- It's implement a Perceptron. The learning rate is :
$$\bm{W}^{new} = \bm{W}^{old} + \alpha \bm{e}\bm{p}^T$$ $$\bm{b}^{new} = \bm{b}^{old} + \alpha \bm{e}\b$$
- The program could draw the classes distribution after finishing the learning.
- The program can also learning at different learning rates to compare within them.
- Run it simply by
python perceptron.py
- It's an implementation of MLQP. The activation function is
$$y_i = sigmoid(\sum_i (u_{ji}y_i^2+v_{ji}y_i) +b_j)$$
- The program could draw the classes distribution after finishing the learning.
- The program can also learning at different learning rates to compare within them.
There are several functions it provides. You can change modes in network.py
net = MLQP.MLQP([2,10,2])
# 1. batch mode
itr = net.SGD_batch(training_data, 10, 0.8) # (., epochs, batch size, learning rate)
# 2. on-line mode
itr = net.SGD_online(training_data, 0.8) # (., epochs, batch size, learning rate)
# verify the test_data
result = net.evaluate(test_data)
# draw the classes distribution
net.plot()
# test the performance of different learning rate
test_performance(training_data, test_data,1.0,1.5,0.05)- Run it simply by
python network.py - When test the performance of different learning rate, it executes 10 times each eta. And it will calculate the average value (not include the max and min ones).
For each classes, the program randomly split it into two samples of the same size, say they are t11, t12, t01 and t02. By merging them together, we get