entire header layers, by reducing the channel numbers.
Beta Was this translation helpful? Give feedback.
All reactions