2211275
/
gnn
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

updated iRProp-, added comments perceptron

main
romanamo 2024-05-07 11:39:43 +02:00
parent 919eed574a
commit 115d78e063
2 changed files with 14 additions and 13 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

5
uebungen/uebung1.py
View File

@ -15,9 +15,11 @@ grad = np.zeros(3)
def sigmoid(summe): # Transferfunktion def sigmoid(summe): # Transferfunktion
return 1.0/(1.0+np.exp(-1.0*summe)) return 1.0/(1.0+np.exp(-1.0*summe))
# Schwellwertfunktion
def perceptron(output): def perceptron(output):
return max(np.sign(output), 0) return max(np.sign(output), 0)
# Vektorisieren der Schwellwertfunktion
vperceptron = np.vectorize(perceptron) vperceptron = np.vectorize(perceptron)
def learn(): def learn():
@ -25,10 +27,13 @@ def learn():
global train, weight, out, target, learnrate global train, weight, out, target, learnrate
# Neuronenausgabe für alle 4 Trainingsmuster berechnen # Neuronenausgabe für alle 4 Trainingsmuster berechnen
# Ausgabe des Neurons
out = vperceptron(np.matmul(train, weight)) out = vperceptron(np.matmul(train, weight))
for j in range(4): for j in range(4):
for i in range(3): for i in range(3):
# Anpassung der Gewichte nach Perzeptronlernregel
if train[j][i] == 1 and out[j] == 0 and target[j] == 1: if train[j][i] == 1 and out[j] == 0 and target[j] == 1:
weight[i] = weight[i] + train[j][i] weight[i] = weight[i] + train[j][i]
elif train[j][i] == 1 and out[j] == 1 and target[j] == 0: elif train[j][i] == 1 and out[j] == 1 and target[j] == 0:

22
uebungen/uebung2.py
View File

@ -23,11 +23,9 @@ inp_size = 3 # Eingabeneuronen
hid_size = 4 # Hidden-Neuronen hid_size = 4 # Hidden-Neuronen
out_size = 1 # Ausgabeneuron out_size = 1 # Ausgabeneuron
delta_L1 = np.ones((inp_size, hid_size)) * 0.125 delta_L1 = np.ones((inp_size, hid_size)) * 0.1
delta_L2 = np.ones((hid_size, out_size)) * 0.125 delta_L2 = np.ones((hid_size, out_size)) * 0.1
delta_min = 0
delta_max = 50
# Gewichte zufällig initialisieren (Mittelwert = 0) # Gewichte zufällig initialisieren (Mittelwert = 0)
w0 = np.random.random((inp_size, hid_size)) - 0.5 w0 = np.random.random((inp_size, hid_size)) - 0.5
@ -41,14 +39,13 @@ def multiply_learnrate(old, new):
return 0.5 return 0.5
return 1 return 1
v_multiply_learnrate = np.vectorize(multiply_learnrate) v_multiply_learnrate = np.vectorize(multiply_learnrate)
L2_grad_old = np.zeros((4, 1)) L1_grad_old = np.zeros((inp_size, hid_size))
L1_grad_old = np.zeros((3, 4)) L2_grad_old = np.zeros((hid_size, out_size))
# Netzwerk trainieren # Netzwerk trainieren
for i in range(100): for i in range(1000):
# Vorwärtsaktivierung # Vorwärtsaktivierung
L0 = inp L0 = inp
@ -65,19 +62,18 @@ for i in range(100):
L1_delta = L1_error * deriv_sigmoid(L1) L1_delta = L1_error * deriv_sigmoid(L1)
# Gradienten # Gradienten
L2_grad_new = np.matmul(L1.T, L2_delta) L2_grad_new = np.matmul(L1.T, L2_error)
L1_grad_new = np.matmul(L0.T, L1_delta) L1_grad_new = np.matmul(L0.T, L1_error)
# Gewichte aktualisieren # Gewichte aktualisieren
learnrate = 0.1
delta_L1 = np.clip( delta_L1 = np.clip(
delta_L1 * v_multiply_learnrate(L1_grad_old, L1_grad_new), 0, 50) delta_L1 * v_multiply_learnrate(L1_grad_old, L1_grad_new), 0, 50)
delta_L2 = np.clip( delta_L2 = np.clip(
delta_L2 * v_multiply_learnrate(L2_grad_old, L2_grad_new), 0, 50) delta_L2 * v_multiply_learnrate(L2_grad_old, L2_grad_new), 0, 50)
w1 -= learnrate * np.sign(L2_grad_new) * delta_L2 w1 -= np.sign(L2_grad_new) * delta_L2
w0 -= learnrate * np.sign(L1_grad_new) * delta_L1 w0 -= np.sign(L1_grad_new) * delta_L1
# Gradienten aktualisieren # Gradienten aktualisieren