Computer Vision Experimentation Framework

Overview

This project provides a framework for running computer vision experiments involving image classification. It includes components for data loading, preprocessing, model training, and result reporting.

Project Structure

project/
├──src/
├──── config/
│       ├── __init__.py
│       └── config.py
├──── custom_model/
│       ├── __init__.py
│       └── custom_model.py
├──── data_loader/
│       ├── __init__.py
│       ├── data_generator.py
│       ├── data_loader.py
│       └── image_preprocessing.py
├──── experiment/
│       ├── __init__.py
│       ├── experiment.py
│       └── reporting.py
├──── trainer/
│       ├── __init__.py
│       └── trainer.py
├──── utils/
│       ├── __init__.py
│       └── utils.py
└── config.json

Configuration

The configuration file config.json allows you to specify various parameters for your experiments, including:

• data_path: The path to the directory containing the image data.
• seed: A random seed for reproducibility.
• batch_size: The size of the batches used during training.
• patience: The patience for early stopping during training.
• epochs: The number of epochs to train the model.
• target_size: The target size for input images.
• learning_rate: The learning rate for model training.
• distributions: A list of training/validation/test splits to use in experiments.
• segmentations: A list of segmentation methods to apply to the images.
• models: A list of model architectures to use in experiments.

Components

Data Loading and Preprocessing

• DataLoader: Handles loading image data and splitting it into training, validation, and test sets.
• ImagePreprocessing: Contains methods for segmenting images.
• DataGenerator: Creates data generators with augmentation for training, validation, and test sets.

Model Training

• CustomModel: Defines methods to create and compile different CNN architectures.
• Trainer: Handles the training process with early stopping.

Reporting

• Reporting: Generates confusion matrices and classification reports for model performance.

Experiment Management

• Experiment: Manages the setup and execution of experiments, including training, evaluating, and saving models.

Running an Experiment

1. Create a configuration file (config.json) with the desired parameters.
2. Initialize and run the experiment:

from experiment import Experiment

# Initialize Experiment with the path to your config.json
experiment = Experiment(config_path='path/to/config.json')

# Run the experiment
results = experiment.run_experiment()

Saving a Model

To save a specific model based on its model_name, segmentation, and distribution, use the save_model method:

# Save the model
experiment.save_model('ResNet152V2', 'segment_image_by_color', '0.7/0.15', '../model')

Reporting Results

To generate reports from the experiment results:

from reporting import Reporting

# Initialize Reporting
reporting = Reporting()

# Generate reports
reporting.plot_confusion_matrix_and_classification_report(results)