Training neural networks in DeepSqueak

[Anna-V-Kl]

Hello,
I want to detect mouse USVs recorded over long periods of time in semi-natural conditions using DeepSqueak, but the recordings contain a lot of extraneous noise and the quality of calls recorded are not always good. I want to try training a neural network based on the Yolo P2 detector to better distinguishing of mouse USVs from extraneous noise contained in the recordings. But may I ask, if there is any general algorithm for training? In particular, does it make sense to train a neural network on one large data set once or is it better to split the data into several modules and train the neural network on them sequentially, step by step, each time correcting new errors that arise?
I also wanted to ask if the number of images generated at the beginning of the network training process should match the number of USVs manually labeled in DeepSqueak? During my attempts, the number of generated images is always much less than the number of labeled sounds… Is there any limit to the number of images that can be used to train the network? And can we expect that the more images included in the training process, the better the neural network will learn?
I would be grateful for any advice on this matter!