ELM-pytorch

Extreme Learning Machine (ELM) implemented in Pytorch.

It's MNIST tutorial with basic ELM algorithm, Online Sequential ELM (OS-ELM), and Convolutional ELM.

You can run the code using cpu or gpu mode.

Requirements

• Python 3.5+
• Pytorch 0.3.1+

Extreme Learning Machine

Usage:

cd mnist

GPU mode: python main_ELM.py

CPU mode: python main_ELM.py --no-cuda

The training was completed in 2.0sec and the accuracy reached 97.77%. (Geforce GTX1080Ti 11GB, #hidden neurons=7000)

In CPU mode, the training was completed in 26.92sec and the accuracy was the same. (intel Core i7-6700K CPU 4.00GHz x 8 64GB RAM, #hidden neurons=7000)

If you do not have enough memory for the training process, reduce the number of hidden neurons and try again.

Online Sequential Extreme Learning Machine

Usage:

cd mnist

GPU mode: python main_ELM.py

CPU mode: python main_ELM.py --no-cuda

The training was completed in 10.0sec and the accuracy reached 97.77%. (Geforce GTX1080Ti 11GB, #hidden neurons=7000, batch_size=1000)

In CPU mode, the training was completed in 100.92sec and the accuracy was the same. (intel Core i7-6700K CPU 4.00GHz x 8 64GB RAM, #hidden neurons=7000, batch_size=1000)

If you do not have enough memory for the training process, reduce the number of hidden neurons and try again.

Convolutional Extreme Learning Machine

Usage:

cd mnist

GPU mode: python main_CNNELM.py

CPU mode: python main_CNNELM.py --no-cuda

The training was completed in 7.2sec and the accuracy reached 98.01%. (Geforce GTX1080Ti 11GB, the code used almost all RAM.)

Network configuration

ConvLayer1: kernel_size=5, #channel=10, padding=1 PoolLayer1: kernel_size=2 ReluLayer1: ConvLayer2: kernel_size=4, #channel=80, padding=1 PoolLayer2: kernel_size=2 ReluLayer2: FCLayer:

In CPU mode, the training was completed in 177.92sec and the accuracy was 98.80%. (intel Core i7-6700K CPU 4.00GHz x 8 64GB RAM, the code used almost all RAM.)

Network configuration

ConvLayer1: kernel_size=5, #channel=10, padding=1 PoolLayer1: kernel_size=2 ReluLayer1: ConvLayer2: kernel_size=4, #channel=450, padding=1 PoolLayer2: kernel_size=2 ReluLayer2: FCLayer:

If you do not have enough memory for the training process, reduce the number of hidden neurons and try again.