59 lines
2.0 KiB
Python
59 lines
2.0 KiB
Python
import numpy as np
|
|
import nnwplot
|
|
import matplotlib.pyplot as plt
|
|
|
|
|
|
def error_rate(Y, T):
|
|
if Y.ndim == 1 or Y.shape[0] == 1:
|
|
errors = Y - T
|
|
return errors.sum() / Y.size
|
|
else: # für mehrere Ausgaben in one-hot Kodierung:
|
|
errors = Y.argmax(0) != T.argmax(0)
|
|
return errors.sum() / Y.shape[1]
|
|
|
|
|
|
class SLN:
|
|
def __init__(self, dIn, cOut):
|
|
self.dIn = dIn
|
|
self.cOut = cOut
|
|
np.random.seed(42)
|
|
self._b = np.zeros((cOut, 1))
|
|
self._W = np.random.randn(dIn, cOut) / np.sqrt(dIn + 1)
|
|
|
|
def neuron(self, X):
|
|
N = X.shape[1]
|
|
net = np.zeros(N)
|
|
|
|
for n in range(0, N):
|
|
value = 0
|
|
for i in range(self.dIn):
|
|
value += X[i, n] * self._W[i][0]
|
|
net[n] += value + self._b[0][0]
|
|
return net
|
|
|
|
def delta_train(self, X, T, eta, maxIter, maxErrorRate):
|
|
N = X.shape[1] # Anzahl Trainingsdaten
|
|
plt.ion() # interactive mode on
|
|
for it in range(maxIter):
|
|
Y = self.neuron(X) # nur ein Aufruf
|
|
err = error_rate(Y, T)
|
|
if (it % 10) == 0:
|
|
print('#{} err:{}\n{}\n{}'.format(it, err, self._W, self._b))
|
|
nnwplot.plotTwoFeatures(X, T, self.neuron)
|
|
plt.pause(0.05) # warte auf GUI event loop (je nach python Version notwendig oder nicht)
|
|
if err <= maxErrorRate:
|
|
break
|
|
# update _W
|
|
for n in range(N):
|
|
for j in range(self.dIn):
|
|
weighted_error = err * X[j, n]
|
|
to_learn = weighted_error * eta
|
|
#print("{:.2f} = {} * {} * {}".format(to_learn, err, X[j, n], eta))
|
|
self._W[j, 0] += to_learn
|
|
self._b[0][0] -= to_learn
|
|
# update _b
|
|
...
|
|
print('#{} err:{}\n{}\n{}'.format(it, err, self._W, self._b))
|
|
nnwplot.plotTwoFeatures(X, T, self.neuron)
|
|
plt.pause(0.05) # warte auf GUI event loop (je nach python Version notwendig oder nicht)
|