def gradient(self, x, d):
# forward
self.loss(x, d, train_flg=True)
# backward
dout = 1
dout = self.last_layer.backward(dout)
layers = list(self.layers.values())
layers.reverse()
for layer in layers:
dout = layer.backward(dout)
# 設定
grads = {}
for idx in range(1, self.hidden_layer_num + 2):
grads['W' + str(idx)] = self.layers['Affine' + str(
idx)].dW + self.weight_decay_lambda * self.params['W' +
str(idx)]
grads['b' + str(idx)] = self.layers['Affine' + str(idx)].db
if self.use_batchnorm and idx != self.hidden_layer_num + 1:
grads['gamma' + str(idx)] = self.layers['BatchNorm' +
str(idx)].dgamma
grads['beta' + str(idx)] = self.layers['BatchNorm' +
str(idx)].dbeta
return grads
def gradient(self, x, t):
"""勾配を求める（誤差逆伝播法）
Parameters
----------
x : 入力データ
t : 教師ラベル
Returns
-------
各層の勾配を持ったディクショナリ変数
grads['W1']、grads['W2']、...は各層の重み
grads['b1']、grads['b2']、...は各層のバイアス
"""
# forward
self.loss(x, t)
# backward
dout = 1
dout = self.last_layer.backward(dout)
layers = list(self.layers.values())
layers.reverse()
for layer in layers:
dout = layer.backward(dout)
# 設定
grads = {}
for idx in range(1, self.hidden_layer_num + 2):
grads['W' + str(idx)] = self.layers['Affine' + str(
idx)].d_W + self.weight_decay_lambda * self.layers['Affine' +
str(idx)].W
grads['b' + str(idx)] = self.layers['Affine' + str(idx)].d_b
return grads
def gradient(self, x, d):
# forward
self.loss(x, d)
# backward
dout = 1
dout = self.last_layer.backward(dout)
layers = list(self.layers.values())
layers.reverse()
for layer in layers:
dout = layer.backward(dout)
# 設定
grad = {}
grad['W1'], grad['b1'] = self.layers['Conv1'].dW, self.layers[
'Conv1'].db
grad['W2'], grad['b2'] = self.layers['Conv2'].dW, self.layers[
'Conv2'].db
grad['W3'], grad['b3'] = self.layers['Affine1'].dW, self.layers[
'Affine1'].db
grad['W4'], grad['b4'] = self.layers['Affine2'].dW, self.layers[
'Affine2'].db
return grad
def gradient(self, x_batch, t_batch):
# forward
self.loss(x_batch, t_batch, train_flag=True)
# backward
dout = 1
dout = self.last_layer.backward(d_y=dout)
layers = list(
self.layers.values()) # self.layers is a dict not a list!!!
layers.reverse()
for layer in layers:
dout = layer.backward(dout)
grads = {}
for idx in range(1, len(self.hidden_size_list) + 2):
grads['W' + str(idx)] = self.layers['Affine' + str(
idx)].d_W + self.weight_decay_lambda * self.layers['Affine' +
str(idx)].W
grads['b' + str(idx)] = self.layers['Affine' + str(idx)].d_b
# calculate gradients of gamma & beta
# last Affine layer need no BN
if self.use_batchnorm:
for idx in range(1, self.hidden_layer_num + 1):
grads['gamma' + str(idx)] = self.layers['BatchNorm' +
str(idx)].dgamma
grads['beta' + str(idx)] = self.layers['BatchNorm' +
str(idx)].dbeta
return grads
def gradient(self, x, t):
"""기울기를 구한다
Parameters
----------
x : 입력 데이터
t : 정답 레이블
Returns
-------
각 층의 기울기를 담은 사전(dictionary) 변수
grads['W1']、grads['W2']、... 각 층의 가중치
grads['b1']、grads['b2']、... 각 층의 편향
"""
# forward
loss = self.loss(x, t, train_flg=True)
# backward
dout = 1
dout = self.last_layer.backward(dout)
layers = list(self.layers.values())
layers.reverse()
for layer in layers:
dout = layer.backward(dout)
# 결과 저장
grads = {}
# hidden_layer_num + 1 만큼 (출력층 Affine 포함)
for idx in range(1, self.hidden_layer_num + 2):
# 각 Affine 층의 가중치 매개변수에 가중치 감소의 미분값을 더해준다.
grads['W' + str(idx)] = self.layers['Affine' + str(
idx)].dW + self.weight_decay_lambda * self.params['W' +
str(idx)]
grads['b' + str(idx)] = self.layers['Affine' + str(idx)].db
# < BatchNormalization 계층 사용한다면 매개변수 갱신해준다.>
if self.use_batchnorm and idx != self.hidden_layer_num + 1:
grads['gamma' + str(idx)] = self.layers['BatchNorm' +
str(idx)].dgamma
grads['beta' + str(idx)] = self.layers['BatchNorm' +
str(idx)].dbeta
return grads
def gradient(self, x, d):
# forward
self.loss(x, d)
# backward
dout = 1
dout = self.last_layer.backward(dout)
layers = list(self.layers.values())
layers.reverse()
for layer in layers:
dout = layer.backward(dout)
# 設定
grad = {}
for idx in range(1, self.hidden_layer_num+2):
grad['W' + str(idx)] = self.layers['Affine' + str(idx)].dW
grad['b' + str(idx)] = self.layers['Affine' + str(idx)].db
return grad
def gradient(self, x, t):
# forward
self.loss(x, t)
# backward
dout = 1
dout = self.lastLayer.backward(dout)
# 반대 순서로 각 계층의 backward() 메서드를 호출하기만 하면 처리된다
layers = list(self.layers.values())
layers.reverse()
for layer in layers:
dout = layer.backward(dout)
# 결과 저장
grads = {}
grads['W1'], grads['b1'] = self.layers['Affine1'].dW, self.layers[
'Affine1'].db
grads['W2'], grads['b2'] = self.layers['Affine2'].dW, self.layers[
'Affine2'].db
return grads
def gradient(self, x_batch, t_batch):
# forward
logging.info('Forward Start...')
self.loss(x_batch, t_batch)
logging.info('Forward End.')
# backward
logging.info('Backward Start...')
dout = 1
logging.info('Loss Layer> {}'.format(self.last_layer))
dout = self.last_layer.backward(d_y=dout)
layers = list(
self.layers.values()) # self.layers is a dict not a list!!!
layers.reverse()
for layer in layers:
dout = layer.backward(dout)
logging.info('Backward Layer> {}'.format(layer))
logging.info('Backward End.')
grad = {}
for idx in range(1, len(self.hidden_size_list) + 2):
grad['W' + str(idx)] = self.layers['Affine' + str(idx)].d_W
grad['b' + str(idx)] = self.layers['Affine' + str(idx)].d_b
return grad