#!/usr/bin/env python

import os
import numpy as np
from matplotlib import pyplot

def plot_weight_boundary(w, color, label, linetype='-'):
    pyplot.quiver(0,0, w[0],w[1], angles='xy', scale_units='xy', scale=1, color=color) # draw arrow (0,0)->w
    x = np.arange(-10,10,0.01) # like range()
    pyplot.plot(x, -w[0]/w[1] * x, linetype, color=color, label=label) # perc separation boundary (line)
    pyplot.legend(loc=2) # top left location
    
def gen(): # generate random separable data set
    w = np.random.rand(2) - 0.5 # from [0,1) to [-0.5, +0.5)
    w /= np.linalg.norm(w) # normalize to unit vector
    print "unit oracle w=", w

    plot_weight_boundary(w, 'blue', 'oracle', '--') # oracle weight vector

    X = np.random.randn(100, 2) * 10 # [-10,10]x[-10,10] range
    Y = X.dot(w) # dot product 
    Y_filter = np.abs(Y) > 1 # # filter: functional margin > 2
    X = X[Y_filter] # the subset that meets margin requirement
    Y = Y[Y_filter] # the corresponding Y subset
    
    # draw data points
    xx, yy = X[Y>1].T # positive points
    pyplot.plot(xx, yy, 'ob') # blue circle
    xx, yy = X[Y<-1].T # negative points
    pyplot.plot(xx, yy, 'sr') # red square 

    pyplot.plot([-10,10],[0,0], lw=0.3, color='black') # from (-10,0) to (10,0), i.e., draw x-axis
    pyplot.plot([0,0],[-10,10], lw=0.3, color='black') # from (0,-10) to (0,10), i.e., draw y-axis
    pyplot.xlim(-10, 10)
    pyplot.ylim(-10, 10)
    pyplot.axes().set_aspect('equal') # equal xy scale! 

    return X, Y
    #return np.concatenate((X[Y>1], X[Y<-1])), np.concatenate((Y[Y>1], Y[Y<-1])) # clustered pos/neg

def perc(data):
    w = np.zeros(2) # (0,0)
    epoch, total, updates = 0, 0, 1
    while updates > 0: # keep working until convergence
        epoch += 1
        updates = 0
        for (x, y) in data:
            if x.dot(w) * y <= 0:
                w += y * x
                updates += 1
        print "epoch %d, updates %d" % (epoch, updates)
        total += updates
    print "converged after %d updates, w: %s" % (total, w)
    w /= np.linalg.norm(w) # normalize to unit vector

    plot_weight_boundary(w, 'black', 'perc')
    
    print "unit perceptron w=", w
    print
    return w

pyplot.ion() # turn interactive mode on

while True:
    X, Y = gen() # generate random data set
    w = perc(zip(X, Y)) # train perceptron
    pyplot.show()
    try:
        a = raw_input() # keyboard prompt
    except:
        break
    pyplot.clf() # clear figure