def stochasticGradient(self, train_X, train_y, trainLossVec):
changeWeights = [0] * (self.L - 1)
changeBiases = [0] * (self.L - 1)
lossFuncSum = 0
for i in range(0, BATCH_SIZE):
aVec = []
DVec = []
deltaVec = []
k = random.randint(0, len(train_X) - 1)
xk = train_X[k]
yk = train_y[k]
a = hf.flatten(xk)
aVec.append(a)
# Performs back-propagation for all layers
for l in range(0, self.L - 1):
z = np.matmul(self.weights[l], a) + self.biases[l]
a = hf.relu(z)
D = np.diag(hf.reluPrim(z))
#a = hf.sigmoid(z)
#D = np.diag(hf.sigmoidPrim(z))
aVec.append(a)
DVec.append(D)
delta_L = np.matmul(DVec[-1], (a - hf.formatY(yk)))
deltaVec.append(delta_L)
for l in reversed(range(-self.L + 1, -1)):
delta_l = np.matmul(
DVec[l],
np.matmul(np.transpose(self.weights[l + 1]),
deltaVec[l + 1]))
deltaVec.insert(0, delta_l)
for l in reversed(range(-self.L + 1, 0)):
changeBiases[l] += deltaVec[l]
changeWeights[l] += np.outer(deltaVec[l], aVec[l - 1])
prediction = max(aVec[-1])
index = aVec[-1].index(prediction)
if (index == int(yk)):
self.correct += 1
lossFuncSum += hf.lossFunc(aVec[-1], yk)
trainLossVec.append(lossFuncSum / BATCH_SIZE)
# Calculates average values
dw = [cw / BATCH_SIZE for cw in changeWeights]
db = [cb / BATCH_SIZE for cb in changeBiases]
return dw, db
def predict(self, image):
a = hf.flatten(image)
for l in range(0, self.L - 1):
a = self.nextLayer(a, l)
return a