Birdsong Classification

Introduction

In this lesson you will classify bird by their songs. The instructor will give (or already gave) you a short introduction to the topic of neural networks and machine learning. After that you will create the minimal code to run the classification.

Lesson Steps

Part 1: Create a Model class

We will now implement a Python class that is managing the bird classification model¹

A class is an abstract data type that defines a blueprint for objects. It contains attributes and methods that are shared by all objects of this class.

1. Download the files

   • BirdNET_GLOBAL_6K_V2.4_Model_FP32.tflite
   • BirdNET_GLOBAL_6K_V2.4_Labels.txt

   and store it in your package folder birdnet_mini under a new folder you name models. Also create an empty __init__.py file in the models folder.

2. Create a file model.py in the birdnet_mini folder. This will contain all your model class. Your directory should now look like this:

   └───birdnet-mini
       ├───birdnet_mini
       |       audio.py
       │       main.py
       |       model.py
       |       __init__.py
       ├───models
       │       BirdNET_GLOBAL_6K_V2.4_Labels.txt
       │       BirdNET_GLOBAL_6K_V2.4_Model_FP32.tflite
       │       __init__.py
       └───testdata
               test_1min.wav
   

3. Add the following code to the model.py file. This is the basic structure of the model class.

   class Model:
       def __init__(self, model_path=None, num_threads=1):
           raise NotImplementedError
   
       def predict(self, samples):
           raise NotImplementedError
   
       @staticmethod
       def _flat_sigmoid(x, sensitivity=-1.0):
           return 1 / (1.0 + np.exp(sensitivity * np.clip(x, -15, 15)))
   

4. The __init__ method initializes the object and sets all important class members, this is usually called constructor. We will create the interpreter that is running the model. We use Tensorflow Lite, since this is also available on mini computers such as the RaspberryPi. Load the model file using the tflite.Interpreter function and store the interpreter as the member variable self._interpreter.

5. Allocate the tensors for the interpreter an then retrieve input and output information from the model and store it in the member variables

   • self._input_layer_index
   • self._output_layer_index

6. Once we have loaded the model in the __init__function we will now implement the predict function. This function will take a single chunk of samples as input and return the prediction of the model. The steps are:

   • convert the samples to a Numpy array of type float32
   • set the input tensor of the interpreter with the sample data via the set_tensor function
   • invoke the interpreter with the TFLite invoke function
   • retrieve the result tensor via the get_tensor function
   • apply the sigmoid function self._flat_sigmoid to the result tensor and return the result ().

7. Now it is time to test your model. In the main.py file import the Model class and create an instance of it. Load the test file and split it into chunks of 3 seconds. Then predict the class of the first chunk and print the results.

---

1. All models used in this Tutorial are based on the BirdNET model by Stefan Kahl. The model is licensed under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License. You can find more information about the model here and on this GitHub page . ↩