Galaxy Zoo Classifier

Overview

The Galaxy Zoo Classifier is a framework to build, compile and train deep learning models designed to classify galaxies based on their images. It uses convolutional neural network (CNN) architectures to learn and predict the classes of galaxies. While this was the original aim of the project for which I developed the code, it could serve as a more general framework for .jpg images classification with CNNs.

For practical reasons, I am not including here the datasets used for model training and model testing. You can find the relevant galaxy images at the webpage of the Galaxy Zoo Kaggle challenge.

The Galaxy Challenge

Setup

Before using the Galaxy Zoo Classifier, it is recommended to set up a Conda environment with TensorFlow. This ensures that the necessary dependencies are isolated and managed efficiently. Here are the steps to set up the environment:

1. First, install Anaconda or Miniconda if you haven't already. Anaconda is a distribution of Python and R for scientific computing, while Miniconda is a smaller, "minimal installer" version that only includes Conda and its dependencies.

2. Once you have Anaconda or Miniconda installed, create a new Conda environment. You can name it whatever you want, but in this example, we'll call it galaxy_env. To create the environment, open your terminal and run the following command:

   conda create --name galaxy_env

3. Activate the newly created environment by running:

   conda activate galaxy_env

4. Install TensorFlow in the Conda environment. You might prefer to choose the latest stable version:

   conda install -c conda-forge tensorflow

5. Now that TensorFlow is installed, clone the Galaxy Zoo Classifier project into your desired directory and navigate to the project's root directory in the terminal.

6. Install the other necessary dependencies by running the following command. You might need to modify the requirements.txt file to specify the version you want to install:

   pip install -r requirements.txt

Usage

To run the Galaxy Zoo Classifier without saving the model or training history, you can use the following instructions:

1. Prepare your data:

   • Ensure that your galaxy images are stored in the directory specified by paths.images_path.
   • Create a YAML configuration file (config.yaml) and update it with your specific paths and parameters. An example configuration file is provided.
   • Modify the configuration file according to your needs, including data preprocessing settings, model parameters, and training parameters.

2. Run the code by executing the run.py script with the configuration file as a command-line argument:

   python run.py --config config.yaml --no-save_model --no-save_history --seed 48

3. The code will load the configuration, preprocess the data, build the model, train it, evaluate its performance, and generate plots and a confusion matrix. However, the model and training history will not be saved.

Note that the --no-save_model and --no-save_history flags are used to disable the saving of the model and training history, respectively. By default, both saving options are enabled (--save_model and --save_history), so you need to explicitly specify the flags to disable them. The --seed flag is used to set the random seed for all the randomized processes. The value 48 is the default value, being the value that must be chosen for reproducibility of the results.

Configuration

The configuration file (config.yaml) contains the following sections. For more information about the role of each configuration parameter, please see the code documentation. A few important notes:

1. The task parameter defines the task to solve. We have only considered the first and second tasks of the proposed project, with values 1 and 2 respectively.

2. One has to set by hand the parameter one_hot_labels to True for each task if desired.

3. The conv_layers and pool_size must be passed as a list of lists. They are transformed to tuples in the run.py file. Being the configuration file .yaml, it does not recognize tuples by default.

4. The parameters which require a number as input such as crop_size, dropout_rate, early_stop_patience and so on must be passed as null if not wanted. The same applies to the data_augmentation_params dictionary. The code was written to take those as None if not desired, but .yaml files require null which then is transformed to None in Python.

5. If class_weights set to True the class weights are calculated. However, if you don't use the custom losses the weights are not applied.

paths:
  images_path: 'galaxy-zoo_data/images_training_rev1/'
  labels_path: 'galaxy-zoo_data/training_solutions_rev1.csv'
  plots_path: 'metrics_plots/'
  models_path: 'saved_models/'

data_preprocessing:
  task: 2
  min: 0.5
  one_hot_labels: False
  crop: True
  crop_size: 256
  img_size: [64, 64]
  normalize: True
  grayscale: False
  training_size: 0.80
  test_size: 0.2

model_name: 'Base_ModelII_regression'

model_params:
  model_type: 'base'
  conv_layers: [[32, [3,3]], [64, [3,3]], [128, [3,3]]]
  dense_units: [256, 128]
  batch_normalization: False
  activation: 'relu'
  pool_size: [2, 2]
  flattening: 'Flatten'
  class_weights: False
  out_activation: 'sigmoid'
  dropout_rate: 0.25
  max_out: True
  early_stop_patience: 15
  monitor: 'val_loss'

data_augmentation_params:
  rotation_range: 90
  width_shift_range: 0.01
  height_shift_range: 0.01
  horizontal_flip: True
  vertical_flip: True
  shear_range: 0.015
  zoom_range: 0.15

train_params:
  learning_rate: 0.001
  loss_function: 'mean_squared_error'
  metrics: ['mse', 'accuracy']
  batch_size: 100
  epochs: 150
  threshold: 0.0
  take_weights_log: True

Customization

Feel free to customize the code to fit your specific needs. You can modify the model architecture, experiment with different data preprocessing techniques, or adjust the training parameters. Additionally, you can extend the code by adding new functionality or implementing advanced features.

Dependencies

The Galaxy Zoo Classifier relies on the following dependencies:

• TensorFlow
• NumPy
• argparse

Ensure that these dependencies are installed before running the code.

References

• Example configuration file: config.yaml
• run.py: Main script for running the code
• eval.py: Evaluation functions for model performance
• models.py: Definition of the GalaxyZooClassifier model
• utils.py: Utility functions for data loading and preprocessing

License

The Galaxy Zoo Classifier is open source and distributed under the MIT License.