deforce: Derivative-Free Algorithms for Optimizing Cascade Forward Neural Networks

deforce (DErivative Free Optimization foR Cascade forward nEural networks) is a Python library that implements variants and the traditional version of Cascade Forward Neural Networks. These include Derivative Free-optimized CFN models (such as genetic algorithm, particle swarm optimization, whale optimization algorithm, teaching learning optimization, differential evolution, ...) and Gradient Descent-optimized CFN models (such as stochastic gradient descent, Adam optimizer, Adelta optimizer, ...). It provides a comprehensive list of optimizers for training CFN models and is also compatible with the Scikit-Learn library. With deforce, you can perform searches and hyperparameter tuning for traditional CFN networks using the features provided by the Scikit-Learn library.

• Free software: GNU General Public License (GPL) V3 license
• Provided Estimator: CfnRegressor, CfnClassifier, DfoCfnRegressor, DfoCfnClassifier, DfoTuneCfn
• Total DFO-based CFN Regressor: > 200 Models
• Total DFO-based CFN Classifier: > 200 Models
• Total GD-based CFN Regressor: 12 Models
• Total GD-based CFN Classifier: 12 Models
• Supported performance metrics: >= 67 (47 regressions and 20 classifications)
• Supported objective functions: >= 67 (47 regressions and 20 classifications)
• Documentation: https://deforce.readthedocs.io
• Python versions: >= 3.8.x
• Dependencies: numpy, scipy, scikit-learn, pandas, mealpy, permetrics, torch, skorch

Citation Request

If you want to understand how to use Derivative Free-optimized Cascade Forward Neural Network, you need to read the paper titled "Optimization of neural-network model using a meta-heuristic algorithm for the estimation of dynamic Poisson’s ratio of selected rock types". The paper can be accessed at the following link

Please include these citations if you plan to use this library:

@software{thieu_deforce_2024,
  author = {Van Thieu, Nguyen},
  title = {{deforce: Derivative-Free Algorithms for Optimizing Cascade Forward Neural Networks}},
  url = {https://github.com/thieu1995/deforce},
  doi = {10.5281/zenodo.10935437},
  year = {2024}
}

@article{van2023mealpy,
  title={MEALPY: An open-source library for latest meta-heuristic algorithms in Python},
  author={Van Thieu, Nguyen and Mirjalili, Seyedali},
  journal={Journal of Systems Architecture},
  year={2023},
  publisher={Elsevier},
  doi={10.1016/j.sysarc.2023.102871}
}

@article{van2023groundwater,
  title={Groundwater level modeling using Augmented Artificial Ecosystem Optimization},
  author={Van Thieu, Nguyen and Barma, Surajit Deb and Van Lam, To and Kisi, Ozgur and Mahesha, Amai},
  journal={Journal of Hydrology},
  volume={617},
  pages={129034},
  year={2023},
  publisher={Elsevier}
}

Installation

• Install the current PyPI release:

$ pip install deforce

After installation, check the installed version by:

$ python
>>> import deforce
>>> deforce.__version__

Examples

Please check documentation website and examples folder.

1. deforce provides this useful classes

from deforce import DataTransformer, Data
from deforce import CfnRegressor, CfnClassifier
from deforce import DfoCfnRegressor, DfoCfnClassifier

2. What can you do with all model classes

from deforce import CfnRegressor, CfnClassifier, DfoCfnRegressor, DfoCfnClassifier

## Use standard CFN model for regression problem
regressor = CfnRegressor(hidden_size=50, act1_name="tanh", act2_name="sigmoid", obj_name="MSE",
                         max_epochs=1000, batch_size=32, optimizer="SGD", optimizer_paras=None, verbose=False, seed=42)

## Use standard CFN model for classification problem 
classifier = CfnClassifier(hidden_size=50, act1_name="tanh", act2_name="sigmoid", obj_name="NLLL",
                           max_epochs=1000, batch_size=32, optimizer="SGD", optimizer_paras=None, verbose=False, seed=42)

## Use Metaheuristic-optimized CFN model for regression problem
print(DfoCfnClassifier.SUPPORTED_OPTIMIZERS)
print(DfoCfnClassifier.SUPPORTED_REG_OBJECTIVES)

opt_paras = {"name": "WOA", "epoch": 100, "pop_size": 30}
regressor = DfoCfnRegressor(hidden_size=50, act1_name="tanh", act2_name="sigmoid",
                            obj_name="MSE", optimizer="OriginalWOA", optimizer_paras=opt_paras, verbose=True, seed=42)

## Use Metaheuristic-optimized CFN model for classification problem
print(DfoCfnClassifier.SUPPORTED_OPTIMIZERS)
print(DfoCfnClassifier.SUPPORTED_CLS_OBJECTIVES)

opt_paras = {"name": "WOA", "epoch": 100, "pop_size": 30}
classifier = DfoCfnClassifier(hidden_size=50, act1_name="tanh", act2_name="softmax",
                              obj_name="CEL", optimizer="OriginalWOA", optimizer_paras=opt_paras, verbose=True, seed=42)

3. After you define the model, do something with it

• Use provides functions to train, predict, and evaluate model

from deforce import CfnRegressor, Data

data = Data()  # Assumption that you have provide this object like above

model = CfnRegressor(hidden_size=50, act1_name="tanh", act2_name="sigmoid", obj_name="MSE",
                     max_epochs=1000, batch_size=32, optimizer="SGD", optimizer_paras=None, verbose=False)

## Train the model
model.fit(data.X_train, data.y_train)

## Predicting a new result
y_pred = model.predict(data.X_test)

## Calculate metrics using score or scores functions.
print(model.score(data.X_test, data.y_test, method="MAE"))
print(model.scores(data.X_test, data.y_test, list_methods=["MAPE", "NNSE", "KGE", "MASE", "R2", "R", "R2S"]))

## Calculate metrics using evaluate function
print(model.evaluate(data.y_test, y_pred, list_metrics=("MSE", "RMSE", "MAPE", "NSE")))

## Save performance metrics to csv file
model.save_evaluation_metrics(data.y_test, y_pred, list_metrics=("RMSE", "MAE"), save_path="history",
                              filename="metrics.csv")

## Save training loss to csv file
model.save_training_loss(save_path="history", filename="loss.csv")

## Save predicted label
model.save_y_predicted(X=data.X_test, y_true=data.y_test, save_path="history", filename="y_predicted.csv")

## Save model
model.save_model(save_path="history", filename="traditional_CFN.pkl")

## Load model 
trained_model = CfnRegressor.load_model(load_path="history", filename="traditional_CFN.pkl")