def f_df(self, theta, x):
"""
Compute objective and gradient of the CAE objective using the
examples {\bf x}.
Parameters
----------
x: array-like, shape (n_examples, n_inputs)
Parameters
----------
loss: array-like, shape (n_examples,)
Value of the loss function for each example before the step.
"""
self.set_params(theta)
h = self.encode(x)
r = self.decode(h)
def _contraction_jacobian():
"""
Compute the gradient of the contraction cost w.r.t parameters.
"""
a = 2*(h * (1 - h))**2
d = ((1 - 2 * h) * a * (self.W**2).sum(0)[None, :])
b = np.dot(x.T / x.shape[0], d)
c = a.mean(0) * self.W
return (b + c), d.mean(0)
def _reconstruction_jacobian():
"""
Compute the gradient of the reconstruction cost w.r.t parameters.
"""
dr = (r - x) / x.shape[0]
dr *= r * (1-r)
dd = np.dot(dr.T, h)
dh = np.dot(dr, self.W) * h * (1. - h)
de = np.dot(x.T, dh)
return (dd + de), dr.sum(0), dh.sum(0)
W_rec, c_rec, b_rec = _reconstruction_jacobian()
W_con, b_con = _contraction_jacobian()
dW = W_rec + self.jacobi_penalty * W_con
dhidbias = b_rec + self.jacobi_penalty * b_con
dvisbias = c_rec
return self.loss(x, h, r), vector_from_args([dW, dhidbias, dvisbias]);
def get_params(self):
return vector_from_args(self.params)
def get_params(self):
return vector_from_args([self.W, self.hidbias, self.visbias])