def optimize(self, layer_i: int, optimizer: Optimizer):
# 算出weight和bias在優化器中的識別碼後，放入優化器進行優化。
self.linear.weight = optimizer.update(
layer_i * Layer._max_params_n + 0,
self.linear.weight, self.linear.grad_weight)
self.linear.bias = optimizer.update(
layer_i * Layer._max_params_n + 1,
self.linear.bias, self.linear.grad_bias)
def fit(self,
train_images,
train_labels,
test_images,
test_labels,
optimization_method=OptimizationMethod.Adam,
optimization_params={
'learning_rate': 0.01,
'momentum': 0.9,
'beta1': 0.9,
'beta2': 0.999
},
iters_num=10000,
mini_batch_size=100,
samples_num_evaluated_per_epoc=100,
is_activation_check=False):
"""Fits weight paramters by using optimization algorithm.
"""
costs = []
train_accuracies = []
test_accuracies = []
ITERS_NUM = iters_num
MINI_BATCH_SIZE = mini_batch_size
TRAIN_IMAGES_NUM = train_images.shape[0]
ITER_PER_EPOC = max(TRAIN_IMAGES_NUM / MINI_BATCH_SIZE, 1)
optimizer = Optimizer().optimizer(
optimization_method=optimization_method,
learning_rate=optimization_params.get('learning_rate'),
momentum=optimization_params.get('momentum'),
beta1=optimization_params.get('beta1'),
beta2=optimization_params.get('beta2'),
)
for i in range(ITERS_NUM):
batch_mask = np.random.choice(TRAIN_IMAGES_NUM, MINI_BATCH_SIZE)
x_batch = train_images[batch_mask]
t_batch = train_labels[batch_mask]
grads = nn.computeGradientWithBackPropagation(
x_batch, t_batch, is_activation_check=is_activation_check)
optimizer.update(nn.params, grads)
costs.append(nn.computeCost(nn.forward(x_batch), t_batch))
print('cost {}'.format(costs[-1]))
# check accuracy
if i % ITER_PER_EPOC == 0:
print('=========ITERATION {}=========='.format(i))
if samples_num_evaluated_per_epoc is None:
samples_num_evaluated_per_epoc = -1
train_accuracies.append(
nn.computeAccuracy(
train_images[:samples_num_evaluated_per_epoc],
train_labels[:samples_num_evaluated_per_epoc]))
test_accuracies.append(
nn.computeAccuracy(
test_images[:samples_num_evaluated_per_epoc],
test_labels[:samples_num_evaluated_per_epoc]))
print("train accuracy {}, test accuracy {}".format(
train_accuracies[-1], test_accuracies[-1]))
return costs, train_accuracies, test_accuracies