def _core_network(self, l_p, h_p, x_t):
"""
Parameters:
x_t - 28x28 image
l_p - 2x1 focus vector
h_p - 256x1 vector
Returns:
h_t, 256x1 vector
"""
g_t = self._glimpse_network(x_t, l_p)
h_t = self._tanh(
T.dot(g_t, self.W_h_g) + T.dot(h_p, self.W_h) + self.B_h)
l_t = self._location_network(h_t)
if not self.disable_reinforce:
sampled_l_t = self._sample_gaussian(l_t, self.cov)
sampled_pdf = self._multi_gaussian_pdf(
disconnected_grad(sampled_l_t), l_t)
wl_grad = T.grad(T.log(sampled_pdf), self.W_l)
else:
sampled_l_t = l_t
wl_grad = self.W_l
if self.random_glimpse and self.disable_reinforce:
sampled_l_t = self.srng.uniform((2, ), low=-1.7, high=1.7)
a_t = self._action_network(h_t)
return sampled_l_t, h_t, a_t, wl_grad
def _core_network(self, l_p, h_p, x_t):
"""
Parameters:
x_t - 28x28 image
l_p - 2x1 focus vector
h_p - 256x1 vector
Returns:
h_t, 256x1 vector
"""
g_t = self._glimpse_network(x_t, l_p)
h_t = self._tanh(T.dot(g_t, self.W_h_g) + T.dot(h_p, self.W_h) + self.B_h)
l_t = self._location_network(h_t)
if not self.disable_reinforce:
sampled_l_t = self._sample_gaussian(l_t, self.cov)
sampled_pdf = self._multi_gaussian_pdf(disconnected_grad(sampled_l_t), l_t)
wl_grad = T.grad(T.log(sampled_pdf), self.W_l)
else:
sampled_l_t = l_t
wl_grad = self.W_l
if self.random_glimpse and self.disable_reinforce:
sampled_l_t = self.srng.uniform((2,), low=-1.7, high=1.7)
a_t = self._action_network(h_t)
return sampled_l_t, h_t, a_t, wl_grad
def _first_glimpse_sensor(self, x_t):
"""
Compute first glimpse position using down-sampled image.
"""
downsampled_img = theano.tensor.signal.downsample.max_pool_2d(x_t, (4,4))
downsampled_img = downsampled_img.flatten()
first_l = T.dot(downsampled_img, self.W_f)
if self.disable_reinforce:
wf_grad = self.W_f
if self.random_glimpse:
first_l = self.srng.uniform((2,), low=-1.7, high=1.7)
else:
sampled_l_t = self._sample_gaussian(first_l, self.cov)
sampled_pdf = self._multi_gaussian_pdf(disconnected_grad(sampled_l_t), first_l)
wf_grad = T.grad(T.log(sampled_pdf), self.W_f)
first_l = sampled_l_t
return first_l, wf_grad
def _first_glimpse_sensor(self, x_t):
"""
Compute first glimpse position using down-sampled image.
"""
downsampled_img = theano.tensor.signal.downsample.max_pool_2d(
x_t, (4, 4))
downsampled_img = downsampled_img.flatten()
first_l = T.dot(downsampled_img, self.W_f)
if self.disable_reinforce:
wf_grad = self.W_f
if self.random_glimpse:
first_l = self.srng.uniform((2, ), low=-1.7, high=1.7)
else:
sampled_l_t = self._sample_gaussian(first_l, self.cov)
sampled_pdf = self._multi_gaussian_pdf(
disconnected_grad(sampled_l_t), first_l)
wf_grad = T.grad(T.log(sampled_pdf), self.W_f)
first_l = sampled_l_t
return first_l, wf_grad