Python numerical_gradient_net Exemples

Exemple #1

# 4. training
train_losses = []
train_accuracies = []
test_accuracies = []

for idx in range(1, numiters+1):

    # 4-1. fetch mini-batch
    batch_mask = np.random.choice(sztrain, szbatch)
    train_x_batch = train_x[batch_mask]                 # 100 x 784
    train_t_batch = train_t[batch_mask]                 # 100 x 10

    # 4-2 gradient
    stime = time.time()                 # stopwatch: start
    gradient = network.numerical_gradient_net(train_x_batch, train_t_batch)
    elapsed = time.time() - stime       # stopwatch: end

    # 4-3. update parameter
    for key in network.params:
        network.params[key] -= ratelearning * gradient[key]

    # 4-4. train loss
    loss = network.loss(train_x_batch, train_t_batch)
    train_losses.append(loss)

    # 4-5. accuracy per epoch
    if idx / szepoch == 0:
        train_accuracy = network.accuracy(train_x, train_t)
        train_accuracies.append(train_accuracy)

Exemple #2

# 2. hyperparameter

#==Numerical Gradient=============================

# 3. initialize network
network.initialize(3, 2, 3)

_layers = [
    Affine(network.params['w1'], network.params['b1']),
    ReLU(),
    Affine(network.params['w2'], network.params['b2']),
    SoftmaxWithLoss()
]

grad = network.numerical_gradient_net(_x, _t)
print(grad)

#== Backpropagation Gradient=============================
# 3. initialize network
np.random.seed(3)
network.initialize(3, 2, 3)

_layers = [
    Affine(network.params['w1'], network.params['b1']),
    ReLU(),
    Affine(network.params['w2'], network.params['b2']),
    SoftmaxWithLoss()
]

grad = backpropagation_gradient_net(_x, _t)

Exemple #3

# 4. training
train_losses = []
train_accuracies = []
test_accuracies = []

for idx in range(1, numiters+1):

    # 4-1. fetch mini-batch
    batch_mask = np.random.choice(sztrain, szbatch)     # 60,000 => szbatch
    train_x_batch = train_x[batch_mask]                 # szbatch * 784
    train_t_batch = train_t[batch_mask]                 # szbatch * 10

    # 4.-2. gradient
    start = time.time()                 # stopwatch:start
    gradient = network.numerical_gradient_net(train_x_batch, train_t_batch)     # 모든 파라메터들에 대한 기울기 행렬
    elapsed = time.time() - start       # stopwatch:end

    # 4-3. update parameter
    for key in network.params:
        network.params[key] -= ratelearning * gradient[key]                     # 파라메터들을 각각의 기울기값으로 교정

    # 4-4. train loss
    loss = network.loss(train_x_batch, train_t_batch)
    train_losses.append(loss)

    # 4-5 accuracy per epoch
    if idx % szepoch == 0:
        train_accuracy = network.accuracy(train_x, train_t)
        train_accuracies.append(train_accuracy)