A Simple Perceptron Implementation in Python
During my final term of classes for my degree, we were required to learn how perceptron neural networks function. In my free time I came up with the following Python implementation.
class Perceptron(object): def __init__(self, weights=None, threshold=0.5): self.threshold = threshold self.weights = weights self.verbose = False def output(self, input_vector): if self._total(input_vector) < self.threshold: return 0 else: return 1 def train(self, training_set, alpha=0.1, end_after=100): if self.weights is None: self.weights = [0 for _ in range(len(training_set.keys()[0]))] n = 0 updated = True while(updated): n += 1 updated = False for xv, t in training_set.items(): y = self.output(xv) if(y != t): self._update_weights(alpha, t, y, xv) self._update_threshold(alpha, t, y) updated = True if end_after is not None and n >= end_after: break return n def test(self, training_set): for xv, t in training_set.items(): if(self.output(xv) != t): return False return True def _total(self, input_vector): total = 0 for w,x in zip(self.weights, input_vector): total += (w * x) return total def _update_weights(self, alpha, t, y, xv): for i in range(len(self.weights)): self.weights[i] = (alpha * (t - y) * xv[i]) + self.weights[i] def _update_threshold(self, alpha, t, y): self.threshold = (alpha * (t - y) * -1) + self.threshold
As an example of use, here we are training the perceptron to learn how to compute values from a NAND truth table.
nand_set = { (0,0): 1, (0,1): 1, (1,0): 1, (1,1): 0 } nn = Perceptron() iterations = nn.train(nand_set, alpha=.2) # alpha is the learning rate print "Trained in {0} iterations.".format(iterations) print "Correct?", nn.test(nand_set) print "Weights: {0}, Threshold: {1}".format(nn.weights, nn.threshold)
Output:
Trained in 9 iterations. Correct? True Weights: [-0.40000000000000002, -0.20000000000000001], Threshold: -0.5
Now that we have the weights (-0.4, -0.2) and threshold (-0.5), we can feed those in later without having to train it.
nand_perceptron = Perceptron((-0.4, -0.2), -0.5) print "Correct?", nand_perceptron.test(nand_set)
Output:
Correct? True
This was mostly just a fun little project of mine in order to help my understanding, but if you have any suggestions for improvement, I welcome it!
Marcel wrote,
Nice Evan, blog!
I also have an implementation with Adaline and Perceptron in Python
http://aimotion.blogspot.com/2008/12/artigo-introduzindo-redes-neurais-e_5497.html
http://aimotion.blogspot.com/2008/12/ol-todos-nesse-post-irei-apresentar.html
It’s in portuguese, but if you use google translate you can understand! Keep with this work!
Nice to meet you,
Marcel
A. wrote,
Nice, but :
[0] * len(training_set.keys()[0])
would be more “pythonic” and readable (i think) than :
[0 for _ in range(len(training_set.keys()[0]))]