def get_output(self, input):
alpha = T.exp(self.alpha_).dimshuffle('x', 0, 'x', 'x')
lam = T.exp(self.lambda_).dimshuffle('x', 0, 'x', 'x')
down = self.__downsampling(input)
down_up = utils.unpool(down, self.ws)
W = alpha + self.__penalty(input - down_up, lam)
output = pool(input * W, self.ws, mode='average_exc_pad')
weight = 1. / pool(W, self.ws, mode='average_exc_pad')
return output * weight
def mean_interp_pad(x, padding):
padding = (padding, padding) if isinstance(padding, int) else tuple(padding)
size = tuple(np.array(padding) * 2 + 1)
resize = ((x.shape[2] + 2 * padding[0], x.shape[2] - 2 * padding[0]),
(x.shape[3] + 2 * padding[1], x.shape[3] - 2 * padding[1]))
y = pool(x, size, (1, 1), mode='average_exc_pad')
z = G.disconnected_grad(nn.utils.frac_bilinear_upsampling(y, resize))
_, _, h, w = z.shape
return T.set_subtensor(z[:, :, padding[0]:h - padding[0], padding[1]:w - padding[1]], x)
def __downsampling(self, input):
output = pool(input, self.ws, stride=self.ws, mode='average_exc_pad')
output = conv(output, self.kern, border_mode='half')
return output
def get_output(self, input, *args, **kwargs):
return pool(input, self.ws, self.stride, self.mode,
self.pad) if self.ignore_border else T.signal.pool.pool_2d(
input, self.ws, self.ignore_border, self.stride,
self.pad, self.mode)