WTB

Automatic classification of birds for this project : http://www.cabane-oiseaux.org/

Installation

This project requires python 3.6+ Then, just clone this project and run this command from the main dir :

pip install -r requirements.txt

Tensorflow optimization

Depending on your hardware configuration, you can get an optimized version of tensorflow. I used a pre-compiled version for Intel CPU :

pip install https://storage.googleapis.com/intel-optimized-tensorflow/tensorflow-1.11.0-cp36-cp36m-linux_x86_64.whl

If are lucky enough to own a NVIDIA GPU card, you will get best performances with a specific version. Then, use the "tensorflow-gpu" package with a single pip install tensorflow-gpu.

Data

Pictures and classes

The training is a supervised one. That means that you must have labellised pictures of birds: each picture of a class must be stored in a folder which name is the class name. All these folders must be stored themselves in a main folder that will be the entry point of the training system.

Balance the data

As the training system relies on test and training data, the project comes with a tool that can split your original data folder into test, validation and training folders. This tool will also generate new pictures for sub-represented classes (using crop, rotation, ... on original files). For this, use the run_prepare_data.py script from the folder that contains your class folders. By default, 10% of your original dataset will be used for both test and validation data.

python run_prepare_data.py --source=data

Training

The training process relies on this folder structure:

photos/
    |__ photos_test/
        |__ accenteur_mouchet/
            |__ photo01.jpg                
            |__ photo02.jpg                
            |__ ...                
        |__ mesange_bleu/
            |__ photo01.jpg                
            |__ photo02.jpg                
            |__ ...                        
        |__ mesange_charbonniere/
            |__ photo01.jpg                
            |__ photo02.jpg                
            |__ ...                
        |__ ...
    |__ photos_train/
        |__ accenteur_mouchet/
            |__ ...                
        |__ mesange_bleu/
            |__ ...                        
        |__ mesange_charbonniere/
            |__ ...                
        |__ ...        
    |__ photos_valid/
        |__ accenteur_mouchet/
            |__ ...                
        |__ mesange_bleu/
            |__ ...                        
        |__ mesange_charbonniere/
            |__ ...                
        |__ ...

From such a main folder, you can train a new model from the command line

python run_train.py --train_dir=data_train --validation_dir=data_valid

This script produces a model that is stored in the models/ folder with a hdf5 extension.

Prediction

To predict the class of a given image, one may use the predict script:

python predict.py --image_path=photos/truc.jpg --model_path=models/bestmodel-09-0.97.hdf5 --mapper_path=models/mapper.json

where:

• photo/truc.jpg is the file_path of the photo to check
• models/bestmodel-09-0.97.hdf5 is the file_path of the model to use

Run the notebook

All the machine learning stuff has been developed step by step using a jupyter notebook. Feel free to run it to have a more summarized view of the deep learning process. Run jupyter from the project root dir and use your browser to open the file /notebook/WTB.ipynb

python jupyter notebook