(Cross-posting here from the data science stack exchange, as my question didn't get any replies. I hope it's okay!)

I've been playing around with YOLOv3 and obtaining some good results on the ~20 custom classes I trained. However, one or two classes look like they can use some additional training data (not a lot, say about 10% more data), which I can provide.

What is the most efficient way to train my model now? Do I need to start training from scratch? Can I just throw in my additional data (with the appropriate changes to the config files etc.) and run the training based on the weight matrix I already acquired, but for a small number of iterations? (1000?) Or is this more like a transfer learning problem now?

Thanks for all tips!


Assuming that you do have a dataset (images + labels/bounding boxes) in a format that's required by the training model, you can fine-tune your existing model. You can choose to unlock the final few layers or leave all the layers unlocked during the training process. When I had performed such an experiment with RetinaNet I chose to unlock final few conv layers and was able to achieve slightly higher accuracy.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.