def _batch_data(batch):
images = float_tensor(batch[0].float())
bsize = len(images)
return m(
images=images,
x1=float_tensor(batch[3].float()),
x2=float_tensor(batch[4].float()),
id_labels=init(batch[1]),
pose_labels=init(batch[2]),
fake_pose_labels=long_tensor(np.random.randint(args.Np,
size=bsize)),
ones=ones(bsize),
zeros=zeros(bsize),
noise=float_tensor(np.random.uniform(-1., 1., (bsize, args.Nz))))
def blank_loglikes(n):
a = ones((NUM_CHARS, n)) * 0.1
a[0, :] = 0.9
a /= sqrt(square(a).sum(axis=0))
return log(a)
def uniform_loglikes(n):
return log(ones((NUM_CHARS, n)) / float(NUM_CHARS))
def bprop(self):
logger.debug('%s backprop' % str(self))
# TODO This can be merged / sped up
self.grad = ones(self.pred[0].out.shape)
self.full_grad = self.grad
def bprop(self):
logger.debug("%s backprop" % str(self))
# TODO This can be merged / sped up
self.grad = ones(self.pred[0].out.shape)
self.full_grad = self.grad
def blank_loglikes(n):
a = ones((NUM_CHARS, n)) * 0.1
a[0, :] = 0.9
a /= sqrt(square(a).sum(axis=0))
return log(a)
def uniform_loglikes(n):
return log(ones((NUM_CHARS, n)) / float(NUM_CHARS))
