https://github.com/burnpiro/elm-pure

Pure implementation of ELM (Extreme Learning Machine) in python (just with numpy)
https://github.com/burnpiro/elm-pure

extreme-learning-machine machine-learning ml

Last synced: 14 days ago
JSON representation

Pure implementation of ELM (Extreme Learning Machine) in python (just with numpy)

• Host: GitHub
• URL: https://github.com/burnpiro/elm-pure
• Owner: burnpiro
• Created: 2020-03-01T21:36:00.000Z (over 4 years ago)
• Default Branch: master
• Last Pushed: 2020-05-13T20:25:14.000Z (over 4 years ago)
• Last Synced: 2023-02-28T15:01:04.138Z (over 1 year ago)
• Topics: extreme-learning-machine, machine-learning, ml
• Language: Jupyter Notebook
• Homepage:
• Size: 180 KB
• Stars: 25
• Watchers: 3
• Forks: 18
• Open Issues: 2
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

        
# Minimalist implementation of Extreme Learning Machine

![ELM structure](./ELM.png)

This is an implementation of Extreme Learning Machine as defined in [Extreme Learning Machine: A New LearningScheme of Feedforward Neural Networks](https://www.researchgate.net/publication/4116697_Extreme_learning_machine_A_new_learning_scheme_of_feedforward_neural_networks) paper by Guang-Bin Huang, Qin-Yu Zhu, and Chee-Kheong Siew

## Dependencies

Please always check `requirements.txt` for current dependencies

- Python 3.7

- Numpy 1.17

- Keras 2.3

Keras are not used to design model. It's just a great source of datasets :P Feel free remove it and use your own dataset.

## Usage

Currently tests are run on MNIST dataset (it's a hand-written digits dataset). You can change that inside `test.py` file. 

```bash

python test.py

```

If you want, you can load weights into model by passing them as arguments:

- `beta_init`

- `w_init`

- `bias_init`

You can also change `activation` and `loss` function just pass:

- `activation` - `sigmoid`, `fourier`, `hardlimit`

- `loss` - `mse` (mean square error), `mae` (mean absolute error)

### Important

Watch out for computation complexity. Each time you try to __fit__ the model it has to do expensive matrix inversion [Moore–Penrose inverse](https://en.wikipedia.org/wiki/Moore%E2%80%93Penrose_inverse). MNIST dataset has 60k images (__H__ matrix has size of __60000x1024__) and takes around 8.5s to inverse on i7-7820X CPU. Remember about it when changing dataset or number of hidden layers

## Todo

- Implement saving/loading model (`h5py`)

- Implement tests

- Implement performance metric