RNN and CNN hyperparameter optimization with genetic algorithms using TF and DEAP

This is part of my Thesis work. Some ideas were based on the book 'Hands-On Genetic Algorithms with Python' by Eyal Wirsansky. You can use it for educational purpose.

Setting up evironment and programm execution

In order to run this program you need to:

1. Create venv from environment.yml file. It's recommended to use conda:

  conda env create -f environment.yml

2. Activate this environment:

  conda activate ./venv

3. Define experiment by setting type of testing model (RRN or CNN) and datasets in configurational file ./src/configs/ga-config.json. When setting datasets check its order in ./src/configs/data-config.json.

4. Set current directory to ./src:

  cd src

5. To run the programm use:

  python main.py

6. When execution is done you'll receive graphs with confusion matrix, loss function and classification accuracy calculated for the NN model with optimized hyperparameters.

Specific runnung modes

If you want to do some specific test, you'll need to comment the whole ---Automated optimization for all datasets block and you can chose some test from those block by uncommenting them:

• ---Generating baseline CNN and RNN models. Train and evaluate RNN and CNN models with baseline hyperparameters.
• ---Specific dataset optimization. Allows you to set manually type of testing NN and dataset.
• ---EDA for timeseries datasets. Will make a EDA(Exploratory Data Analysis) and print a diagramms for every dataset.
• ---Time series analysis. Allows you to plot plot specific batch of data from time series dataset.

NN module description

The NN module located in ./src/models/ directory and has 3 files:

1. File base_model.py has BaseModel class which allows to create NN model no matter its type.
2. File cnn_model.py has cnnModel class which is fully inherits the whole functionality from BaseModel class and add configuration for hidden convolution laeyers. In addiction this class has methods that allow to convert genetic chromosome representation into the list of the hyperparameters of CNN and also methods to build, test and evaluate CNN model.
3. File rnn_model.py has rnnModel class which is fully inherits the whole functionality from BaseModel class and add configuration for hidden recurrent laeyers. In addiction this class has methods that allow to convert genetic chromosome representation into the list of the hyperparameters of RNN and also methods to build, test and evaluate CNN model.

Genetic algorithm module

The whole GA module consists of one file ./src/evolutionary_search/genetic_optimization.py and one class GeneticSearch which allows to finetune the hyperparameters of testing NN model for specific dataset.

Visualizing experiments with TensorBoard

All experiments are logging to the specific directory ./src/experiments/<NN_TYPE>_<DATASET_NAME>. In order to use tensorboard inside root project directory, you'll need to run following script:

  tensorboard --logdir src/experiment

Once it's running you'll receive message in your terminal with a localhost URL that allows to open TensorBoard tool.