def gradient(self, x, t):
"""Calculate gradient to weight params using backpropagation.
This can calculate faster than numerical gradient function.
Args:
x (numpy.ndarray): image data which mean input to NN
t (numpy.ndarray): labels
Return:
dictionary: dictionary of gradient to each param.
"""
W1, W2 = self.params['W1'], self.params['W2']
b1, b2 = self.params['b1'], self.params['b2']
grads = {}
batch_num = x.shape[0]
# forward
a1 = np.dot(x, W1) + b1
z1 = sigmoid(a1)
a2 = np.dot(z1, W2) + b2
y = softmax(a2)
# backward
dy = (y - t) / batch_num
grads['W2'] = np.dot(z1.T, dy)
grads['b2'] = np.sum(dy, axis=0)
dz1 = np.dot(dy, W2.T)
da1 = sigmoid_grad(a1) * dz1
grads['W1'] = np.dot(x.T, da1)
grads['b1'] = np.sum(da1, axis=0)
return grads
def backward(self, dz):
# 出力部の逆伝搬（シグモイド版）
dy = fn.sigmoid_grad(self.z) * dz
#print("dy:", dy)
#print("x.T:", self.x.T)
self.b_grad = dy
self.W_grad = np.dot(self.x.T, dy)
dx = np.dot(dy, self.W.T)
#print("dw:", self.W_grad)
#print("W:", self.W)
W_tmp = self.W
#print("dW", np.sum(self.W_grad))
# オプティマイザーによりself.W、self.bの値を更新
self.W = self.W - self.lr * self.W_grad
self.b = self.b - self.lr * self.b_grad
#print("Wn-W", np.sum(self.W - Wn))
#print("W:", self.W)
return dx
def gradient(self, x, t):
w1, w2 = self.dict['w1'], self.dict['w2']
b1, b2 = self.dict['b1'], self.dict['b2']
grads = {}
a1 = np.dot(x, w1) + b1
z1 = sigmoid(a1)
a2 = np.dot(z1, w2) + b2
y = softmax(a2)
num = x.shape[0]
dy = (y - t) / num
grads['w2'] = np.dot(z1.T, dy)
grads['b2'] = np.sum(dy, axis=0)
da1 = np.dot(dy, w2.T)
dz1 = sigmoid_grad(a1) * da1
grads['w1'] = np.dot(x.T, dz1)
grads['b1'] = np.sum(dz1, axis=0)
return grads
def gradient(self, x, t):
W1, W2 = self.params['W1'], self.params['W2']
b1, b2 = self.params['b1'], self.params['b2']
grads = {}
batch_num = x.shape[0]
# forward
a1 = np.dot(x, W1) + b1
z1 = f.sigmoid(a1)
a2 = np.dot(z1, W2) + b2
y = f.softmax(a2)
# backward
dy = (y - t) / batch_num
grads['W2'] = np.dot(z1.T, dy)
grads['b2'] = np.sum(dy, axis=0)
dz1 = np.dot(dy, W2.T)
da1 = f.sigmoid_grad(a1) * dz1
grads['W1'] = np.dot(x.T, da1)
grads['b1'] = np.sum(da1, axis=0)
return grads
def backward(self, dout):
return dout * fun.sigmoid_grad(self.x)