def grad(self, params, x, y):
w, alpha = self.inflate(params)
d_prev = self.decide(0., x[0], w, alpha)
d_prev_by_w = sech2(dot(w, x[0])) * x[0]
r_prev_by_w = d_prev_by_w * y[0] - self.delta * sign(
d_prev) * d_prev_by_w
A_prev_by_w = r_prev_by_w
d_prev_by_alpha = r_prev_by_alpha = A_prev_by_alpha = 0.
for i in range(1, len(y)):
x_t = x[i]
q_t = y[i]
d_t = self.decide(d_prev, x_t, w, alpha)
d_i = dot(w, x_t) + alpha * d_prev
d_t_by_w = sech2(d_i) * (x_t + alpha * d_prev_by_w)
r_t_by_w = d_t_by_w * q_t - self.delta * sign(d_t - d_prev) * (
d_t_by_w - d_prev_by_w)
A_t_by_w = A_prev_by_w + (1. / (i + 1)) * (r_t_by_w - A_prev_by_w)
if i == 1:
d_t_by_alpha = sech2(d_i) * d_prev
else:
d_t_by_alpha = sech2(d_i) * (alpha * d_prev_by_alpha + d_prev)
r_t_by_alpha = d_t_by_alpha * q_t - (
self.delta * sign(d_t - d_prev) *
(d_t_by_alpha - d_prev_by_alpha))
A_t_by_alpha = A_prev_by_alpha + (1. / (i + 1)) * (r_t_by_alpha -
A_prev_by_alpha)
d_prev = d_t
d_prev_by_w = d_t_by_w
r_prev_by_w = r_t_by_w
A_prev_by_w = A_t_by_w
d_prev_by_alpha = d_t_by_alpha
r_prev_by_alpha = r_t_by_alpha
A_prev_by_alpha = A_t_by_alpha
return -1 * append(A_t_by_w, A_t_by_alpha) + 2 * self.lmb * params
def grad(self, params, x, y):
w, alpha = self.inflate(params)
d_prev = self.decide(0., x[0], w, alpha)
d_prev_by_w = sech2(dot(w, x[0])) * x[0]
r_prev_by_w = d_prev_by_w*y[0] - self.delta*sign(d_prev)*d_prev_by_w
A_prev_by_w = r_prev_by_w
d_prev_by_alpha = r_prev_by_alpha = A_prev_by_alpha = 0.
for i in range(1, len(y)):
x_t = x[i]
q_t = y[i]
d_t = self.decide(d_prev, x_t, w, alpha)
d_i = dot(w, x_t) + alpha*d_prev
d_t_by_w = sech2(d_i) * (x_t + alpha*d_prev_by_w)
r_t_by_w = d_t_by_w*q_t - self.delta*sign(d_t - d_prev)*(d_t_by_w - d_prev_by_w)
A_t_by_w = A_prev_by_w + (1./(i+1))*(r_t_by_w - A_prev_by_w)
if i == 1:
d_t_by_alpha = sech2(d_i) * d_prev
else:
d_t_by_alpha = sech2(d_i) * (alpha * d_prev_by_alpha + d_prev)
r_t_by_alpha = d_t_by_alpha*q_t - (
self.delta*sign(d_t - d_prev)*(d_t_by_alpha - d_prev_by_alpha))
A_t_by_alpha = A_prev_by_alpha + (1./(i+1))*(r_t_by_alpha - A_prev_by_alpha)
d_prev = d_t
d_prev_by_w = d_t_by_w
r_prev_by_w = r_t_by_w
A_prev_by_w = A_t_by_w
d_prev_by_alpha = d_t_by_alpha
r_prev_by_alpha = r_t_by_alpha
A_prev_by_alpha = A_t_by_alpha
return -1 * append(A_t_by_w, A_t_by_alpha) + 2*self.lmb*params
def grad(self, params, x, y):
W, w, alpha = self.inflate(params)
x_t = x[0]
q_t = y[0]
h_t = tanh(dot(W, x_t))
d_prev = self.decide(0., x_t, W, w, alpha)
d_prev_by_w = sech2(dot(w, h_t)) * h_t
r_prev_by_w = d_prev_by_w * y[0] - self.delta * sign(
d_prev) * d_prev_by_w
A_prev_by_w = r_prev_by_w
h_prev_by_W = zeros((w.size, W.shape[0], W.shape[1]))
for i in range(self.hidden):
h_prev_by_W[i, :, :] = sech2(dot(W[i, :], x_t)) * x_t
d_prev_by_W = zeros(W.shape) + sech2(dot(w, h_t))
for i in range(self.hidden):
d_prev_by_W[i, :] *= w[i] * h_prev_by_W[i, i, :]
r_prev_by_W = d_prev_by_W * y[0] - self.delta * sign(
d_prev) * d_prev_by_W
A_prev_by_W = r_prev_by_W
d_prev_by_alpha = zeros(alpha.shape)
r_prev_by_alpha = zeros(alpha.shape)
A_prev_by_alpha = zeros(alpha.shape)
for step in range(1, len(y)):
x_t = x[step]
q_t = y[step]
h_t = tanh(dot(W, x_t) + alpha * d_prev)
d_t = self.decide(d_prev, x_t, W, w, alpha)
h_t_by_w = dot(outer(d_prev_by_w, alpha),
diag(sech2(dot(W, x_t) + alpha * d_prev)))
d_t_by_w = sech2(dot(w, h_t)) * (dot(h_t_by_w, w) + h_t)
r_t_by_w = d_t_by_w * q_t - self.delta * sign(d_t - d_prev) * (
d_t_by_w - d_prev_by_w)
A_t_by_w = A_prev_by_w + (1. /
(step + 1)) * (r_t_by_w - A_prev_by_w)
h_t_by_W = zeros((w.size, W.shape[0], W.shape[1]))
for k in range(w.size):
a = sech2(dot(W[k, :], x_t) + alpha[k] * d_prev)
for i in range(W.shape[0]):
for j in range(W.shape[1]):
b = alpha[k] * d_prev_by_W[i][j]
if k == i:
b += x_t[j]
h_t_by_W[k][i][j] = a * b
d_t_by_W = zeros(W.shape)
for i in range(self.hidden):
for j in range(self.exlen):
d_t_by_W[i][j] = sech2(dot(w, h_t)) * dot(
w, h_t_by_W[:, i, j])
r_t_by_W = d_t_by_W * q_t - self.delta * sign(d_t - d_prev) * (
d_t_by_W - d_prev_by_W)
A_t_by_W = A_prev_by_W + (1. /
(step + 1)) * (r_t_by_W - A_prev_by_W)
a = sech2(dot(W, x_t) + alpha * d_prev)
b = outer(d_prev_by_alpha, alpha) + eye(alpha.size) * d_prev
h_t_by_alpha = dot(b, diag(a))
d_t_by_alpha = sech2(dot(w, h_t)) * dot(h_t_by_alpha, w)
r_t_by_alpha = d_t_by_alpha * q_t - self.delta * sign(
d_t - d_prev) * (d_t_by_alpha - d_prev_by_alpha)
A_t_by_alpha = A_prev_by_alpha + (1. /
(step + 1)) * (r_t_by_alpha -
A_prev_by_alpha)
d_prev = d_t
d_prev_by_w = d_t_by_w
r_prev_by_w = r_t_by_w
A_prev_by_w = A_t_by_w
d_prev_by_W = d_t_by_W
r_prev_by_W = r_t_by_W
A_prev_by_W = A_t_by_W
d_prev_by_alpha = d_t_by_alpha
r_prev_by_alpha = r_t_by_alpha
A_prev_by_alpha = A_t_by_alpha
return -1 * concatenate((A_t_by_W.reshape((A_t_by_W.size, )), A_t_by_w,
A_t_by_alpha)) + 2 * self.lmb * params
def grad(self, params, x, y):
W, w, alpha = self.inflate(params)
x_t = x[0]
q_t = y[0]
h_t = tanh(dot(W, x_t))
d_prev = self.decide(0., x_t, W, w, alpha)
d_prev_by_w = sech2(dot(w, h_t)) * h_t
r_prev_by_w = d_prev_by_w*y[0] - self.delta*sign(d_prev)*d_prev_by_w
A_prev_by_w = r_prev_by_w
h_prev_by_W = zeros((w.size, W.shape[0], W.shape[1]))
for i in range(self.hidden):
h_prev_by_W[i,:,:] = sech2(dot(W[i,:], x_t)) * x_t
d_prev_by_W = zeros(W.shape) + sech2(dot(w, h_t))
for i in range(self.hidden):
d_prev_by_W[i,:] *= w[i]*h_prev_by_W[i,i,:]
r_prev_by_W = d_prev_by_W*y[0] - self.delta*sign(d_prev)*d_prev_by_W
A_prev_by_W = r_prev_by_W
d_prev_by_alpha = zeros(alpha.shape)
r_prev_by_alpha = zeros(alpha.shape)
A_prev_by_alpha = zeros(alpha.shape)
for step in range(1, len(y)):
x_t = x[step]
q_t = y[step]
h_t = tanh(dot(W, x_t) + alpha*d_prev)
d_t = self.decide(d_prev, x_t, W, w, alpha)
h_t_by_w = dot(outer(d_prev_by_w, alpha), diag(sech2(dot(W, x_t) + alpha*d_prev)))
d_t_by_w = sech2(dot(w, h_t)) * (dot(h_t_by_w, w) + h_t)
r_t_by_w = d_t_by_w*q_t - self.delta*sign(d_t - d_prev)*(d_t_by_w - d_prev_by_w)
A_t_by_w = A_prev_by_w + (1./(step+1))*(r_t_by_w - A_prev_by_w)
h_t_by_W = zeros((w.size, W.shape[0], W.shape[1]))
for k in range(w.size):
a = sech2(dot(W[k,:], x_t) + alpha[k]*d_prev)
for i in range(W.shape[0]):
for j in range(W.shape[1]):
b = alpha[k]*d_prev_by_W[i][j]
if k == i:
b += x_t[j]
h_t_by_W[k][i][j] = a*b
d_t_by_W = zeros(W.shape)
for i in range(self.hidden):
for j in range(self.exlen):
d_t_by_W[i][j] = sech2(dot(w, h_t)) * dot(w, h_t_by_W[:,i,j])
r_t_by_W = d_t_by_W*q_t - self.delta*sign(d_t - d_prev)*(d_t_by_W - d_prev_by_W)
A_t_by_W = A_prev_by_W + (1./(step+1))*(r_t_by_W - A_prev_by_W)
a = sech2(dot(W, x_t) + alpha*d_prev)
b = outer(d_prev_by_alpha, alpha) + eye(alpha.size)*d_prev
h_t_by_alpha = dot(b, diag(a))
d_t_by_alpha = sech2(dot(w, h_t)) * dot(h_t_by_alpha, w)
r_t_by_alpha = d_t_by_alpha*q_t - self.delta*sign(d_t - d_prev)*(d_t_by_alpha -
d_prev_by_alpha)
A_t_by_alpha = A_prev_by_alpha + (1./(step+1))*(r_t_by_alpha - A_prev_by_alpha)
d_prev = d_t
d_prev_by_w = d_t_by_w
r_prev_by_w = r_t_by_w
A_prev_by_w = A_t_by_w
d_prev_by_W = d_t_by_W
r_prev_by_W = r_t_by_W
A_prev_by_W = A_t_by_W
d_prev_by_alpha = d_t_by_alpha
r_prev_by_alpha = r_t_by_alpha
A_prev_by_alpha = A_t_by_alpha
return -1 * concatenate((A_t_by_W.reshape((A_t_by_W.size,)), A_t_by_w,
A_t_by_alpha)) + 2*self.lmb*params
