Niveau 1 : Mettre en œuvre la propagation vers l'avant du modèle de réseau de neurones
importer numpy
à partir des couches, importez Convolution, Relu, FullyConnected, MaxPool, SoftmaxWithLoss
classe TinyNet :
def __init__(soi, W_conv1, b_conv1, W_conv2, b_conv2, W_fc, b_fc) :
########## Commencer ##########
self.conv1 = Convolution(W_conv1, b_conv1, stride=1, pad=1)
self.relu1 = Relu()
self.pool1 = MaxPool(2, 2, foulée=2, pad=0)
self.conv2 = Convolution(W_conv2, b_conv2, stride=1, pad=1)
self.relu2 = Relu()
self.pool2 = MaxPool(2, 2, foulée=2, pad=0)
self.fc = Entièrement Connecté(W_fc, b_fc)
self.loss = SoftmaxWithLoss()
########## Fin ##########
def avant (soi, x, t):
########## Commencer ##########
x = self.conv1.forward(x)
x = self.relu1.forward(x)
x = self.pool1.forward(x)
x = self.conv2.forward(x)
x = self.relu2.forward(x)
x = self.pool2.forward(x)
x = self.fc.forward(x)
perte = self.loss.forward(x, t)
retour x, perte
########## Fin ##########
Niveau 2 : Mettre en œuvre la rétropropagation du modèle de réseau de neurones
importer numpy
à partir des couches, importez Convolution, Relu, FullyConnected, MaxPool, SoftmaxWithLoss
classe TinyNet :
def __init__(soi, W_conv1, b_conv1, W_conv2, b_conv2, W_fc, b_fc) :
self.conv1 = Convolution(W_conv1, b_conv1, stride=1, pad=1)
self.relu1 = Relu()
self.pool1 = MaxPool(2, 2, foulée=2, pad=0)
self.conv2 = Convolution(W_conv2, b_conv2, stride=1, pad=1)
self.relu2 = Relu()
self.pool2 = MaxPool(2, 2, foulée=2, pad=0)
self.fc = Entièrement Connecté(W_fc, b_fc)
self.loss = SoftmaxWithLoss()
def avant (soi, x, t):
x = self.conv1.forward(x)
x = self.relu1.forward(x)
x = self.pool1.forward(x)
x = self.conv2.forward(x)
x = self.relu2.forward(x)
x = self.pool2.forward(x)
x = self.fc.forward(x)
perte = self.loss.forward(x, t)
retour x, perte
def vers l'arrière (soi):
########## Commencer ##########
dx = self.loss.backward()
dx = self.fc.backward(dx)
dx = self.pool2.backward(dx)
dx = self.relu2.backward(dx)
dx = self.conv2.backward(dx)
dx = self.pool1.backward(dx)
dx = self.relu1.backward(dx)
dx = self.conv1.backward(dx)
########## Fin ##########
retour self.conv1.dW, self.conv1.db, self.conv2.dW, self.conv2.db, self.fc.dW, self.fc.db
Niveau 3 : Mise en œuvre de la formation de descente de gradient pour les réseaux de neurones
importer numpy
à partir des couches, importez Convolution, Relu, FullyConnected, MaxPool, SoftmaxWithLoss
classe TinyNet :
def __init__(soi, W_conv1, b_conv1, W_conv2, b_conv2, W_fc, b_fc) :
self.conv1 = Convolution(W_conv1, b_conv1, stride=1, pad=1)
self.relu1 = Relu()
self.pool1 = MaxPool(2, 2, foulée=2, pad=0)
self.conv2 = Convolution(W_conv2, b_conv2, stride=1, pad=1)
self.relu2 = Relu()
self.pool2 = MaxPool(2, 2, foulée=2, pad=0)
self.fc = Entièrement Connecté(W_fc, b_fc)
self.loss = SoftmaxWithLoss()
def avant (soi, x, t):
x = self.conv1.forward(x)
x = self.relu1.forward(x)
x = self.pool1.forward(x)
x = self.conv2.forward(x)
x = self.relu2.forward(x)
x = self.pool2.forward(x)
x = self.fc.forward(x)
perte = self.loss.forward(x, t)
retour x, perte
def vers l'arrière (soi):
dx = self.loss.backward()
dx = self.fc.backward(dx)
dx = self.pool2.backward(dx)
dx = self.relu2.backward(dx)
dx = self.conv2.backward(dx)
dx = self.pool1.backward(dx)
dx = self.relu1.backward(dx)
dx = self.conv1.backward(dx)
retour self.conv1.dW, self.conv1.db, self.conv2.dW, self.conv2.db, self.fc.dW, self.fc.db
def train_one_iter(W_conv1, b_conv1, W_conv2, b_conv2, W_fc, b_fc, x, t, learning_rate):
réseau = TinyNet(W_conv1, b_conv1, W_conv2, b_conv2, W_fc, b_fc)
out, loss = network.forward(x, t)
dW_conv1, db_conv1, dW_conv2, db_conv2, dW_fc, db_fc = network.backward()
########## Commencer ##########
new_W_conv1 = W_conv1 - dW_conv1 * learning_rate
new_b_conv1 = b_conv1 - db_conv1 * learning_rate
new_W_conv2 = W_conv2 - dW_conv2 * learning_rate
new_b_conv2 = b_conv2 - db_conv2 * learning_rate
nouveau_W_fc = W_fc - dW_fc * taux_d'apprentissage
nouveau_b_fc = b_fc - db_fc * taux_d'apprentissage
########## Fin ##########
renvoie new_W_conv1, new_b_conv1, new_W_conv2, new_b_conv2, new_W_fc, new_b_fc