Darknet: one more question about fine-tuning vs transfer-learning

[shootingliu]

Hello @AlexeyAB
A quick question:
I have modified the (yolo.cfg's) last filters into 45(for 4 classes) and also the classes into 4 to train my own dataset with yolo.weights: ./darknet detector train cfg/yolo4.data cfg/yolo.cfg yolo.weights.
It really worked. I do not know weather it means fine-tuning or not. #245

[AlexeyAB]

Hi @shootingliu

• If all layers changed - this is transfer learning
• If only some top layers changed - this is fine tuning

For fine tuning you can add here: https://github.com/AlexeyAB/darknet/blob/ea09a6e0b38e1ddf43ffcd81d27f0506411eb8e4/cfg/yolo.cfg#L232
this line: stopbackward=1

I added this in some last commit: https://github.com/AlexeyAB/darknet/commit/9ac78d8b84f6a059d2cefe22a10aa60de5b3feaf

[VanitarNordic]

@AlexeyAB

So which of them is better? fine tuning or transfer learning?

Please add some info in the readme about this parameter, stopbackward=1

[AlexeyAB]

@VanitarNordic Yes, I'll add it to readme.

• Transfer-learning (without stopbackward=1) - achieves better accuracy, but trains longer and harder (can lead to Nan), also without stopbackward=1 you can train from scratch without pre-trained weights

• Fine-tuning (with stopbackward=1) - has less accuracy, but trains faster and easier, you can't fine-tune without pre-trained weights

Difference between Trainsfer-learning and Fine-Tuning: https://groups.google.com/d/msg/darknet/WvBFz4zSSH4/k4wZeDuOAgAJ