def __init__(self, input_width, state_width, learning_rate):
self.input_width = input_width
self.state_width = state_width
self.learning_rate = learning_rate
# 门的激活函数
self.gate_activator = SigmoidActivator()
# 输出的激活函数
self.output_activator = TanhActivator()
# 当前时刻初始化为t0
self.times = 0
# 各个时刻的单元状态向量c
self.c_list = self.init_state_vec()
# 各个时刻的输出向量h
self.h_list = self.init_state_vec()
# 各个时刻的遗忘门f
self.f_list = self.init_state_vec()
# 各个时刻的输入门i
self.i_list = self.init_state_vec()
# 各个时刻的输出门o
self.o_list = self.init_state_vec()
# 各个时刻的即时状态c~
self.ct_list = self.init_state_vec()
# 遗忘门权重矩阵Wfh, Wfx, 偏置项bf
self.Wfh, self.Wfx, self.bf = (self.init_weight_mat())
# 输入门权重矩阵Wfh, Wfx, 偏置项bf
self.Wih, self.Wix, self.bi = (self.init_weight_mat())
# 输出门权重矩阵Wfh, Wfx, 偏置项bf
self.Woh, self.Wox, self.bo = (self.init_weight_mat())
# 单元状态权重矩阵Wfh, Wfx, 偏置项bf
self.Wch, self.Wcx, self.bc = (self.init_weight_mat())
def __init__(self, input_width, state_width,
learning_rate):
self.input_width = input_width
self.state_width = state_width
self.learning_rate = learning_rate
# 门的激活函数
self.gate_activator = SigmoidActivator()
# 输出的激活函数
self.output_activator = TanhActivator()
# 当前时刻初始化为t0
self.times = 0
# 各个时刻的单元状态向量c
self.c_list = self.init_state_vec()
# 各个时刻的输出向量h
self.h_list = self.init_state_vec()
# 各个时刻的遗忘门f
self.f_list = self.init_state_vec()
# 各个时刻的输入门i
self.i_list = self.init_state_vec()
# 各个时刻的输出门o
self.o_list = self.init_state_vec()
# 各个时刻的即时状态c~
self.ct_list = self.init_state_vec()
# 遗忘门权重矩阵Wfh, Wfx, 偏置项bf
self.Wfh, self.Wfx, self.bf = (
self.init_weight_mat())
# 输入门权重矩阵Wfh, Wfx, 偏置项bf
self.Wih, self.Wix, self.bi = (
self.init_weight_mat())
# 输出门权重矩阵Wfh, Wfx, 偏置项bf
self.Woh, self.Wox, self.bo = (
self.init_weight_mat())
# 单元状态权重矩阵Wfh, Wfx, 偏置项bf
self.Wch, self.Wcx, self.bc = (
self.init_weight_mat())
def __init__(self, input_width, state_width, learning_rate):
self.input_width = input_width
self.state_width = state_width
self.learning_rate = learning_rate
self.gate_activator = SigmoidActivator()
self.output_activator = TanhActivator()
self.times = 0
self.c_list = self.init_state_vec()
self.h_list = self.init_state_vec()
self.f_list = self.init_state_vec()
self.i_list = self.init_state_vec()
self.o_list = self.init_state_vec()
self.ct_list = self.init_state_vec()
self.Wfh, self.Wfx, self.bf = (self.init_weight_mat())
self.Wih, self.Wix, self.bi = (self.init_weight_mat())
self.Woh, self.Wox, self.bo = (self.init_weight_mat())
self.Wch, self.Wcx, self.bc = (self.init_weight_mat())
def __init__(self, input_width, state_width, output_width,
learning_rate, penaltyL2, momentum):
self.input_width = input_width
self.state_width = state_width
self.output_width= output_width
self.learning_rate = learning_rate
self.penaltyL2 = penaltyL2
self.momentum = momentum
# 门的激活函数
self.gate_activator = SigmoidActivator()
# 输出的激活函数
self.output_activator = TanhActivator()
self.class_activator = SoftmaxActivator()
# 遗忘门权重矩阵Wfh, Wfx, 偏置项bf
self.Wfh, self.Wfx, self.bf, self.vWfh, self.vWfx, self.vbf =(self.init_weight_mat(0))
# 输入门权重矩阵Wfh, Wfx, 偏置项bf
self.Wih, self.Wix, self.bi, self.vWih, self.vWix, self.vbi =(self.init_weight_mat(0))
# 输出门权重矩阵Wfh, Wfx, 偏置项bf
self.Woh, self.Wox, self.bo, self.vWoh, self.vWox, self.vbo =(self.init_weight_mat(0))
# 单元状态权重矩阵Wfh, Wfx, 偏置项bf
self.Wch, self.Wcx, self.bc, self.vWch, self.vWcx, self.vbc =(self.init_weight_mat(0))
# 下一层权重Wy, 偏值by
self.Wy, self.by, self.vWy, self.vby =(self.init_weight_mat(1))
class LstmLayer(object):
def __init__(self, input_width, state_width, learning_rate):
self.input_width = input_width
self.state_width = state_width
self.learning_rate = learning_rate
# 门的激活函数
self.gate_activator = SigmoidActivator()
# 输出的激活函数
self.output_activator = TanhActivator()
# 当前时刻初始化为t0
self.times = 0
# 各个时刻的单元状态向量c
self.c_list = self.init_state_vec()
# 各个时刻的输出向量h
self.h_list = self.init_state_vec()
# 各个时刻的遗忘门f
self.f_list = self.init_state_vec()
# 各个时刻的输入门i
self.i_list = self.init_state_vec()
# 各个时刻的输出门o
self.o_list = self.init_state_vec()
# 各个时刻的即时状态c~
self.ct_list = self.init_state_vec()
# 遗忘门权重矩阵Wfh, Wfx, 偏置项bf
self.Wfh, self.Wfx, self.bf = (self.init_weight_mat())
# 输入门权重矩阵Wfh, Wfx, 偏置项bf
self.Wih, self.Wix, self.bi = (self.init_weight_mat())
# 输出门权重矩阵Wfh, Wfx, 偏置项bf
self.Woh, self.Wox, self.bo = (self.init_weight_mat())
# 单元状态权重矩阵Wfh, Wfx, 偏置项bf
self.Wch, self.Wcx, self.bc = (self.init_weight_mat())
def init_state_vec(self):
'''
初始化保存状态的向量
'''
state_vec_list = []
state_vec_list.append(np.zeros((self.state_width, 1)))
return state_vec_list
def init_weight_mat(self):
'''
初始化权重矩阵
'''
Wh = np.random.uniform(-1e-4, 1e-4,
(self.state_width, self.state_width))
Wx = np.random.uniform(-1e-4, 1e-4,
(self.state_width, self.input_width))
b = np.zeros((self.state_width, 1))
return Wh, Wx, b
def forward(self, x):
'''
根据式1-式6进行前向计算
'''
self.times += 1
# 遗忘门
fg = self.calc_gate(x, self.Wfx, self.Wfh, self.bf,
self.gate_activator)
self.f_list.append(fg)
# 输入门
ig = self.calc_gate(x, self.Wix, self.Wih, self.bi,
self.gate_activator)
self.i_list.append(ig)
# 输出门
og = self.calc_gate(x, self.Wox, self.Woh, self.bo,
self.gate_activator)
self.o_list.append(og)
# 即时状态
ct = self.calc_gate(x, self.Wcx, self.Wch, self.bc,
self.output_activator)
self.ct_list.append(ct)
# 单元状态
c = fg * self.c_list[self.times - 1] + ig * ct
self.c_list.append(c)
# 输出
h = og * self.output_activator.forward(c)
self.h_list.append(h)
def calc_gate(self, x, Wx, Wh, b, activator):
'''
计算门
'''
h = self.h_list[self.times - 1] # 上次的LSTM输出
net = np.dot(Wh, h) + np.dot(Wx, x) + b
gate = activator.forward(net)
return gate
def backward(self, x, delta_h, activator):
'''
实现LSTM训练算法
'''
self.calc_delta(delta_h, activator)
self.calc_gradient(x)
def update(self):
'''
按照梯度下降，更新权重
'''
self.Wfh -= self.learning_rate * self.Whf_grad
self.Wfx -= self.learning_rate * self.Whx_grad
self.bf -= self.learning_rate * self.bf_grad
self.Wih -= self.learning_rate * self.Whi_grad
self.Wix -= self.learning_rate * self.Whi_grad
self.bi -= self.learning_rate * self.bi_grad
self.Woh -= self.learning_rate * self.Wof_grad
self.Wox -= self.learning_rate * self.Wox_grad
self.bo -= self.learning_rate * self.bo_grad
self.Wch -= self.learning_rate * self.Wcf_grad
self.Wcx -= self.learning_rate * self.Wcx_grad
self.bc -= self.learning_rate * self.bc_grad
def calc_delta(self, delta_h, activator):
# 初始化各个时刻的误差项
self.delta_h_list = self.init_delta() # 输出误差项
self.delta_o_list = self.init_delta() # 输出门误差项
self.delta_i_list = self.init_delta() # 输入门误差项
self.delta_f_list = self.init_delta() # 遗忘门误差项
self.delta_ct_list = self.init_delta() # 即时输出误差项
# 保存从上一层传递下来的当前时刻的误差项
self.delta_h_list[-1] = delta_h
# 迭代计算每个时刻的误差项
for k in range(self.times, 0, -1):
self.calc_delta_k(k)
def init_delta(self):
'''
初始化误差项
'''
delta_list = []
for i in range(self.times + 1):
delta_list.append(np.zeros((self.state_width, 1)))
return delta_list
def calc_delta_k(self, k):
'''
根据k时刻的delta_h，计算k时刻的delta_f、
delta_i、delta_o、delta_ct，以及k-1时刻的delta_h
'''
# 获得k时刻前向计算的值
ig = self.i_list[k]
og = self.o_list[k]
fg = self.f_list[k]
ct = self.ct_list[k]
c = self.c_list[k]
c_prev = self.c_list[k - 1]
tanh_c = self.output_activator.forward(c)
delta_k = self.delta_h_list[k]
# 根据式9计算delta_o
delta_o = (delta_k * tanh_c * self.gate_activator.backward(og))
delta_f = (delta_k * og * (1 - tanh_c * tanh_c) * c_prev *
self.gate_activator.backward(fg))
delta_i = (delta_k * og * (1 - tanh_c * tanh_c) * ct *
self.gate_activator.backward(ig))
delta_ct = (delta_k * og * (1 - tanh_c * tanh_c) * ig *
self.output_activator.backward(ct))
delta_h_prev = (np.dot(delta_o.transpose(), self.Woh) +
np.dot(delta_i.transpose(), self.Wih) +
np.dot(delta_f.transpose(), self.Wfh) +
np.dot(delta_ct.transpose(), self.Wch)).transpose()
# 保存全部delta值
self.delta_h_list[k - 1] = delta_h_prev
self.delta_f_list[k] = delta_f
self.delta_i_list[k] = delta_i
self.delta_o_list[k] = delta_o
self.delta_ct_list[k] = delta_ct
def calc_gradient(self, x):
# 初始化遗忘门权重梯度矩阵和偏置项
self.Wfh_grad, self.Wfx_grad, self.bf_grad = (
self.init_weight_gradient_mat())
# 初始化输入门权重梯度矩阵和偏置项
self.Wih_grad, self.Wix_grad, self.bi_grad = (
self.init_weight_gradient_mat())
# 初始化输出门权重梯度矩阵和偏置项
self.Woh_grad, self.Wox_grad, self.bo_grad = (
self.init_weight_gradient_mat())
# 初始化单元状态权重梯度矩阵和偏置项
self.Wch_grad, self.Wcx_grad, self.bc_grad = (
self.init_weight_gradient_mat())
# 计算对上一次输出h的权重梯度
for t in range(self.times, 0, -1):
# 计算各个时刻的梯度
(Wfh_grad, bf_grad, Wih_grad, bi_grad, Woh_grad, bo_grad, Wch_grad,
bc_grad) = (self.calc_gradient_t(t))
# 实际梯度是各时刻梯度之和
self.Wfh_grad += Wfh_grad
self.bf_grad += bf_grad
self.Wih_grad += Wih_grad
self.bi_grad += bi_grad
self.Woh_grad += Woh_grad
self.bo_grad += bo_grad
self.Wch_grad += Wch_grad
self.bc_grad += bc_grad
# 计算对本次输入x的权重梯度
xt = x.transpose()
self.Wfx_grad = np.dot(self.delta_f_list[-1], xt)
self.Wix_grad = np.dot(self.delta_i_list[-1], xt)
self.Wox_grad = np.dot(self.delta_o_list[-1], xt)
self.Wcx_grad = np.dot(self.delta_ct_list[-1], xt)
def init_weight_gradient_mat(self):
'''
初始化权重矩阵
'''
Wh_grad = np.zeros((self.state_width, self.state_width))
Wx_grad = np.zeros((self.state_width, self.input_width))
b_grad = np.zeros((self.state_width, 1))
return Wh_grad, Wx_grad, b_grad
def calc_gradient_t(self, t):
'''
计算每个时刻t权重的梯度
'''
h_prev = self.h_list[t - 1].transpose()
Wfh_grad = np.dot(self.delta_f_list[t], h_prev)
bf_grad = self.delta_f_list[t]
Wih_grad = np.dot(self.delta_i_list[t], h_prev)
bi_grad = self.delta_f_list[t]
Woh_grad = np.dot(self.delta_o_list[t], h_prev)
bo_grad = self.delta_f_list[t]
Wch_grad = np.dot(self.delta_ct_list[t], h_prev)
bc_grad = self.delta_ct_list[t]
return Wfh_grad, bf_grad, Wih_grad, bi_grad, \
Woh_grad, bo_grad, Wch_grad, bc_grad
def reset_state(self):
# 当前时刻初始化为t0
self.times = 0
# 各个时刻的单元状态向量c
self.c_list = self.init_state_vec()
# 各个时刻的输出向量h
self.h_list = self.init_state_vec()
# 各个时刻的遗忘门f
self.f_list = self.init_state_vec()
# 各个时刻的输入门i
self.i_list = self.init_state_vec()
# 各个时刻的输出门o
self.o_list = self.init_state_vec()
# 各个时刻的即时状态c~
self.ct_list = self.init_state_vec()
class LstmLayer(object):
'''
隐藏层相比rnn（只有一个hidden state）,额外增加cell state
怎样控制长期状态cell state？有三个开关，（每个开关的实现涉及到门的输入到输出的激活计算）:
第一个开关，负责控制继续保存长期状态c；
遗忘门：决定上一个时刻的单元状态Ct-1有多少保留到当前时刻的Ct
1.遗忘门：Ft=f(Wf*[Ht-1,Xt]+Bf)
第二个开关，负责控制把即时状态输入到长期状态c；
输入门：决定当前时刻网络的输入Xt有多少保存到单元状态Ct
1.输入门:It=f(Wc*[Ht-1,Xt]+Bi)
2.当前输入的单元状态Ct' = tanh(Wc*[Ht-1,Xt]+Bc)
3.当前时刻的单元状态（当前的记忆Ct'和长期的记忆Ct-1融合一起）Ct = Ft。Ct-1+It。Ct'
第三个开关，负责控制是否把长期状态c作为当前的LSTM的输出。
输出门：控制单元状态Ct有多少输出到LSTM的当前输出值ht;
1.输出门：Ot = f(Wo*[Ht-1,Xt]+Bo)
2.LSTM最终的输出：Ht = Ot。tanh(Ct)
'''
def __init__(self, input_width, state_width, learning_rate):
self.input_width = input_width
self.state_width = state_width
self.learning_rate = learning_rate
# 门的激活函数
self.gate_activator = SigmoidActivator()
# 输出的激活函数
self.output_activator = TanhActivator()
# 当前时刻初始化为t0
self.times = 0
# 各个时刻的单元cell状态向量c
self.c_list = self.init_state_vec()
# 各个时刻的输出向量h
self.h_list = self.init_state_vec()
# 各个时刻的遗忘门f
self.f_list = self.init_state_vec()
# 各个时刻的输入门i
self.i_list = self.init_state_vec()
# 各个时刻的输出门o
self.o_list = self.init_state_vec()
# 各个时刻的即时状态c~
self.ct_list = self.init_state_vec()
# 遗忘门权重矩阵Wfh, Wfx, 偏置项bf
self.Wfh, self.Wfx, self.bf = (self.init_weight_mat())
# 输入门权重矩阵Wfh, Wfx, 偏置项bf
self.Wih, self.Wix, self.bi = (self.init_weight_mat())
# 输出门权重矩阵Wfh, Wfx, 偏置项bf
self.Woh, self.Wox, self.bo = (self.init_weight_mat())
# 单元状态权重矩阵Wfh, Wfx, 偏置项bf
self.Wch, self.Wcx, self.bc = (self.init_weight_mat())
def init_state_vec(self):
'''
初始化保存状态的向量
'''
state_vec_list = []
state_vec_list.append(np.zeros((self.state_width, 1)))
return state_vec_list
def init_weight_mat(self):
'''
初始化权重矩阵
'''
Wh = np.random.uniform(-1e-4, 1e-4,
(self.state_width, self.state_width))
Wx = np.random.uniform(-1e-4, 1e-4,
(self.state_width, self.input_width))
b = np.zeros((self.state_width, 1))
return Wh, Wx, b
def forward(self, x):
'''
前向传播
'''
self.times += 1
# 遗忘门
fg = self.calc_gate(x, self.Wfx, self.Wfh, self.bf,
self.gate_activator)
self.f_list.append(fg)
# 输入门
ig = self.calc_gate(x, self.Wix, self.Wih, self.bi,
self.gate_activator)
self.i_list.append(ig)
# 输出门
og = self.calc_gate(x, self.Wox, self.Woh, self.bo,
self.gate_activator)
self.o_list.append(og)
# 即时状态
ct = self.calc_gate(x, self.Wcx, self.Wch, self.bc,
self.output_activator)
self.ct_list.append(ct)
# 单元状态
c = fg * self.c_list[self.times - 1] + ig * ct
self.c_list.append(c)
# 输出
h = og * self.output_activator(c)
self.h_list.append(h)
def calc_gate(self, x, Wx, Wh, b, activator):
'''
计算门
'''
h = self.h_list[self.times - 1] # 上次的LSTM输出
net = np.dot(Wh, h) + np.dot(Wx, x) + b
gate = activator.forward(net)
return gate
def backward(self, x, delta_h, activator):
'''
实现lstm训练算法
'''
self.calc_delta(delta_h, activator)
self.calc_gradient(x)
def calc_delta(self, delta_h, activator):
##初始化各个时刻误差
self.delta_h_list = self.init_delta() # 输出误差项
self.delta_o_list = self.init_delta() # 输出门误差项
self.delta_i_list = self.init_delta() # 输入门误差项
self.delta_f_list = self.init_delta() # 遗忘门误差项
self.delta_ct_list = self.init_delta() # 即时输出误差项
# 保存上一层传递下来的当前时刻的误差项
self.delta_h_list[-1] = delta_h
# 迭代计算每个时刻的误差项
for k in range(self.times, 0, -1):
self.calc_delta_k(k)
def calc_delta_k(self, k):
'''
t时刻的误差沿着时间的反向传播公式：
根据k时刻的delta_h，计算k时刻的delta_f、delta_i、delta_o、delta_ct，以及k-1时刻的delta_h
'''
ig = self.i_list[k]
og = self.o_list[k]
fg = self.f_list[k]
ct = self.ct_list[k]
c = self.c_list[k]
c_prev = self.c_list[k - 1]
tanh_c = self.output_activator.forward(c)
delta_k = self.delta_h_list[k]
delta_o = (delta_k * tanh_c * self.gate_activator.backward(og))
delta_f = (delta_k * og * (1 - tanh_c * tanh_c) * c_prev *
self.gate_activator.backward(fg))
delta_i = (delta_k * og * (1 - tanh_c * tanh_c) * ct *
self.gate_activator.backward(ig))
delta_ct = (delta_k * og * (1 - tanh_c * tanh_c) * ig *
self.output_activator.backward(ct))
delta_h_prev = (np.dot(delta_o.transpose(), self.Woh) +
np.dot(delta_i.transpose(), self.Wih) +
np.dot(delta_f.transpose(), self.Wfh) +
np.dot(delta_ct.transpose(), self.Wch)).transpose()
# 保存全部的delta值
self.delta_h_list[k - 1] = delta_h_prev
self.delta_f_list[k] = delta_f
self.delta_i_list[k] = delta_i
self.delta_o_list[k] = delta_o
self.delta_ct_list[k] = delta_ct
def init_delta(self):
'''
初始化误差项
'''
delta_list = []
for i in range(self.times + 1):
delta_list.append(np.zeros((self.state_width, 1)))
return delta_list
def calc_gradient(self, x):
# 初始化遗忘门权重梯度矩阵和偏置项
self.Wfh_grad, self.Wfx_grad, self.bf_grad = (
self.init_weight_gradient_mat())
# 初始化输入门权重梯度矩阵和偏置项
self.Wih_grad, self.Wix_grad, self.bi_grad = (
self.init_weight_gradient_mat())
# 初始化输出门权重梯度矩阵和偏置项
self.Woh_grad, self.Wox_grad, self.bo_grad = (
self.init_weight_gradient_mat())
# 初始化单元状态权重梯度矩阵和偏置项
self.Wch_grad, self.Wcx_grad, self.bc_grad = (
self.init_weight_gradient_mat())
# 计算对上一次输出h的权重梯度
for t in range(self.times, 0, -1):
# 计算各个时刻的梯度
(Wfh_grad, bf_grad, Wih_grad, bi_grad, Woh_grad, bo_grad, Wch_grad,
bc_grad) = (self.calc_gradient_t(t))
def init_weight_gradient_mat(self):
'''
初始化权重矩阵
'''
Wh_grad = np.zeros((self.state_width, self.state_width))
Wx_grad = np.zeros((self.state_width, self.input_width))
b_grad = np.zeros((self.state_width, 1))
return Wh_grad, Wx_grad, b_grad
def calc_gradient_t(self, t):
'''
计算每个时刻t权重的梯度
'''
h_prev = self.h_list[t - 1].transpose()
Wf
class LstmLayer(object):
def __init__(self, input_width, state_width,
learning_rate):
self.input_width = input_width
self.state_width = state_width
self.learning_rate = learning_rate
# 门的激活函数
self.gate_activator = SigmoidActivator()
# 输出的激活函数
self.output_activator = TanhActivator()
# 当前时刻初始化为t0
self.times = 0
# 各个时刻的单元状态向量c
self.c_list = self.init_state_vec()
# 各个时刻的输出向量h
self.h_list = self.init_state_vec()
# 各个时刻的遗忘门f
self.f_list = self.init_state_vec()
# 各个时刻的输入门i
self.i_list = self.init_state_vec()
# 各个时刻的输出门o
self.o_list = self.init_state_vec()
# 各个时刻的即时状态c~
self.ct_list = self.init_state_vec()
# 遗忘门权重矩阵Wfh, Wfx, 偏置项bf
self.Wfh, self.Wfx, self.bf = (
self.init_weight_mat())
# 输入门权重矩阵Wfh, Wfx, 偏置项bf
self.Wih, self.Wix, self.bi = (
self.init_weight_mat())
# 输出门权重矩阵Wfh, Wfx, 偏置项bf
self.Woh, self.Wox, self.bo = (
self.init_weight_mat())
# 单元状态权重矩阵Wfh, Wfx, 偏置项bf
self.Wch, self.Wcx, self.bc = (
self.init_weight_mat())
def init_state_vec(self):
'''
初始化保存状态的向量
'''
state_vec_list = []
state_vec_list.append(np.zeros(
(self.state_width, 1)))
return state_vec_list
def init_weight_mat(self):
'''
初始化权重矩阵
'''
Wh = np.random.uniform(-1e-4, 1e-4,
(self.state_width, self.state_width))
Wx = np.random.uniform(-1e-4, 1e-4,
(self.state_width, self.input_width))
b = np.zeros((self.state_width, 1))
return Wh, Wx, b
def forward(self, x):
'''
根据式1-式6进行前向计算
'''
self.times += 1
# 遗忘门
fg = self.calc_gate(x, self.Wfx, self.Wfh,
self.bf, self.gate_activator)
self.f_list.append(fg)
# 输入门
ig = self.calc_gate(x, self.Wix, self.Wih,
self.bi, self.gate_activator)
self.i_list.append(ig)
# 输出门
og = self.calc_gate(x, self.Wox, self.Woh,
self.bo, self.gate_activator)
self.o_list.append(og)
# 即时状态
ct = self.calc_gate(x, self.Wcx, self.Wch,
self.bc, self.output_activator)
self.ct_list.append(ct)
# 单元状态
c = fg * self.c_list[self.times - 1] + ig * ct
self.c_list.append(c)
# 输出
h = og * self.output_activator.forward(c)
self.h_list.append(h)
def calc_gate(self, x, Wx, Wh, b, activator):
'''
计算门
'''
h = self.h_list[self.times - 1] # 上次的LSTM输出
net = np.dot(Wh, h) + np.dot(Wx, x) + b
gate = activator.forward(net)
return gate
def backward(self, x, delta_h, activator):
'''
实现LSTM训练算法
'''
self.calc_delta(delta_h, activator)
self.calc_gradient(x)
def update(self):
'''
按照梯度下降，更新权重
'''
self.Wfh -= self.learning_rate * self.Whf_grad
self.Wfx -= self.learning_rate * self.Whx_grad
self.bf -= self.learning_rate * self.bf_grad
self.Wih -= self.learning_rate * self.Whi_grad
self.Wix -= self.learning_rate * self.Whi_grad
self.bi -= self.learning_rate * self.bi_grad
self.Woh -= self.learning_rate * self.Wof_grad
self.Wox -= self.learning_rate * self.Wox_grad
self.bo -= self.learning_rate * self.bo_grad
self.Wch -= self.learning_rate * self.Wcf_grad
self.Wcx -= self.learning_rate * self.Wcx_grad
self.bc -= self.learning_rate * self.bc_grad
def calc_delta(self, delta_h, activator):
# 初始化各个时刻的误差项
self.delta_h_list = self.init_delta() # 输出误差项
self.delta_o_list = self.init_delta() # 输出门误差项
self.delta_i_list = self.init_delta() # 输入门误差项
self.delta_f_list = self.init_delta() # 遗忘门误差项
self.delta_ct_list = self.init_delta() # 即时输出误差项
# 保存从上一层传递下来的当前时刻的误差项
self.delta_h_list[-1] = delta_h
# 迭代计算每个时刻的误差项
for k in range(self.times, 0, -1):
self.calc_delta_k(k)
def init_delta(self):
'''
初始化误差项
'''
delta_list = []
for i in range(self.times + 1):
delta_list.append(np.zeros(
(self.state_width, 1)))
return delta_list
def calc_delta_k(self, k):
'''
根据k时刻的delta_h，计算k时刻的delta_f、
delta_i、delta_o、delta_ct，以及k-1时刻的delta_h
'''
# 获得k时刻前向计算的值
ig = self.i_list[k]
og = self.o_list[k]
fg = self.f_list[k]
ct = self.ct_list[k]
c = self.c_list[k]
c_prev = self.c_list[k-1]
tanh_c = self.output_activator.forward(c)
delta_k = self.delta_h_list[k]
# 根据式9计算delta_o
delta_o = (delta_k * tanh_c *
self.gate_activator.backward(og))
delta_f = (delta_k * og *
(1 - tanh_c * tanh_c) * c_prev *
self.gate_activator.backward(fg))
delta_i = (delta_k * og *
(1 - tanh_c * tanh_c) * ct *
self.gate_activator.backward(ig))
delta_ct = (delta_k * og *
(1 - tanh_c * tanh_c) * ig *
self.output_activator.backward(ct))
delta_h_prev = (
np.dot(delta_o.transpose(), self.Woh) +
np.dot(delta_i.transpose(), self.Wih) +
np.dot(delta_f.transpose(), self.Wfh) +
np.dot(delta_ct.transpose(), self.Wch)
).transpose()
# 保存全部delta值
self.delta_h_list[k-1] = delta_h_prev
self.delta_f_list[k] = delta_f
self.delta_i_list[k] = delta_i
self.delta_o_list[k] = delta_o
self.delta_ct_list[k] = delta_ct
def calc_gradient(self, x):
# 初始化遗忘门权重梯度矩阵和偏置项
self.Wfh_grad, self.Wfx_grad, self.bf_grad = (
self.init_weight_gradient_mat())
# 初始化输入门权重梯度矩阵和偏置项
self.Wih_grad, self.Wix_grad, self.bi_grad = (
self.init_weight_gradient_mat())
# 初始化输出门权重梯度矩阵和偏置项
self.Woh_grad, self.Wox_grad, self.bo_grad = (
self.init_weight_gradient_mat())
# 初始化单元状态权重梯度矩阵和偏置项
self.Wch_grad, self.Wcx_grad, self.bc_grad = (
self.init_weight_gradient_mat())
# 计算对上一次输出h的权重梯度
for t in range(self.times, 0, -1):
# 计算各个时刻的梯度
(Wfh_grad, bf_grad,
Wih_grad, bi_grad,
Woh_grad, bo_grad,
Wch_grad, bc_grad) = (
self.calc_gradient_t(t))
# 实际梯度是各时刻梯度之和
self.Wfh_grad += Wfh_grad
self.bf_grad += bf_grad
self.Wih_grad += Wih_grad
self.bi_grad += bi_grad
self.Woh_grad += Woh_grad
self.bo_grad += bo_grad
self.Wch_grad += Wch_grad
self.bc_grad += bc_grad
# 计算对本次输入x的权重梯度
xt = x.transpose()
self.Wfx_grad = np.dot(self.delta_f_list[-1], xt)
self.Wix_grad = np.dot(self.delta_i_list[-1], xt)
self.Wox_grad = np.dot(self.delta_o_list[-1], xt)
self.Wcx_grad = np.dot(self.delta_ct_list[-1], xt)
def init_weight_gradient_mat(self):
'''
初始化权重矩阵
'''
Wh_grad = np.zeros((self.state_width,
self.state_width))
Wx_grad = np.zeros((self.state_width,
self.input_width))
b_grad = np.zeros((self.state_width, 1))
return Wh_grad, Wx_grad, b_grad
def calc_gradient_t(self, t):
'''
计算每个时刻t权重的梯度
'''
h_prev = self.h_list[t-1].transpose()
Wfh_grad = np.dot(self.delta_f_list[t], h_prev)
bf_grad = self.delta_f_list[t]
Wih_grad = np.dot(self.delta_i_list[t], h_prev)
bi_grad = self.delta_f_list[t]
Woh_grad = np.dot(self.delta_o_list[t], h_prev)
bo_grad = self.delta_f_list[t]
Wch_grad = np.dot(self.delta_ct_list[t], h_prev)
bc_grad = self.delta_ct_list[t]
return Wfh_grad, bf_grad, Wih_grad, bi_grad, \
Woh_grad, bo_grad, Wch_grad, bc_grad
def reset_state(self):
# 当前时刻初始化为t0
self.times = 0
# 各个时刻的单元状态向量c
self.c_list = self.init_state_vec()
# 各个时刻的输出向量h
self.h_list = self.init_state_vec()
# 各个时刻的遗忘门f
self.f_list = self.init_state_vec()
# 各个时刻的输入门i
self.i_list = self.init_state_vec()
# 各个时刻的输出门o
self.o_list = self.init_state_vec()
# 各个时刻的即时状态c~
self.ct_list = self.init_state_vec()
class LstmLayer(object):
def __init__(self, input_width, state_width, learning_rate):
self.input_width = input_width
self.state_width = state_width
self.learning_rate = learning_rate
self.gate_activator = SigmoidActivator()
self.output_activator = TanhActivator()
self.times = 0
self.c_list = self.init_state_vec()
self.h_list = self.init_state_vec()
self.f_list = self.init_state_vec()
self.i_list = self.init_state_vec()
self.o_list = self.init_state_vec()
self.ct_list = self.init_state_vec()
self.Wfh, self.Wfx, self.bf = (self.init_weight_mat())
self.Wih, self.Wix, self.bi = (self.init_weight_mat())
self.Woh, self.Wox, self.bo = (self.init_weight_mat())
self.Wch, self.Wcx, self.bc = (self.init_weight_mat())
def init_state_vec(self):
state_vec_list = []
state_vec_list.append(np.zeros((self.state_width, 1)))
return state_vec_list
def init_weight_mat(self):
Wh = np.random.uniform(-1e-4, 1e-4,
(self.state_width, self.state_width))
Wx = np.random.uniform(-1e-4, 1e-4,
(self.state_width, self.input_width))
b = np.zeros((self.state_width, 1))
return Wh, Wx, b
def forward(self, x):
self.times += 1
fg = self.calc_gate(x, self.Wfx, self.Wfh, self.bf,
self.gate_activator)
self.f_list.append(fg)
ig = self.calc_gate(x, self.Wix, self.Wih, self.bi,
self.gate_activator)
self.i_list.append(ig)
og = self.calc_gate(x, self.Wox, self.Woh, self.bo,
self.gate_activator)
self.o_list.append(og)
ct = self.calc_gate(x, self.Wcx, self.Wch, self.bc,
self.output_activator)
self.ct_list.append(ct)
c = fg * self.c_list[self.times - 1] + ig * ct
self.c_list.append(c)
h = og * self.output_activator.forward(c)
self.h_list.append(h)
def calc_gate(self, x, Wx, Wh, b, activator):
h = self.h_list[self.times - 1]
net = np.dot(Wh, h) + np.dot(Wx, x) + b
gate = activator.forward(net)
return gate
def backward(self, x, delta_h, activator):
self.calc_delta(delta_h, activator)
self.calc_gradient(x)
def update(self):
self.Wfh -= self.learning_rate * self.Whf_grad
self.Wfx -= self.learning_rate * self.Whx_grad
self.bf -= self.learning_rate * self.bf_grad
self.Wih -= self.learning_rate * self.Whi_grad
self.Wix -= self.learning_rate * self.Whi_grad
self.bi -= self.learning_rate * self.bi_grad
self.Woh -= self.learning_rate * self.Wof_grad
self.Wox -= self.learning_rate * self.Wox_grad
self.bo -= self.learning_rate * self.bo_grad
self.Wch -= self.learning_rate * self.Wcf_grad
self.Wcx -= self.learning_rate * self.Wcx_grad
self.bc -= self.learning_rate * self.bc_grad
def calc_delta(self, delta_h, activator):
self.delta_h_list = self.init_delta()
self.delta_o_list = self.init_delta()
self.delta_i_list = self.init_delta()
self.delta_f_list = self.init_delta()
self.delta_ct_list = self.init_delta()
self.delta_h_list[-1] = delta_h
for k in range(self.times, 0, -1):
self.calc_delta_k(k)
def init_delta(self):
delta_list = []
for i in range(self.times + 1):
delta_list.append(np.zeros((self.state_width, 1)))
return delta_list
def calc_delta_k(self, k):
ig = self.i_list[k]
og = self.o_list[k]
fg = self.f_list[k]
ct = self.ct_list[k]
c = self.c_list[k]
c_prev = self.c_list[k - 1]
tanh_c = self.output_activator.forward(c)
delta_k = self.delta_h_list[k]
delta_o = (delta_k * tanh_c * self.gate_activator.backward(og))
delta_f = (delta_k * og * (1 - tanh_c * tanh_c) * c_prev *
self.gate_activator.backward(fg))
delta_i = (delta_k * og * (1 - tanh_c * tanh_c) * ct *
self.gate_activator.backward(ig))
delta_ct = (delta_k * og * (1 - tanh_c * tanh_c) * ig *
self.output_activator.backward(ct))
delta_h_prev = (np.dot(delta_o.transpose(), self.Woh) +
np.dot(delta_i.transpose(), self.Wih) +
np.dot(delta_f.transpose(), self.Wfh) +
np.dot(delta_ct.transpose(), self.Wch)).transpose()
self.delta_h_list[k - 1] = delta_h_prev
self.delta_f_list[k] = delta_f
self.delta_i_list[k] = delta_i
self.delta_o_list[k] = delta_o
self.delta_ct_list[k] = delta_ct
def calc_gradient(self, x):
self.Wfh_grad, self.Wfx_grad, self.bf_grad = (
self.init_weight_gradient_mat())
self.Wih_grad, self.Wix_grad, self.bi_grad = (
self.init_weight_gradient_mat())
self.Woh_grad, self.Wox_grad, self.bo_grad = (
self.init_weight_gradient_mat())
self.Wch_grad, self.Wcx_grad, self.bc_grad = (
self.init_weight_gradient_mat())
for t in range(self.times, 0, -1):
(Wfh_grad, bf_grad, Wih_grad, bi_grad, Woh_grad, bo_grad, Wch_grad,
bc_grad) = (self.calc_gradient_t(t))
self.Wfh_grad += Wfh_grad
self.bf_grad += bf_grad
self.Wih_grad += Wih_grad
self.bi_grad += bi_grad
self.Woh_grad += Woh_grad
self.bo_grad += bo_grad
self.Wch_grad += Wch_grad
self.bc_grad += bc_grad
xt = x.transpose()
self.Wfx_grad = np.dot(self.delta_f_list[-1], xt)
self.Wix_grad = np.dot(self.delta_i_list[-1], xt)
self.Wox_grad = np.dot(self.delta_o_list[-1], xt)
self.Wcx_grad = np.dot(self.delta_ct_list[-1], xt)
def init_weight_gradient_mat(self):
Wh_grad = np.zeros((self.state_width, self.state_width))
Wx_grad = np.zeros((self.state_width, self.input_width))
b_grad = np.zeros((self.state_width, 1))
return Wh_grad, Wx_grad, b_grad
def calc_gradient_t(self, t):
h_prev = self.h_list[t - 1].transpose()
Wfh_grad = np.dot(self.delta_f_list[t], h_prev)
bf_grad = self.delta_f_list[t]
Wih_grad = np.dot(self.delta_i_list[t], h_prev)
bi_grad = self.delta_f_list[t]
Woh_grad = np.dot(self.delta_o_list[t], h_prev)
bo_grad = self.delta_f_list[t]
Wch_grad = np.dot(self.delta_ct_list[t], h_prev)
bc_grad = self.delta_ct_list[t]
return Wfh_grad, bf_grad, Wih_grad, bi_grad, \
Woh_grad, bo_grad, Wch_grad, bc_grad
def reset_state(self):
self.times = 0
self.c_list = self.init_state_vec()
self.h_list = self.init_state_vec()
self.f_list = self.init_state_vec()
self.i_list = self.init_state_vec()
self.o_list = self.init_state_vec()
self.ct_list = self.init_state_vec()
def __str__(self):
result = 'Wfh:\n%s\nWfx:\n%s\nbf:\n%s\n' % (self.Wfh, self.Wfx,
self.bf)
result += 'Wih:\n%s\nWix:\n%s\nbi:\n%s\n' % (self.Wih, self.Wix,
self.bi)
result += 'Woh:\n%s\nWox:\n%s\nbo:\n%s\n' % (self.Woh, self.Wox,
self.bo)
result += 'Wch:\n%s\nWcx:\n%s\nbc:\n%s\n' % (self.Wch, self.Wcx,
self.bc)
result += 'Wfh_grad:\n%s\nWfx_grad:\n%s\nbf_grad:\n%s\n' % (
self.Wfh_grad, self.Wfx_grad, self.bf_grad)
result += 'Wih_grad:\n%s\nWix_grad:\n%s\nbi_grad:\n%s\n' % (
self.Wih_grad, self.Wix_grad, self.bi_grad)
result += 'Woh_grad:\n%s\nWox_grad:\n%s\nbo_grad:\n%s\n' % (
self.Woh_grad, self.Wox_grad, self.bo_grad)
result += 'Wch_grad:\n%s\nWcx_grad:\n%s\nbc_grad:\n%s\n' % (
self.Wch_grad, self.Wcx_grad, self.bc_grad)
return result
class LstmLayer(object):
def __init__(self, input_width, state_width, learning_rate):
self.input_width = input_width
self.state_width = state_width
self.learning_rate = learning_rate
# 门的激活函数
self.gate_activator = SigmoidActivator()
# 输出的激活函数
self.output_activator = TanhActivator()
# 当前时刻初始化为t0
self.times = 0
# 各个时刻的单元状态向量c
self.c_list = self.init_state_vec()
# 各个时刻的输出向量h
self.h_list = self.init_state_vec()
# 各个时刻的遗忘门f
self.f_list = self.init_state_vec()
# 各个时刻的输入门i
self.i_list = self.init_state_vec()
# 各个时刻的输出门o
self.o_list = self.init_state_vec()
# 各个时刻的即时状态c~
self.ct_list = self.init_state_vec()
# 遗忘门权重矩阵Wfh, Wfx, 偏置项bf
self.Wfh, self.Wfx, self.bf = (self.init_weight_mat())
# 输入门权重矩阵Wih, Wix, 偏置项bi
self.Wih, self.Wix, self.bi = (self.init_weight_mat())
# 输出门权重矩阵Woh, Wox, 偏置项bo
self.Woh, self.Wox, self.bo = (self.init_weight_mat())
# 单元状态权重矩阵Wch, Wcx, 偏置项bc
self.Wch, self.Wcx, self.bc = (self.init_weight_mat())
## 初始化保存各类中间状态的向量
def init_state_vec(self):
'''
初始化保存状态的向量
'''
state_vec_list = []
state_vec_list.append(np.zeros((self.state_width, 1)))
return state_vec_list
## 初始化保存各类权重矩阵
def init_weight_mat(self):
'''
初始化权重矩阵
'''
Wh = np.random.uniform(-1e-4, 1e-4, (self.state_width, self.state_width))
Wx = np.random.uniform(-1e-4, 1e-4, (self.state_width, self.input_width))
b = np.zeros((self.state_width, 1))
return Wh, Wx, b
## forward方法实现了LSTM的前向计算
def forward(self, x):
'''
根据式1-式6进行前向计算
'''
self.times += 1
# 遗忘门
fg = self.calc_gate(x, self.Wfx, self.Wfh, self.bf, self.gate_activator)
self.f_list.append(fg)
# 输入门
ig = self.calc_gate(x, self.Wix, self.Wih, self.bi, self.gate_activator)
self.i_list.append(ig)
# 输出门
og = self.calc_gate(x, self.Wox, self.Woh, self.bo, self.gate_activator)
self.o_list.append(og)
# 即时状态
ct = self.calc_gate(x, self.Wcx, self.Wch, self.bc, self.output_activator)
self.ct_list.append(ct)
# 单元状态: Ct = Ft * Ct-1 + It * c~t
c = fg * self.c_list[self.times - 1] + ig * ct
self.c_list.append(c)
# 输出: Ht = Ot * TANH(Ct)
h = og * self.output_activator.forward(c)
self.h_list.append(h)
## 门的计算都是相同的算法，而门和c~t的计算仅仅是激活函数不同。因此提出了calc_gate方法，这样减少了很多重复代码。
def calc_gate(self, x, Wx, Wh, b, activator):
'''
计算门
'''
h = self.h_list[self.times - 1] # 上次的LSTM输出
net = np.dot(Wh, h) + np.dot(Wx, x) + b
gate = activator.forward(net)
return gate
## backward方法实现了LSTM的反向传播算法。需要注意的是，与backword相关的内部状态变量是在调用backward方法之后才初始化的。
## 这种延迟初始化的一个好处是，如果LSTM只是用来推理，那么就不需要初始化这些变量，节省了很多内存。
def backward(self, x, delta_h, activator):
'''
实现LSTM训练算法,主要包含两部分：
STEP 1: 计算误差项
STEP 2: 计算梯度
'''
self.calc_delta(delta_h, activator)
self.calc_gradient(x)
## 梯度下降算法来更新权重
def update(self):
'''
按照梯度下降，更新权重
'''
self.Wfh -= self.learning_rate * self.Whf_grad
self.Wfx -= self.learning_rate * self.Whx_grad
self.bf -= self.learning_rate * self.bf_grad
self.Wih -= self.learning_rate * self.Whi_grad
self.Wix -= self.learning_rate * self.Whi_grad
self.bi -= self.learning_rate * self.bi_grad
self.Woh -= self.learning_rate * self.Wof_grad
self.Wox -= self.learning_rate * self.Wox_grad
self.bo -= self.learning_rate * self.bo_grad
self.Wch -= self.learning_rate * self.Wcf_grad
self.Wcx -= self.learning_rate * self.Wcx_grad
self.bc -= self.learning_rate * self.bc_grad
##STEP 1: 计算误差项
def calc_delta(self, delta_h, activator):
# 初始化各个时刻的误差项
self.delta_h_list = self.init_delta() # 输出误差项
self.delta_o_list = self.init_delta() # 输出门误差项
self.delta_i_list = self.init_delta() # 输入门误差项
self.delta_f_list = self.init_delta() # 遗忘门误差项
self.delta_ct_list = self.init_delta() # 即时输出误差项
# 保存从上一层传递下来的当前时刻的误差项,将最后一项【重新赋值】为delta_h
self.delta_h_list[-1] = delta_h
# 迭代计算每个时刻的误差项
for k in range(self.times, 0, -1): # [times, times-1, ``` , 1]
self.calc_delta_k(k)
## 初始化各类误差项
def init_delta(self):
'''
初始化误差项，全部初始化为0
'''
delta_list = []
for i in range(self.times + 1):
delta_list.append(np.zeros((self.state_width, 1)))
return delta_list
## 计算k时刻的误差项
def calc_delta_k(self, k):
'''
根据k时刻的delta_h，计算k时刻的delta_f、delta_i、delta_o、delta_ct，以及k-1时刻的delta_h
'''
# 获得k时刻前向计算的值
ig = self.i_list[k]
og = self.o_list[k]
fg = self.f_list[k]
ct = self.ct_list[k]
c = self.c_list[k]
c_prev = self.c_list[k - 1]
tanh_c = self.output_activator.forward(c)
delta_k = self.delta_h_list[k]
# 根据【式9 - 式12】计算delta_o, delta_f, delta_i, delta_ct
delta_o = (delta_k * tanh_c * self.gate_activator.backward(og))
delta_f = (delta_k * og * (1 - tanh_c * tanh_c) * c_prev * self.gate_activator.backward(fg))
delta_i = (delta_k * og * (1 - tanh_c * tanh_c) * ct * self.gate_activator.backward(ig))
delta_ct = (delta_k * og * (1 - tanh_c * tanh_c) * ig * self.output_activator.backward(ct))
# 根据【式8】计算delta_h[k-1]
delta_h_prev = (
np.dot(delta_o.transpose(), self.Woh) +
np.dot(delta_i.transpose(), self.Wih) +
np.dot(delta_f.transpose(), self.Wfh) +
np.dot(delta_ct.transpose(), self.Wch)
).transpose()
# 保存全部delta值
self.delta_h_list[k - 1] = delta_h_prev
self.delta_f_list[k] = delta_f
self.delta_i_list[k] = delta_i
self.delta_o_list[k] = delta_o
self.delta_ct_list[k] = delta_ct
##STEP 2: 计算梯度
def calc_gradient(self, x):
# 初始化遗忘门权重梯度矩阵和偏置项
self.Wfh_grad, self.Wfx_grad, self.bf_grad = (self.init_weight_gradient_mat())
# 初始化输入门权重梯度矩阵和偏置项
self.Wih_grad, self.Wix_grad, self.bi_grad = (self.init_weight_gradient_mat())
# 初始化输出门权重梯度矩阵和偏置项
self.Woh_grad, self.Wox_grad, self.bo_grad = (self.init_weight_gradient_mat())
# 初始化单元状态权重梯度矩阵和偏置项
self.Wch_grad, self.Wcx_grad, self.bc_grad = (self.init_weight_gradient_mat())
# 计算对上一次输出h的权重梯度
for t in range(self.times, 0, -1):
# 计算各个时刻的梯度
(Wfh_grad, bf_grad,
Wih_grad, bi_grad,
Woh_grad, bo_grad,
Wch_grad, bc_grad) = (self.calc_gradient_t(t))
# 实际梯度是各时刻梯度之和
self.Wfh_grad += Wfh_grad
self.bf_grad += bf_grad
self.Wih_grad += Wih_grad
self.bi_grad += bi_grad
self.Woh_grad += Woh_grad
self.bo_grad += bo_grad
self.Wch_grad += Wch_grad
self.bc_grad += bc_grad
# 计算对本次输入x的权重梯度
xt = x.transpose()
self.Wfx_grad = np.dot(self.delta_f_list[-1], xt)
self.Wix_grad = np.dot(self.delta_i_list[-1], xt)
self.Wox_grad = np.dot(self.delta_o_list[-1], xt)
self.Wcx_grad = np.dot(self.delta_ct_list[-1], xt)
## 初始化权重梯度矩阵
def init_weight_gradient_mat(self):
'''
初始化权重梯度矩阵
'''
Wh_grad = np.zeros((self.state_width, self.state_width))
Wx_grad = np.zeros((self.state_width, self.input_width))
b_grad = np.zeros((self.state_width, 1))
return Wh_grad, Wx_grad, b_grad
## 计算每个时刻t权重的梯度
def calc_gradient_t(self, t):
'''
计算每个时刻t权重的梯度
'''
h_prev = self.h_list[t - 1].transpose()
Wfh_grad = np.dot(self.delta_f_list[t], h_prev)
bf_grad = self.delta_f_list[t]
Wih_grad = np.dot(self.delta_i_list[t], h_prev)
bi_grad = self.delta_f_list[t]
Woh_grad = np.dot(self.delta_o_list[t], h_prev)
bo_grad = self.delta_f_list[t]
Wch_grad = np.dot(self.delta_ct_list[t], h_prev)
bc_grad = self.delta_ct_list[t]
return Wfh_grad, bf_grad, Wih_grad, bi_grad, Woh_grad, bo_grad, Wch_grad, bc_grad
## 和RecurrentLayer一样，为了支持梯度检查，我们需要支持重置内部状态：
def reset_state(self):
# 当前时刻初始化为t0
self.times = 0
# 各个时刻的单元状态向量c
self.c_list = self.init_state_vec()
# 各个时刻的输出向量h
self.h_list = self.init_state_vec()
# 各个时刻的遗忘门f
self.f_list = self.init_state_vec()
# 各个时刻的输入门i
self.i_list = self.init_state_vec()
# 各个时刻的输出门o
self.o_list = self.init_state_vec()
# 各个时刻的即时状态c~
self.ct_list = self.init_state_vec()
class LstmLayer():
def __init__(self, input_width, state_width, output_width,
learning_rate, penaltyL2, momentum):
self.input_width = input_width
self.state_width = state_width
self.output_width= output_width
self.learning_rate = learning_rate
self.penaltyL2 = penaltyL2
self.momentum = momentum
# 门的激活函数
self.gate_activator = SigmoidActivator()
# 输出的激活函数
self.output_activator = TanhActivator()
self.class_activator = SoftmaxActivator()
# 遗忘门权重矩阵Wfh, Wfx, 偏置项bf
self.Wfh, self.Wfx, self.bf, self.vWfh, self.vWfx, self.vbf =(self.init_weight_mat(0))
# 输入门权重矩阵Wfh, Wfx, 偏置项bf
self.Wih, self.Wix, self.bi, self.vWih, self.vWix, self.vbi =(self.init_weight_mat(0))
# 输出门权重矩阵Wfh, Wfx, 偏置项bf
self.Woh, self.Wox, self.bo, self.vWoh, self.vWox, self.vbo =(self.init_weight_mat(0))
# 单元状态权重矩阵Wfh, Wfx, 偏置项bf
self.Wch, self.Wcx, self.bc, self.vWch, self.vWcx, self.vbc =(self.init_weight_mat(0))
# 下一层权重Wy, 偏值by
self.Wy, self.by, self.vWy, self.vby =(self.init_weight_mat(1))
def init_weight_mat(self,i):
'''
初始化权重矩阵
'''
if (i<1):
Wh = np.mat(np.random.uniform(-0.5, 0.5,
(self.state_width, self.state_width)))/self.state_width
vWh = np.mat(np.zeros(Wh.shape))
Wx = np.mat(np.random.uniform(-0.5, 0.5,
(self.state_width, self.input_width)))/self.input_width
vWx = np.mat(np.zeros(Wx.shape))
b = np.mat(np.random.uniform(-0.5,0.5,(self.state_width, 1)))/self.state_width
vb = np.mat(np.zeros(b.shape))
return Wh, Wx, b, vWh, vWx, vb
else:
Wy = np.mat(np.random.uniform(-0.5, 0.5,
(self.output_width, self.state_width)))/self.output_width
vWy = np.mat(np.zeros(Wy.shape))
by = np.mat(np.random.uniform(-0.5,0.5,(self.output_width, 1)))/self.output_width
vby = np.mat(np.zeros(by.shape))
return Wy, by, vWy, vby
def forward(self, x):
'''
根据式1-式6进行前向计算
'''
self.x = x
self.reset_state()
n = x.shape[0]
for i in range(n):
# 遗忘门
fg = self.calc_gate(x[i], self.Wfx, self.Wfh,
self.bf, self.gate_activator,i)
self.f_list[i] = fg
# 输入门
ig = self.calc_gate(x[i], self.Wix, self.Wih,
self.bi, self.gate_activator,i)
self.i_list[i] = ig
# 输出门
og = self.calc_gate(x[i], self.Wox, self.Woh,
self.bo, self.gate_activator,i)
self.o_list[i] = og
# 即时状态
ct = self.calc_gate(x[i], self.Wcx, self.Wch,
self.bc, self.output_activator,i)
self.ct_list[i] = ct
# 单元状态
if i==0:
c = np.multiply(ig, ct)
else:
c = np.multiply(fg, self.c_list[i - 1]) + np.multiply(ig, ct)
self.c_list[i] = c
# 输出
h = np.multiply(og, self.output_activator.forward(c))
self.h_list[i] = h
y = self.class_activator.forward(self.Wy * h.T + self.by)
self.y_list[i] = y.T
def reset_state(self):
# 各个时刻的单元状态向量c
self.c_list = np.mat(np.zeros((self.x.shape[0], self.state_width)))
# 各个时刻的输出向量h
self.h_list = np.mat(np.zeros((self.x.shape[0], self.state_width)))
# 各个时刻的遗忘门f
self.f_list = np.mat(np.zeros((self.x.shape[0], self.state_width)))
# 各个时刻的输入门i
self.i_list = np.mat(np.zeros((self.x.shape[0], self.state_width)))
# 各个时刻的输出门o
self.o_list = np.mat(np.zeros((self.x.shape[0], self.state_width)))
# 各个时刻的即时状态c~
self.ct_list = np.mat(np.zeros((self.x.shape[0], self.state_width)))
self.y_list = np.mat(np.zeros((self.x.shape[0], self.output_width)))
def calc_gate(self, x, Wx, Wh, b, activator,i):
'''
计算门
'''
if i==0:
h = np.mat(np.zeros((1, self.state_width)))
else:
h = self.h_list[i-1] # 上次的LSTM输出
net = (Wh * h.T + Wx * x.T + b).T
gate = activator.forward(net)
return gate
def backward(self, e):
'''
实现LSTM训练算法
'''
self.e = e
self.calc_delta()
self.calc_gradient()
self.update()
def update(self):
'''
按照梯度下降，更新权重
'''
self.vWfh = self.momentum * self.vWfh - self.learning_rate * \
(self.Wfh_grad + self.penaltyL2 * self.Wfh)
self.vWfx = self.momentum * self.vWfx - self.learning_rate * \
(self.Wfx_grad + self.penaltyL2 * \
np.concatenate((np.mat(np.zeros((self.Wfx.shape[0],1))),self.Wfx[:,1:]),axis=1))
self.vbf = self.momentum * self.vbf - self.learning_rate * self.bf_grad
self.vWih = self.momentum * self.vWih - self.learning_rate * \
(self.Wih_grad + self.penaltyL2 * self.Wih)
self.vWix = self.momentum * self.vWix - self.learning_rate * \
(self.Wix_grad + self.penaltyL2 * \
np.concatenate((np.mat(np.zeros((self.Wix.shape[0],1))),self.Wix[:,1:]),axis=1))
self.vbi = self.momentum * self.vbi - self.learning_rate * self.bi_grad
self.vWoh = self.momentum * self.vWoh - self.learning_rate * \
(self.Woh_grad + self.penaltyL2 * self.Woh)
self.vWox = self.momentum * self.vWox - self.learning_rate * \
(self.Wox_grad + self.penaltyL2 * \
np.concatenate((np.mat(np.zeros((self.Wox.shape[0],1))),self.Wox[:,1:]),axis=1))
self.vbo = self.momentum * self.vbo - self.learning_rate * self.bo_grad
self.vWch = self.momentum * self.vWch - self.learning_rate * \
(self.Wch_grad + self.penaltyL2 * self.Wch)
self.vWcx = self.momentum * self.vWcx - self.learning_rate * \
(self.Wcx_grad + self.penaltyL2 * \
np.concatenate((np.mat(np.zeros((self.Wcx.shape[0],1))),self.Wcx[:,1:]),axis=1))
self.vbc = self.momentum * self.vbc - self.learning_rate * self.bc_grad
self.vWy = self.momentum * self.vWy - self.learning_rate * \
(self.Wy_grad + self.penaltyL2 * self.Wy)
self.vby = self.momentum * self.vby - self.learning_rate * self.by_grad
self.Wfh += self.vWfh
self.Wfx += self.vWfx
self.bf += self.vbf
self.Wih += self.vWih
self.Wix += self.vWix
self.bi += self.vbi
self.Woh += self.vWoh
self.Wox += self.vWox
self.bo += self.vbo
self.Wch += self.vWch
self.Wcx += self.vWcx
self.bc += self.vbc
self.Wy += self.vWy
self.by += self.vby
def calc_delta(self):
# 初始化各个时刻的误差项
self.delta_h_list = self.init_delta() # 输出误差项
self.delta_o_list = self.init_delta() # 输出门误差项
self.delta_i_list = self.init_delta() # 输入门误差项
self.delta_f_list = self.init_delta() # 遗忘门误差项
self.delta_ct_list = self.init_delta() # 即时输出误差项
self.delta_c_list = self.init_delta() #state c
self.delta_h_list[-1] = self.e[-1] * self.Wy
a = self.output_activator.backward(self.output_activator.forward(self.c_list[-1]))
self.delta_c_list[-1] = np.multiply(self.delta_h_list[-1], self.o_list[-1], a)
m = self.e.shape[0]
for k in range(m-1, 0, -1):
self.calc_delta_k(k)
def init_delta(self):
'''
初始化误差项
'''
delta_list = np.mat(np.zeros((self.e.shape[0],self.state_width)))
return delta_list
def calc_delta_k(self, k):
'''
根据k时刻的delta_h，计算k时刻的delta_f、
delta_i、delta_o、delta_ct，以及k-1时刻的delta_h
'''
# 获得k时刻前向计算的值
ig = self.i_list[k]
og = self.o_list[k]
fg = self.f_list[k]
ct = self.ct_list[k]
c = self.c_list[k]
c_prev = self.c_list[k-1]
tan_c = self.output_activator.forward(c)
delta_h = self.delta_h_list[k]
delta_c = self.delta_c_list[k]
delta_y = self.e[k-1]
# 根据式9计算delta_o
gate_o = np.multiply(tan_c, self.gate_activator.backward(og))
delta_o = np.multiply(delta_h, gate_o)
gate_f = np.multiply(c_prev, self.gate_activator.backward(fg))
delta_f = np.multiply(delta_c, gate_f)
gate_i = np.multiply(ct, self.gate_activator.backward(ig))
delta_i = np.multiply(delta_c, gate_i)
gate_c = np.multiply(ig, self.output_activator.backward(ct))
delta_ct = np.multiply(delta_c, gate_c)
delc = np.multiply(og, self.output_activator.backward(tan_c))
delta_h_prev = np.multiply(delta_h, (gate_o * self.Woh + \
np.multiply(gate_i, delc) * self.Wih + \
np.multiply(gate_f, delc) * self.Wfh + \
np.multiply(gate_c, delc) * self.Wch)) + delta_y * self.Wy
delc1 = np.multiply(self.o_list[k-1],self.output_activator.backward(self.output_activator.forward(c_prev)))
delta_c_prev = np.multiply(delta_c, fg) + np.multiply(delta_h_prev, delc1)
# 保存全部delta值
self.delta_h_list[k-1] = delta_h_prev
self.delta_c_list[k-1] = delta_c_prev
self.delta_f_list[k] = delta_f
self.delta_i_list[k] = delta_i
self.delta_o_list[k] = delta_o
self.delta_ct_list[k] = delta_ct
def calc_gradient(self):
# 初始化遗忘门权重梯度矩阵和偏置项
Wfh_grad, Wfx_grad, bf_grad = (
self.init_weight_gradient_mat(0))
# 初始化输入门权重梯度矩阵和偏置项
Wih_grad, Wix_grad, bi_grad = (
self.init_weight_gradient_mat(0))
# 初始化输出门权重梯度矩阵和偏置项
Woh_grad, Wox_grad, bo_grad = (
self.init_weight_gradient_mat(0))
# 初始化单元状态权重梯度矩阵和偏置项
Wch_grad, Wcx_grad, bc_grad = (
self.init_weight_gradient_mat(0))
Wy_grad, by_grad = (self.init_weight_gradient_mat(1))
m = self.e.shape[0]
# 计算对上一次输出h的权重梯度
for t in range(m-1, 0, -1):
h = self.h_list[t]
h_prev = self.h_list[t - 1]
x = self.x[t]
Wfh_grad += self.delta_f_list[t].T * h_prev
Wfx_grad += self.delta_f_list[t].T * x
bf_grad += self.delta_f_list[t].T
Wih_grad += self.delta_i_list[t].T * h_prev
Wix_grad += self.delta_i_list[t].T * x
bi_grad += self.delta_i_list[t].T
Woh_grad += self.delta_o_list[t].T * h_prev
Wox_grad += self.delta_o_list[t].T * x
bo_grad += self.delta_o_list[t].T
Wch_grad += self.delta_ct_list[t].T * h_prev
Wcx_grad += self.delta_ct_list[t].T * x
bc_grad += self.delta_ct_list[t].T
Wy_grad += self.e[t].T * h
by_grad += self.e[t].T
self.Wfh_grad = Wfh_grad/(m-1)
self.Wfx_grad = Wfx_grad/m
self.bf_grad = bf_grad/m
self.Wih_grad = Wih_grad/(m-1)
self.Wix_grad = Wix_grad/m
self.bi_grad = bi_grad/m
self.Woh_grad = Woh_grad/(m-1)
self.Wox_grad = Wox_grad/m
self.bo_grad = bo_grad/m
self.Wch_grad = Wch_grad/(m-1)
self.Wcx_grad = Wcx_grad/m
self.bc_grad = bc_grad/m
self.Wy_grad = Wy_grad/m
self.by_grad = by_grad/m
def init_weight_gradient_mat(self,i):
'''
初始化权重矩阵
'''
if i<1:
Wh_grad = np.mat(np.zeros((self.state_width, self.state_width)))
Wx_grad = np.mat(np.zeros((self.state_width, self.input_width)))
b_grad = np.mat(np.zeros((self.state_width, 1)))
return Wh_grad, Wx_grad, b_grad
else:
Wy_grad = np.mat(np.zeros((self.output_width, self.state_width)))
by_grad = np.mat(np.zeros((self.output_width, 1)))
return Wy_grad, by_grad