# 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)
# 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)
# 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)