Request for Detailed Layer-by-Layer Architecture of BirdNET v2.4

[AdrienDeverin]

Hello everyone,

I'm currently working with BirdNET v2.4 and am interested in gaining a deeper understanding of its internal structure. While I am aware that the model generally follows a pipeline that includes audio preprocessing (conversion to a log-Mel spectrogram), several convolutional blocks, a global pooling layer, and fully connected layers leading to the output, I have not been able to find detailed, layer-by-layer documentation.

Specifically, I am looking for:

• The exact number and configuration of convolutional layers (e.g., number of filters, kernel sizes, strides, padding, activation functions, etc.).
• Detailed information on the pooling layers used (e.g., Global Average Pooling parameters).
• The architecture of the fully connected layers, including the number of neurons, activation functions, and any regularization or dropout details.
• Any other architectural details or processing layers that might be part of BirdNET v2.4.

If anyone has access to a detailed breakdown or can point me toward any documentation or publications that include this level of detail, I would greatly appreciate it. This information would be extremely valuable for my research/integration efforts.

Thank you for your time and assistance!

Best regards,

[jarnoh]

To see inside the model, you can load it to https://netron.app/ and it will show you the internals, perhaps that helps?

[AdrienDeverin]

Indeed, it helps, thank you.