def gradient_check():
error_function = lambda o: o.sum()
rl = RecurrentLayer(3, 2, IdentityActivator(), 1e-3)
x, d = data_set()
rl.forward(x[0])
rl.forward(x[1])
sensitivity_array = np.ones(rl.state_list[-1].shape, dtype=np.float64)
rl.backward(sensitivity_array, IdentityActivator())
epsilon = 10e-4
for i in range(rl.W.shape[0]):
for j in range(rl.W.shape[1]):
rl.W[i, j] += epsilon
rl.reset_state()
rl.forward(x[0])
rl.forward(x[1])
err1 = error_function(rl.state_list[-1])
rl.W[i, j] -= 2 * epsilon
rl.reset_state()
rl.forward(x[0])
rl.forward(x[1])
err2 = error_function(rl.state_list[-1])
expect_grad = (err1 - err2) / (2 * epsilon)
rl.W[i, j] += epsilon
print 'weights(%d,%d): expected - actural %f - %f' % (
i, j, expect_grad, rl.gradient[i, j])
def gradient_check():
'''
梯度检查
'''
print("gradient_check Test!")
# 设计一个误差函数，取所有节点输出项之和
error_function = lambda o: o.sum()
rl = RecurrentLayer(3, 2, IdentityActivator(), 1e-3)
# 计算forward值
x, d = data_set()
rl.forward(x[0])
rl.forward(x[1])
# 求取sensitivity map
sensitivity_array = np.ones(rl.state_list[-1].shape, dtype=np.float64)
# 计算梯度
rl.backward(sensitivity_array, IdentityActivator())
# 检查梯度
epsilon = 10e-4
for i in range(rl.W.shape[0]):
for j in range(rl.W.shape[1]):
rl.W[i, j] += epsilon
rl.reset_state()
rl.forward(x[0])
rl.forward(x[1])
err1 = error_function(rl.state_list[-1])
rl.W[i, j] -= 2 * epsilon
rl.reset_state()
rl.forward(x[0])
rl.forward(x[1])
err2 = error_function(rl.state_list[-1])
expect_grad = (err1 - err2) / (2 * epsilon)
rl.W[i, j] += epsilon
print('weights(%d,%d): expected - actural %f - %f' %
(i, j, expect_grad, rl.gradient[i, j]))
def init_test():
a = np.array([[[0, 1, 1, 0, 2], [2, 2, 2, 2, 1], [1, 0, 0, 2, 0],
[0, 1, 1, 0, 0], [1, 2, 0, 0, 2]],
[[1, 0, 2, 2, 0], [0, 0, 0, 2, 0], [1, 2, 1, 2, 1],
[1, 0, 0, 0, 0], [1, 2, 1, 1, 1]],
[[
2,
1,
2,
0,
0,
], [1, 0, 0, 1, 0], [0, 2, 1, 0, 1], [0, 1, 2, 2, 2],
[2, 1, 0, 0, 1]]])
b = np.array([[[0, 1, 1], [2, 2, 2], [1, 0, 0]],
[[1, 0, 2], [0, 0, 0], [1, 2, 1]]])
c1 = ConvLayer(5, 5, 3, 3, 3, 2, 1, 2, IdentityActivator(), 0.001)
c1.filters[0].weights = np.array([[[-1, 1, 0], [0, 1, 0], [0, 1, 1]],
[[-1, -1, 0], [0, 0, 0], [0, -1, 0]],
[[0, 0, -1], [0, 1, 0], [1, -1, -1]]],
dtype=np.float64)
c1.filters[0].bias = 1
c1.filters[1].weights = np.array([[[1, 1, -1], [-1, -1, 1], [0, -1, 1]],
[[0, 1, 0], [-1, 0, -1], [-1, 1, 0]],
[[-1, 0, 0], [-1, 0, 1], [-1, 0, 0]]],
dtype=np.float64)
return a, b, c1
def init_test():
# channel=3*height=5*width=5
a = np.array([[[0, 1, 1, 0, 2], [2, 2, 2, 2, 1], [1, 0, 0, 2, 0],
[0, 1, 1, 0, 0], [1, 2, 0, 0, 2]],
[[1, 0, 2, 2, 0], [0, 0, 0, 2, 0], [1, 2, 1, 2, 1],
[1, 0, 0, 0, 0], [1, 2, 1, 1, 1]],
[[2, 1, 2, 0, 0], [1, 0, 0, 1, 0], [0, 2, 1, 0, 1],
[0, 1, 2, 2, 2], [2, 1, 0, 0, 1]]])
#3*3*3
b = np.array([[[0, 1, 1], [2, 2, 2], [1, 0, 0]],
[[1, 0, 2], [0, 0, 0], [1, 2, 1]]])
# def ConvLayer.__init__(self, input_width, input_height,
# channel_number, filter_width,
# filter_height, filter_number,
# zero_padding, stride, activator,
# learning_rate):
cl = ConvLayer(5, 5, 3, 3, 3, 2, 1, 2, IdentityActivator(), 0.001)
cl.filters[0].weights = np.array([[[-1, 1, 0], [0, 1, 0], [0, 1, 1]],
[[-1, -1, 0], [0, 0, 0], [0, -1, 0]],
[[0, 0, -1], [0, 1, 0], [1, -1, -1]]],
dtype=np.float64)
cl.filters[0].bias = 1
cl.filters[1].weights = np.array([[[1, 1, -1], [-1, -1, 1], [0, -1, 1]],
[[0, 1, 0], [-1, 0, -1], [-1, 1, 0]],
[[-1, 0, 0], [-1, 0, 1], [-1, 0, 0]]],
dtype=np.float64)
return a, b, cl
def test_bp():
a, b, cl = init_test()
cl.backward(a, b, IdentityActivator())
cl.update()
print(cl.filters[0])
print(cl.filters[1])
def gradient_check():
'''
梯度检查
'''
# 设计一个误差函数，取所有节点输出项之和
error_function = lambda o: o.sum()
# 计算forward值
a, b, cl = init_test()
cl.forward(a)
# 求取sensitivity map
sensitivity_array = np.ones(cl.output_array.shape, dtype=np.float64)
# 计算梯度
cl.backward(a, sensitivity_array, IdentityActivator())
# 检查梯度
epsilon = 10e-4
for d in range(cl.filters[0].weights_grad.shape[0]):
for i in range(cl.filters[0].weights_grad.shape[1]):
for j in range(cl.filters[0].weights_grad.shape[2]):
cl.filters[0].weights[d, i, j] += epsilon
cl.forward(a)
err1 = error_function(cl.output_array)
cl.filters[0].weights[d, i, j] -= 2 * epsilon
cl.forward(a)
err2 = error_function(cl.output_array)
expect_grad = (err1 - err2) / (2 * epsilon)
cl.filters[0].weights[d, i, j] += epsilon
print('weights(%d,%d,%d): expected - actural %f - %f' %
(d, i, j, expect_grad, cl.filters[0].weights_grad[d, i,
j]))
def gradient_check():
error_function = lambda o: o.sum()
rnn = RecursiveLayer(2, 2, IdentityActivator(), 1e-3)
x, d = data_set()
rnn.forward(x[0], x[1])
rnn.forward(rnn.root, x[2])
sensitivity_array = np.ones((rnn.node_width, 1), dtype=np.float64)
rnn.backward(sensitivity_array)
epsilon = 10e-4
for i in range(rnn.W.shape[0]):
for j in range(rnn.W.shape[1]):
rnn.W[i, j] += epsilon
rnn.reset_state()
rnn.forward(x[0], x[1])
rnn.forward(rnn.root, x[2])
err1 = error_function(rnn.root.data)
rnn.W[i, j] -= 2 * epsilon
rnn.reset_state()
rnn.forward(x[0], x[1])
rnn.forward(rnn.root, x[2])
err2 = error_function(rnn.root.data)
expect_grad = (err1 - err2) / (2 * epsilon)
rnn.W[i, j] += epsilon
print 'weights(%d,%d): expected - actural %.4e - %.4e' % (
i, j, expect_grad, rnn.W_grad[i, j])
return rnn
def test():
l = LstmLayer(3, 2, 1e-3)
x, d = data_set()
l.forward(x[0])
l.forward(x[1])
l.backward(x[1], d, IdentityActivator())
return l
def gradient_check():
error_function = lambda o: o.sum()
lstm = LstmLayer(3, 2, 1e-3)
x, d = data_set()
lstm.forward(x[0])
lstm.forward(x[1])
sensitivity_array = np.ones(lstm.h_list[-1].shape, dtype=np.float64)
lstm.backward(x[1], sensitivity_array, IdentityActivator())
epsilon = 10e-4
for i in range(lstm.Wfh.shape[0]):
for j in range(lstm.Wfh.shape[1]):
lstm.Wfh[i, j] += epsilon
lstm.reset_state()
lstm.forward(x[0])
lstm.forward(x[1])
err1 = error_function(lstm.h_list[-1])
lstm.Wfh[i, j] -= 2 * epsilon
lstm.reset_state()
lstm.forward(x[0])
lstm.forward(x[1])
err2 = error_function(lstm.h_list[-1])
expect_grad = (err1 - err2) / (2 * epsilon)
lstm.Wfh[i, j] += epsilon
print 'weights(%d,%d): expected - actural %.4e - %.4e' % (
i, j, expect_grad, lstm.Wfh_grad[i, j])
return lstm
def test():
lstm = LstmLayer(3, 2, 1e-3)
X, D = data_set()
lstm.forward(X[0])
lstm.forward(X[1])
lstm.backward(X[1], D, IdentityActivator())
return lstm
def init_test():
# 3*5*5
a = np.array([[[0, 1, 1, 0, 2], [2, 2, 2, 2, 1], [1, 0, 0, 2, 0],
[0, 1, 1, 0, 0], [1, 2, 0, 0, 2]],
[[1, 0, 2, 2, 0], [0, 0, 0, 2, 0], [1, 2, 1, 2, 1],
[1, 0, 0, 0, 0], [1, 2, 1, 1, 1]],
[[2, 1, 2, 0, 0], [1, 0, 0, 1, 0], [0, 2, 1, 0, 1],
[0, 1, 2, 2, 2], [2, 1, 0, 0, 1]]])
# 2*3*3
b = np.array([[[0, 1, 1], [2, 2, 2], [1, 0, 0]],
[[1, 0, 2], [0, 0, 0], [1, 2, 1]]])
# parameters:（input_width, input_height, channel_number, filter_width, filter_height, filter_number,zero_padding, stride, activator, learning_rate):
cl = ConvLayer(5, 5, 3, 3, 3, 2, 1, 2, IdentityActivator(), 0.001)
# 卷积核为3*3*3
cl.filters[0].weights = np.array([[[-1, 1, 0], [0, 1, 0], [0, 1, 1]],
[[-1, -1, 0], [0, 0, 0], [0, -1, 0]],
[[0, 0, -1], [0, 1, 0], [1, -1, -1]]],
dtype=np.float64)
cl.filters[0].bias = 1
# 3*3*3
cl.filters[1].weights = np.array([[[1, 1, -1], [-1, -1, 1], [0, -1, 1]],
[[0, 1, 0], [-1, 0, -1], [-1, 1, 0]],
[[-1, 0, 0], [-1, 0, 1], [-1, 0, 0]]],
dtype=np.float64)
return a, b, cl
def gradient_check():
'''
梯度检查
我们只对Wfh做了检查，读者可以自行增加对其他梯度的检查
'''
# 设计一个误差函数，取所有节点输出项之和
error_function = lambda o: o.sum()
lstm = LstmLayer(3, 2, 1e-3)
# 计算forward值
x, d = data_set()
lstm.forward(x[0])
lstm.forward(x[1])
# 求取 sensitivity map
sensitivity_array = np.ones(lstm.h_list[-1].shape, dtype=np.float64)
# 计算梯度
lstm.backward(x[1], sensitivity_array, IdentityActivator())
# 检查梯度
epsilon = 10e-4
for i in range(lstm.Wfh.shape[0]):
for j in range(lstm.Wfh.shape[1]):
lstm.Wfh[i, j] += epsilon
lstm.reset_state()
lstm.forward(x[0])
lstm.forward(x[1])
err1 = error_function(lstm.h_list[-1])
lstm.Wfh[i, j] -= 2 * epsilon
lstm.reset_state()
lstm.forward(x[0])
lstm.forward(x[1])
err2 = error_function(lstm.h_list[-1])
expect_grad = (err1 - err2) / (2 * epsilon)
lstm.Wfh[i, j] += epsilon
print('weights(%d,%d): expected - actural %.4e - %.4e' % (i, j, expect_grad, lstm.Wfh_grad[i, j]))
return lstm
def test():
a, b, cl = init_test()
cl.forward(a)
print "前向传播结果:", cl.output_array
cl.backward(a, b, IdentityActivator())
cl.update()
print "反向传播后更新得到的filter1:", cl.filters[0]
print "反向传播后更新得到的filter2:", cl.filters[1]
def test():
lstm = LstmLayer(3, 2, 1e-3)
x, d = data_set()
lstm.forward(x[0])
lstm.forward(x[1])
sensitivity_array = np.ones(lstm.h_list[-1].shape, dtype=np.float64)
lstm.backward(x[1], sensitivity_array, IdentityActivator())
print lstm
def test_bp():
a, b, cl = init_test()
#完成计算误差 ＋ 计算梯度
cl.backward(a, b, IdentityActivator())
#更新权重参数
cl.update()
print(cl.filters[0])
print(cl.filters[1])
def __init__(self, state_size, action_size, seed):
super(QNetwork, self).__init__() # the line you give simply calls the __init__ method of ClassNames parent class.
self.seed = seed.random(seed)
self.fc1 = FCLayer(state_size, 256,ReluActivator(),learning_rate)
self.fc2 = FCLayer(256,128,ReluActivator(),learning_rate)
self.fc3 = FCLayer(128,64,ReluActivator(),learning_rate)
self.out = FCLayer(64, action_size,IdentityActivator() ,learning_rate)
self.data = [self.fc1.parameters,self.fc2.parameters,self.fc3.parameters,self.out.parameters]
def test_bp():
# a : 输入【5，5，3】
# sensitivity map【2, 3，3】
a, b, cl = init_test()
print(cl)
cl.backward(a, b, IdentityActivator())
cl.update()
print(cl.filters[0])
print(cl.filters[1])
def test():
children, d = data_set()
rnn = RecursiveLayer(2, 2, IdentityActivator(), 1e-3)
rnn.forward(children[0], children[1])
rnn.dump()
rnn.forward(rnn.root, children[2])
rnn.dump()
rnn.backward(d)
rnn.dump(dump_grad='true')
return rnn
def gradient_check():
error_function = lambda o: o.sum()
a, b, c1 = init_test()
c1.forward(a)
sensitivity_array = np.ones(c1.output_array.shape, dtype=np.float64)
c1.backward(a, sensitivity_array, IdentityActivator())
epsilon = 10e-4
for d in range(c1.filters[0].weights_grad.shape[0]):
for i in range(c1.filters[0].weights_grad.shape[1]):
for j in range(c1.filters[0].weights_grad.shape[2]):
c1.filters[0].weights[d, i, j] += epsilon
c1.forward(a)
err1 = error_function(c1.output_array)
c1.filters[0].weights[d, i, j] -= 2 * epsilon
c1.forward(a)
err2 - error_function(c1.output_array)
expect_grad = (err1 - err2) / (2 * epsilon)
c1.filters[0].weights[d, i, j] += epsilon
print('weights(%d,%d,%d): expected - actural %f - %f' %
(d, i, j, expect_grad, c1.filters[0].weights_grad[d, i,
j]))
def init_test():
# 输入【3, 5，5】
a = np.array([[[0, 1, 1, 0, 2], [2, 2, 2, 2, 1], [1, 0, 0, 2, 0],
[0, 1, 1, 0, 0], [1, 2, 0, 0, 2]],
[[1, 0, 2, 2, 0], [0, 0, 0, 2, 0], [1, 2, 1, 2, 1],
[1, 0, 0, 0, 0], [1, 2, 1, 1, 1]],
[[2, 1, 2, 0, 0], [1, 0, 0, 1, 0], [0, 2, 1, 0, 1],
[0, 1, 2, 2, 2], [2, 1, 0, 0, 1]]])
#sensitivity map【2, 3，3】
b = np.array([[[0, 1, 1], [2, 2, 2], [1, 0, 0]],
[[1, 0, 2], [0, 0, 0], [1, 2, 1]]])
cl = ConvLayer(5, 5, 3, 3, 3, 2, 1, 2, IdentityActivator(), 0.001)
# cl = ConvLayer(5, 5, 3, 3, 3, 2, 1, 2, ReluActivator(), 0.001)
#重新对Filter的权重进行赋值
cl.filters[0].weights = np.array([[[-1, 1, 0], [0, 1, 0], [0, 1, 1]],
[[-1, -1, 0], [0, 0, 0], [0, -1, 0]],
[[0, 0, -1], [0, 1, 0], [1, -1, -1]]],
dtype=np.float64)
cl.filters[0].bias = 1
cl.filters[1].weights = np.array([[[1, 1, -1], [-1, -1, 1], [0, -1, 1]],
[[0, 1, 0], [-1, 0, -1], [-1, 1, 0]],
[[-1, 0, 0], [-1, 0, 1], [-1, 0, 0]]],
dtype=np.float64)
return a, b, cl