def _log_prob_with_logits(self, x):
# flatten `x` and `logits` as requirement of
# `tf.nn.softmax_cross_entropy_with_logits`
x = tf.cast(x, dtype=self.param_dtype)
x, logits = maybe_explicit_broadcast(x, self.logits)
if x.get_shape().ndims == 2:
x_2d, logits_2d = x, logits
else:
dynamic_shape = tf.stack([-1, tf.shape(x)[-1]], axis=0)
x_2d = tf.reshape(x, dynamic_shape)
logits_2d = tf.reshape(logits, dynamic_shape)
static_shape = (tf.TensorShape([None
]).concatenate(x.get_shape()[-1:]))
x_2d.set_shape(static_shape)
logits_2d.set_shape(static_shape)
# derive the flatten log p(x)
log_p_2d = -tf.nn.softmax_cross_entropy_with_logits(
labels=x_2d,
logits=logits_2d,
)
# reshape log p(x) back into desired shape
if x.get_shape().ndims == 2:
log_p = log_p_2d
else:
static_shape = x.get_shape()[:-1]
log_p = tf.reshape(log_p_2d, tf.shape(x)[:-1])
log_p.set_shape(static_shape)
return log_p
def _log_prob(self, x):
x = tf.cast(x, dtype=self.param_dtype)
if self._probs_is_derived:
logits = self.logits
x, logits = maybe_explicit_broadcast(x, logits)
return self._check_numerics(
-tf.nn.sigmoid_cross_entropy_with_logits(labels=x,
logits=logits),
'log_prob')
else:
# TODO: check whether this is better than using derived logits.
log_p = tf.log(self._probs_clipped)
log_one_minus_p = tf.log1p(-self._probs_clipped)
return self._check_numerics(x * log_p + (1. - x) * log_one_minus_p,
'log_prob')
def variational(self, x, y=None, z=None, z_samples=NOT_SPECIFIED):
"""Derive the `StochasticTensor` objects of variational network.
The behavior of this method is not affected by `y_in_variational_model`,
thus `y` will always be used if it is specified.
Parameters
----------
x : tf.Tensor
The model inputs.
y : tf.Tensor
Optional conditional input for CVAE. See [1] for more details.
[1] K. Sohn, H. Lee, and X. Yan, “Learning structured output
representation using deep conditional generative models,”
in Advances in Neural Information Processing Systems, 2015,
pp. 3483–3491.
z : tf.Tensor | StochasticTensor
The latent variable observations. (default None)
z_samples : int | tf.Tensor | None
If specified, override `z_samples` specified in the constructor.
Returns
-------
StochasticTensor
The derived z stochastic tensors.
"""
if y is None:
features = x
else:
features = tf.concat(maybe_explicit_broadcast(
x, y, tail_no_broadcast_ndims=1),
axis=-1)
z_params = self._z_net(features)
if z_samples is NOT_SPECIFIED:
z_samples = self._z_samples
return self._z_layer(z_params,
n_samples=z_samples,
observed=z,
group_event_ndims=1,
validate_shape=self._validate_shape)
def model(self,
z=None,
y=None,
x=None,
z_samples=NOT_SPECIFIED,
x_samples=None):
"""Derive the `StochasticTensor` objects of generation network.
The behavior of this method is not affected by `y_in_generative_model`,
thus `y` will always be used if it is specified.
Parameters
----------
z : tf.Tensor | StochasticTensor
Optional latent variable samples.
y : tf.Tensor
Optional conditional input for CVAE. See [1] for more details.
[1] K. Sohn, H. Lee, and X. Yan, “Learning structured output
representation using deep conditional generative models,”
in Advances in Neural Information Processing Systems, 2015,
pp. 3483–3491.
x : tf.Tensor
The observations for the output `StochasticTensor`. (default None)
z_samples : int | tf.Tensor | None
If specified, override `z_samples` specified in the constructor.
x_samples : int | tf.Tensor | None
Specify the number of samples to take for output x.
(default None)
Returns
-------
(StochasticTensor, StochasticTensor)
The derived (z, x) stochastic tensors.
"""
if z_samples is NOT_SPECIFIED:
z_samples = self._z_samples
z = self._z_prior.sample_or_observe(
z_samples,
observed=z,
validate_shape=self._validate_shape,
group_event_ndims=1,
name='z')
if y is None:
features = z
else:
features = tf.concat(maybe_explicit_broadcast(
z, y, tail_no_broadcast_ndims=1),
axis=-1)
x_params = self._x_net(features)
x = self._x_layer(x_params,
n_samples=x_samples,
observed=x,
group_event_ndims=1,
validate_shape=self._validate_shape)
return z, x
def test_maybe_explicit_broadcast(self):
with self.get_session() as session:
# test on equal static shape
x = tf.convert_to_tensor(np.arange(2))
y = tf.convert_to_tensor(np.arange(2, 4))
xx, yy = maybe_explicit_broadcast(x, y)
self.assertIs(xx, x)
self.assertEqual(xx.get_shape().as_list(), [2])
np.testing.assert_equal(xx.eval(), [0, 1])
self.assertIs(yy, y)
self.assertEqual(yy.get_shape().as_list(), [2])
np.testing.assert_equal(yy.eval(), [2, 3])
# test on fully static same dimensional shape
x = tf.convert_to_tensor(np.arange(2).reshape([1, 2]))
y = tf.convert_to_tensor(np.arange(2, 4).reshape([2, 1]))
xx, yy = maybe_explicit_broadcast(x, y)
self.assertEqual(xx.get_shape().as_list(), [2, 2])
np.testing.assert_equal(xx.eval(), [[0, 1], [0, 1]])
self.assertEqual(yy.get_shape().as_list(), [2, 2])
np.testing.assert_equal(yy.eval(), [[2, 2], [3, 3]])
# test on fully static different dimensional shape
x = tf.convert_to_tensor(np.arange(2).reshape([2, 1]))
y = tf.convert_to_tensor(np.arange(2, 4))
xx, yy = maybe_explicit_broadcast(x, y)
self.assertEqual(xx.get_shape().as_list(), [2, 2])
np.testing.assert_equal(xx.eval(), [[0, 0], [1, 1]])
self.assertEqual(yy.get_shape().as_list(), [2, 2])
np.testing.assert_equal(yy.eval(), [[2, 3], [2, 3]])
# test on dynamic same dimensional shape
x = tf.placeholder(tf.float32, (None, 3))
y = tf.placeholder(tf.float32, (2, None))
xx, yy = maybe_explicit_broadcast(x, y)
x_data = np.arange(3).reshape((1, 3))
y_data = np.arange(3, 5).reshape((2, 1))
self.assertEqual(xx.get_shape().as_list(), [2, 3])
np.testing.assert_equal(xx.eval({x: x_data, y: y_data}),
[[0, 1, 2], [0, 1, 2]])
self.assertEqual(yy.get_shape().as_list(), [2, 3])
np.testing.assert_equal(yy.eval({x: x_data, y: y_data}),
[[3, 3, 3], [4, 4, 4]])
# test on dynamic different dimensional shape
x = tf.placeholder(tf.float32, (None,))
y = tf.placeholder(tf.float32, (None, 2))
xx, yy = maybe_explicit_broadcast(x, y)
x_data = np.arange(1, 2)
y_data = np.arange(2, 6).reshape((2, 2))
self.assertEqual(xx.get_shape().as_list(), [None, 2])
np.testing.assert_equal(xx.eval({x: x_data, y: y_data}),
[[1, 1], [1, 1]])
self.assertEqual(yy.get_shape().as_list(), [None, 2])
np.testing.assert_equal(yy.eval({x: x_data, y: y_data}),
[[2, 3], [4, 5]])
# test `tail_no_broadcast_ndims = 0`
x = tf.convert_to_tensor(np.arange(2).reshape([1, 1, 2]))
y = tf.convert_to_tensor(np.arange(2, 4).reshape([2, 1]))
xx, yy = maybe_explicit_broadcast(x, y, tail_no_broadcast_ndims=0)
self.assertEqual(xx.get_shape().as_list(), [1, 2, 2])
np.testing.assert_equal(xx.eval(), [[[0, 1], [0, 1]]])
self.assertEqual(yy.get_shape().as_list(), [1, 2, 2])
np.testing.assert_equal(yy.eval(), [[[2, 2], [3, 3]]])
# test `tail_no_broadcast_ndims = 1`
x = tf.convert_to_tensor(np.arange(2).reshape([1, 1, 2]))
y = tf.convert_to_tensor(np.arange(2, 4).reshape([2, 1]))
xx, yy = maybe_explicit_broadcast(x, y, tail_no_broadcast_ndims=1)
self.assertEqual(xx.get_shape().as_list(), [1, 2, 2])
np.testing.assert_equal(xx.eval(), [[[0, 1], [0, 1]]])
self.assertEqual(yy.get_shape().as_list(), [1, 2, 1])
np.testing.assert_equal(yy.eval(), [[[2], [3]]])
# test `tail_no_broadcast_ndims = 2`
x = tf.convert_to_tensor(np.arange(2).reshape([1, 1, 2]))
y = tf.convert_to_tensor(np.arange(2, 4).reshape([2, 1]))
xx, yy = maybe_explicit_broadcast(x, y, tail_no_broadcast_ndims=2)
self.assertEqual(xx.get_shape().as_list(), [1, 1, 2])
np.testing.assert_equal(xx.eval(), [[[0, 1]]])
self.assertEqual(yy.get_shape().as_list(), [1, 2, 1])
np.testing.assert_equal(yy.eval(), [[[2], [3]]])
# test `tail_no_broadcast_ndims = 1` on dynamic shape
x = tf.placeholder(tf.float32, (None, None, None))
y = tf.placeholder(tf.float32, (None, None))
xx, yy = maybe_explicit_broadcast(x, y, tail_no_broadcast_ndims=1)
x_data = np.arange(2).reshape([1, 1, 2])
y_data = np.arange(2, 4).reshape([2, 1])
self.assertEqual(xx.get_shape().as_list(), [None, None, None])
np.testing.assert_equal(xx.eval({x: x_data, y: y_data}),
[[[0, 1], [0, 1]]])
self.assertEqual(yy.get_shape().as_list(), [None, None, None])
np.testing.assert_equal(yy.eval({x: x_data, y: y_data}),
[[[2], [3]]])
# test `tail_no_broadcast_ndims = 1` on fully dynamic shape
x = tf.placeholder(tf.float32)
y = tf.placeholder(tf.float32)
xx, yy = maybe_explicit_broadcast(x, y, tail_no_broadcast_ndims=1)
x_data = np.arange(2).reshape([1, 1, 2])
y_data = np.arange(2, 4).reshape([2, 1])
self.assertEqual(xx.get_shape().ndims, None)
np.testing.assert_equal(xx.eval({x: x_data, y: y_data}),
[[[0, 1], [0, 1]]])
self.assertEqual(yy.get_shape().ndims, None)
np.testing.assert_equal(yy.eval({x: x_data, y: y_data}),
[[[2], [3]]])
# test error on static shape
with self.assertRaisesRegex(
ValueError, '.* and .* cannot broadcast to match.*'):
_ = maybe_explicit_broadcast(
tf.convert_to_tensor(np.arange(2)),
tf.convert_to_tensor(np.arange(3))
)
# test error on dynamic shape
x = tf.placeholder(tf.float32, (None, 3))
y = tf.placeholder(tf.float32, (2, None))
xx, yy = maybe_explicit_broadcast(x, y)
x_data = np.arange(3).reshape((1, 3))
y_data = np.arange(3, 7).reshape((2, 2))
with self.assertRaisesRegex(Exception, r'.*Incompatible shapes.*'):
session.run([xx, yy], feed_dict={x: x_data, y: y_data})