kahst
Training process stopped early - validation metric not improving
Dear BirdNET Team, I hope this find you well. I am writing this to raise an issue I have been facing with BirdNET for some time. During the training process of my model with BirdNET, whenever I modify the minimum frequency, I receive an error notification stating, "Stopped early – validation metric not improving," which causes the process to stop. However, when I keep the minimum frequency at zero, I can modify the maximum frequency as I wish, and everything works fine. Even with the latest version of BirdNET, the issue persists. Since my target species vocalizes within the frequency range of 40-400 Hz, I would like to filter out low-frequency noise by adjusting the minimum frequency. Any assistance in resolving this issue would be greatly appreciated. Thank you.
The "Stopped early" message pops when the validation is not improving for some epochs, which means the models learning is saturated and the stoppage prevents overfitting, atleast in theory. That means "stopped early" is not an error but a notification as to why training ended before the configured number of training epochs was reached.
Okay great! Thanks Josef for your feedback. That means, if I decrease the epochs, the model could perform properly ?
Let me clarify, the classifier still learns, it's just the training process that gets terminated because no further progress is beeing made (means we've potentially reached a local minimum) I just tried training a classifier with a minimum frequency of 400 Hz and it seems to be working, so there might be an issue with the data at that frequency, but it's hard to tell
@Just177 If your species vocalises in the range 40-400Hz (which is a very low frequency) and you set the minimum bandpass anywhere above 40, you'll be removing the parts of the signal you're looking for from the data. If you take the vocalisation out of the training data, it's not going to learn anything!
On the other hand, if you set the minimum below 40, you will most likely still have a lot of background noise.
