Skip to content

Latest commit

 

History

History
110 lines (83 loc) · 5.05 KB

README.md

File metadata and controls

110 lines (83 loc) · 5.05 KB
Raw

MATLAB/Python Toolbox for Regression/Fitting with ANN

This MATLAB/Python toolbox offers many functionalities for Artificial Neural Network (ANN) regression:

  • Scaling, variable transformation and normalization
  • Checking bounds on the datasets
  • Training/fitting the data
  • Displaying and plotting error metrics**
  • Evaluating the fit for given input data
  • Dumping and reloading the data stored in the object

Currently, the current regression/fitting methods are implemented:

  • ANN regression with MATLAB Deep Learning
  • ANN regression with Python Keras and TensorFlow
  • MATLAB regression with nonlinear least-squares, for benchmark with ANN
  • MATLAB regression with genetic algorithm, for benchmark with ANN

The MATLAB and Python runtimes are communicating over TCP/IP sockets (Python is the server, MATLAB the client). The MATLAB code is the main entry point and is making requests to the Python code.

This toolbox is primarily meant for regression with ANN but is also useful for other methods (such as least-squares or genetic algorithm). This toolbox is made in order to be easily extended by other fitting and regression methods (abstract class).

Getting Started / Example

The following code structure is considered:

  • ann_matlab - MATLAB code, the class "AnnManager" is the main class
  • ann_matlab - Python code, the module "ann_server" is the main module
  • ann_example - MATLAB/Python example
    • run_ann_example.m - the MATLAB main file
    • run_ann_server.py - the Python ANN server main file
    • Shell script (Linux) and (batch) script (MS Windows) for starting the Python ANN server

The examples show the capabilities of the toolbox:

  • ANN regression with MATLAB Deep Learning
  • ANN regression with Python Keras and TensorFlow
  • MATLAB regression with nonlinear least-squares
  • MATLAB regression with genetic algorithm

Compatibility

The toolbox is tested with the following MATLAB setup:

  • Tested with MATLAB R2018a, R2018b and 2019a
  • Deep Learning Toolbox ("neural_network_toolbox")
  • Global Optimization Toolbox ("gads_toolbox")
  • Optimization Toolbox ("optimization_toolbox")
  • Signal Processing Toolbox ("signal_toolbox")

The toolbox is tested with the following Python setup:

  • Python 3.6.8
  • Numpy 1.18.1
  • TensorFlow 2.1.0

However, the toolbox can work without Python (only MATLAB), as long as the Python ANN engine is not used.

The toolbox is known to run with the following operating systems:

  • Linux Ubuntu 18.04.4 LTS
  • Linux CentOS 7.5
  • Microsoft Windows 10

FAQ

Why a custom TCP/IP communication is used and not the provided Python interface of MATLAB?

  • Keras and TensorFlow has specific Python version requirements, they are huge libraries.
  • The Python interface of MATLAB has also requirements for the Python version.
  • Additionally, importing TensorFlow from MATLAB causes crashes with some Python environnement.
  • Finally, this choice limits the coupling between the MATLAB and Python code.

Can this toolbox be used for ANN with unstructured data?

  • No, this library is meant to work for structured data.
  • Unstructured data (text, image, etc.) cannot be used.

Can this toolbox be used for binary classification ANN?

  • The toolbox is mainly meant for regression/fitting.
  • However, with very small adaptation, binary classification can be performed.

Should I use the Keras/TensorFlow or the MATLAB Deep Learning ANN engine?

  • The MATLAB Deep Learning ANN engine is easier to use (fewer parameters, no installation of a Python system).
  • The Keras/TensorFlow ANN engine is more flexible, allowing very advanced tuning.

Can this toolbox handle big data?

  • Depending what is big data, few 10 millions of samples are definitely OK.
  • The memory management model (everything is stored in RAM) does not allow billions of samples.

Why a custom variable scaling and transformation is implemented?

  • To offer more flexibility than the standard scaling.
  • To offer the same options with Keras/TensorFlow and the MATLAB Deep Learning ANN engine.

The variable scaling and transformation looks complex, is it required?

  • Yes, proper scaling of the data is critical for ANNs.
  • For some datasets, without scaling, no proper training is possible.

Can this code run with GNU Octave?

  • The MATLAB Deep Learning toolbox is not compatible with GNU Octave.
  • The TCP/IP communication with Python (serialization) is not compatible with GNU Octave.

Author

  • Thomas Guillod, ETH Zurich, Power Electronic Systems Laboratory - GitHub Profile

Acknowledgement

  • Prof. J.W. Kolar, ETH Zurich, Power Electronic Systems Laboratory
  • P. Papamanolis, ETH Zurich, Power Electronic Systems Laboratory
  • Tim, from MATLAB File Exchange for the inspiration for the serialization code
  • Keras and TensorFlow communities

License

  • This project is licensed under the BSD License, see LICENSE.md.
  • This project is copyrighted by: (c) 2019-2021, ETH Zurich, Power Electronic Systems Laboratory, T. Guillod.