Cnn Keras Image Classifier (with NVIDIA cuda)

Autor: Gabriel Kirsten Menezes (gabriel.kirsten@hotmail.com)

A simple and generic image classifier built with Keras using cuda libraries.

Requirements:

You must Install:

1. Python 2.7;
2. Nvidia cuda libraries;
3. Nvidia cuDCNN libraries;
4. Tensorflow or Theano*;
5. Keras;
6. Sklearn;
7. h5py.

note:
* never tested on Theano.

How to use:

• Install requirements above;
• Prepare the images for processing (the script allows .png imgages, there are some scripts thats can help with tasks to prepare the image database in folder utils, please read the directory structure);
• Run main.py python script;

$ sudo python ./main.py -a your_architecture -f your_finetuning_rate

(you must specify the architecture and fineTuningRate, please read the --help)

Architectures available (-a parameter)

    * VGG16
    * VGG19
    * Xception
    * InceptionV3
    * ResNet50
    * MobileNet

Directory structure

    cnn_keras/   
    ├── data (database images)  
    |   ├── train (training images directory)  
    |   |   └── .gitignore (git ignore)
    |   └── validation (validation images directory)  
    |       └── .gitignore (git ignore)
    ├── models_checkpoints (where the model weights are located)
    |   └── .gitignore (git ignore)
    ├── output_images (where the confusion matrix are located) 
    ├── src (source files) 
    |   └── main.py (python source code)  
    ├── utils (some scripts)  
    |   ├── script_convertall.py (script that converts tiff to png in database image folders)  
    |   └── script_split_data.py (script that splits images between train and validation in database image folders)  
    ├── .gitignore (git ignore)  
    └── README.md (some infos)