Training with own dataset

[terpsiphone]

First, I would like to thank the project contributors for your effort on Bat Detective. It is very effective at detecting bat calls.

Is it hopeful that the finetune function will be repaired in the near future? We work with Asian species, and this function would be extremely helpful for our work.

All the best,
Mei-Ling

[chrmue44]

Hi Mei-Ling,
sorry for the late response, for some reason I did not see your question. I had the same problem and with the help of Oisin I figured out which older version to choose in the repository to do the training on own data.

The training works with the version from around the date of the publishing of the paper. I forked the repo at this point and added 2 batch files to prepare the training data (prep_data.bat) and perform the training process (train.bat):

https://github.com/chrmue44/batdetect2 and switch to branch "oldfashioned_training"

You have to prepare a virtual environment with python 3.10 and install all the old library versions defined in requirements.txt. It does NOT work with currently available versions! When this is done the training is possible. The resulting data is still compatible with the current verion of batdetect2.

Hope this helps.
Cheers
Christian

[chrmue44]

Concerning the training I also have a question. I am using Google Colab and have the problem that after some time the execution of the training gets interrupted (usually after around 100 epochs, 4 - 5 hours).

Every 5 epochs I backup the weights. How can I continue the training starting with the last backup of weights?

Beginning from scratch is not an option because I could not manage to run the script until 200 epochs were reached.

Any hint on this would be highly appreciated!

Best
Christian