BirdSoundRecognize

Introduce

Here is the training stage for bird sound recognition

I have prepared two versions which are resnet and efficientnet versions

The resnet version refers to the code provided by the author Stefan Kahl on github to modify it to the tensorflow version

And efficientnet is a reference SOUND-BASED BIRD CLASSIFICATION. How group of Polish women used deep… | by Magdalena Kortas | Towards Data Science

Both work in the context of win10

python version is 3.7.10

Equipment
System	Window 10
CPU	intel i7-6700
GPU	GTX 1070
CUDA	8.0
cuDNN	7.5
RAM	24GB
Development Platform	Anaconda
Programming Language	Python

If there is a situation where it cannot be implemented

Welcome to leave a message, I will try my best to answer

The flow chart of the system is as follows:

Model used in smartphone / APP link Link

Dataset

I use crawler to grab all the bird sound files of xeno-canto

ResNet

Process

Please prepare the materials you want to train first and organize them

The folder structure is as follows:

\DATASET-DOWNLOAD\DATASET_INPUT_FOLDER_NAME
|
+---species1
|       species1_1.wav
|       species1_2.wav     
+---species2
|       species2_1.wav
|       species2_2.wav

Then extract features (the settings here use the settings of Stefan Kahl)

First open config.py to change the path of data set input and output

Then you can execute spec.py to extract features

Note: The file for extracting features must be in wav format

If it is mp3 format, it needs to be converted

Please go to the dataset dataset-download and execute mp3_to_wav.py

Remember to modify the path of the mp3 folder

After extracting features

Then can be trained

There are two approaches here, the difference is the speed of training and accuracy

The first method is to use the image file just output directly as input to train

The second method is to make the image file just output into tfrecord for training

However, the floating values of my method 2 training are large and I haven't figured out why

Method 1

Execute mytrain_resnet18_ver1.py

will start training

Method 2

not update

After

Then I'll put the model on my phone and make instant predictions

APP will be introduced in another project (https://github.com/Bobo100/BirdSoundRecognizeAPP)