def fprop(self, x, h):
self._tmp_x = x
self._tmp_h_tm1 = h
x_stack = ca.dot(self.w_x.array.T, x.T)
h_stack = ca.dot(self.w_h.array.T, h.T)
n = self.n_hidden
x_r = x_stack[:n, :]
x_u = x_stack[n : n * 2, :]
x_c = x_stack[n * 2 : n * 3, :]
h_r = h_stack[:n, :]
h_u = h_stack[n : n * 2, :]
h_c = h_stack[n * 2 : n * 3, :]
r = self.act_r.fprop(x_r + h_r + self.b_r.array)
u = self.act_u.fprop(x_u + h_u + self.b_u.array)
c = self.act_c.fprop(x_c + r * h_c + self.b_c.array)
u = ca.ascontiguousarray(ca.transpose(u))
c = ca.ascontiguousarray(ca.transpose(c))
h_tp1 = 1 - u
h_tp1 *= h
h_tp1 += u * c
self._tmp_r = r
self._tmp_u = u
self._tmp_c = c
self._tmp_h_c = h_c
return {"y": h_tp1, "h": h_tp1}
def bprop(self, y_grad, h_grad):
n = self.n_hidden
h_grad = h_grad + y_grad
c_grad = h_grad * self._tmp_u
u_grad = h_grad * (self._tmp_c - self._tmp_h_tm1)
h_grad *= 1 - self._tmp_u
c_grad = ca.ascontiguousarray(ca.transpose(c_grad))
u_grad = ca.ascontiguousarray(ca.transpose(u_grad))
c_grad = self.act_c.bprop(c_grad)
ca.sum(c_grad, axis=1, keepdims=True, out=self.b_c.grad_array)
u_grad = self.act_u.bprop(u_grad)
ca.sum(u_grad, axis=1, keepdims=True, out=self.b_u.grad_array)
r_grad = c_grad * self._tmp_h_c
r_grad = self.act_r.bprop(r_grad)
ca.sum(r_grad, axis=1, keepdims=True, out=self.b_r.grad_array)
stack_grad = ca.empty((self.n_hidden * 3, y_grad.shape[0]))
stack_grad[:n, :] = r_grad
stack_grad[n : n * 2, :] = u_grad
stack_grad[n * 2 : n * 3, :] = c_grad
ca.dot(self._tmp_x.T, stack_grad.T, out=self.w_x.grad_array)
x_grad = ca.dot(stack_grad.T, self.w_x.array.T)
stack_grad[n * 2 : n * 3, :] *= self._tmp_r
ca.dot(self._tmp_h_tm1.T, stack_grad.T, out=self.w_h.grad_array)
h_grad += ca.dot(stack_grad.T, self.w_h.array.T)
ca.clip(h_grad, -self.clip, self.clip, out=h_grad)
return {"x_grad": x_grad, "h_grad": h_grad}
def bprop(self, y_grad, h_grad):
n = self.n_hidden
h_grad = h_grad + y_grad
c_grad = h_grad * self._tmp_u
u_grad = h_grad * (self._tmp_c - self._tmp_h_tm1)
h_grad *= (1 - self._tmp_u)
c_grad = ca.ascontiguousarray(ca.transpose(c_grad))
u_grad = ca.ascontiguousarray(ca.transpose(u_grad))
c_grad = self.act_c.bprop(c_grad)
ca.sum(c_grad, axis=1, keepdims=True, out=self.b_c.grad_array)
u_grad = self.act_u.bprop(u_grad)
ca.sum(u_grad, axis=1, keepdims=True, out=self.b_u.grad_array)
r_grad = c_grad * self._tmp_h_c
r_grad = self.act_r.bprop(r_grad)
ca.sum(r_grad, axis=1, keepdims=True, out=self.b_r.grad_array)
stack_grad = ca.empty((self.n_hidden*3, y_grad.shape[0]))
stack_grad[:n, :] = r_grad
stack_grad[n:n*2, :] = u_grad
stack_grad[n*2:n*3, :] = c_grad
ca.dot(self._tmp_x.T, stack_grad.T, out=self.w_x.grad_array)
x_grad = ca.dot(stack_grad.T, self.w_x.array.T)
stack_grad[n*2:n*3, :] *= self._tmp_r
ca.dot(self._tmp_h_tm1.T, stack_grad.T, out=self.w_h.grad_array)
h_grad += ca.dot(stack_grad.T, self.w_h.array.T)
ca.clip(h_grad, -self.clip, self.clip, out=h_grad)
return {'x_grad': x_grad, 'h_grad': h_grad}
def fprop(self, x, h):
self._tmp_x = x
self._tmp_h_tm1 = h
x_stack = ca.dot(self.w_x.array.T, x.T)
h_stack = ca.dot(self.w_h.array.T, h.T)
n = self.n_hidden
x_r = x_stack[:n, :]
x_u = x_stack[n:n*2, :]
x_c = x_stack[n*2:n*3, :]
h_r = h_stack[:n, :]
h_u = h_stack[n:n*2, :]
h_c = h_stack[n*2:n*3, :]
r = self.act_r.fprop(x_r + h_r + self.b_r.array)
u = self.act_u.fprop(x_u + h_u + self.b_u.array)
c = self.act_c.fprop(x_c + r*h_c + self.b_c.array)
u = ca.ascontiguousarray(ca.transpose(u))
c = ca.ascontiguousarray(ca.transpose(c))
h_tp1 = 1-u
h_tp1 *= h
h_tp1 += u*c
self._tmp_r = r
self._tmp_u = u
self._tmp_c = c
self._tmp_h_c = h_c
return {'y': h_tp1, 'h': h_tp1}
def matrix_factorization(R, P, Q, mask, steps=200000000, alpha=0.00005, beta=0.02):
Q = ca.transpose(Q)
for step in xrange(steps):
E = ca.subtract(R, ca.multiply(ca.dot(P,Q), mask))
rmse = ca.sqrt(ca.sum(ca.power(E,2)) / ca.sum(mask))
rmse = np.array(rmse)[0]
print 'step: %i RMSE: %f' % (step, rmse)
if rmse < 0.65:
break
P = ca.add(ca.multiply(P,(1-alpha*beta)),ca.multiply(ca.dot(E,ca.transpose(Q)), 2*alpha))
Q = ca.add(ca.multiply(Q,(1-alpha*beta)),ca.multiply(ca.dot(ca.transpose(P),E),2*alpha))
return P, Q
def getSimpleDot(test,train,gpuFlag=False,normalize=True):
if normalize:
test = util.normalize(test,gpuFlag=gpuFlag);
train = util.normalize(train,gpuFlag=gpuFlag);
if gpuFlag:
distances=ca.dot(test,ca.transpose(train));
distances=np.array(distances);
else:
distances=np.dot(test,train.T);
return distances
def matrix_factorization(R,
P,
Q,
mask,
steps=200000000,
alpha=0.00005,
beta=0.02):
Q = ca.transpose(Q)
for step in xrange(steps):
E = ca.subtract(R, ca.multiply(ca.dot(P, Q), mask))
rmse = ca.sqrt(ca.sum(ca.power(E, 2)) / ca.sum(mask))
rmse = np.array(rmse)[0]
print 'step: %i RMSE: %f' % (step, rmse)
if rmse < 0.65:
break
P = ca.add(ca.multiply(P, (1 - alpha * beta)),
ca.multiply(ca.dot(E, ca.transpose(Q)), 2 * alpha))
Q = ca.add(ca.multiply(Q, (1 - alpha * beta)),
ca.multiply(ca.dot(ca.transpose(P), E), 2 * alpha))
return P, Q
def bprop(self):
self.x.out_grad = ca.transpose(self.out_grad)
if self.contiguous:
self.out = ca.ascontiguousarray(self.x.out_grad)
def fprop(self):
self.out = ca.transpose(self.x.out)
if self.contiguous:
self.out = ca.ascontiguousarray(self.out)
def bprop(self):
self.x.grad_array = ca.transpose(self.grad_array)
if self.contiguous:
self.array = ca.ascontiguousarray(self.x.grad_array)
def bprop(self):
self.x.out_grad = ca.transpose(self.out_grad)
if self.contiguous:
self.out = ca.ascontiguousarray(self.x.out_grad)
def fprop(self):
self.out = ca.transpose(self.x.out)
if self.contiguous:
self.out = ca.ascontiguousarray(self.out)
