def inverse_log_det_jacobian(
self, y: Array, event_ndims: Optional[int] = None) -> Array:
"""See `Bijector.inverse_log_det_jacobian`."""
extra_event_ndims = self._check_ndims(
"Inverse", event_ndims, base_bijector.event_ndims_out)
ildj = base_bijector.inverse_log_det_jacobian(y)
return math.sum_last(ildj, extra_event_ndims)
def forward_log_det_jacobian(
self, x: Array, event_ndims: Optional[int] = None) -> Array:
"""See `Bijector.forward_log_det_jacobian`."""
extra_event_ndims = self._check_ndims(
"Forward", event_ndims, base_bijector.event_ndims_in)
fldj = base_bijector.forward_log_det_jacobian(x)
return math.sum_last(fldj, extra_event_ndims)
def log_prob(bijector, inputs):
# forward transformation
outputs, log_det = bijector.forward_and_log_det(inputs)
# probability in the latent space
latent_prob = base_dist.log_prob(outputs)
# log probability
log_prob = sum_last(latent_prob, ndims=latent_prob.ndim -
1) + sum_last(log_det, ndims=log_det.ndim - 1)
# # log probability
# log_prob = sum_last(latent_prob, ndims=latent_prob.ndim) + sum_last(
# log_det, ndims=latent_prob.ndim
# )
return log_prob
def inverse_and_log_det(self, y: Array) -> Tuple[Array, Array]:
"""Computes x = f^{-1}(y) and log|det J(f^{-1})(y)|."""
self._check_inverse_input_shape(y)
masked_y = self._mask * y
params = self._conditioner(masked_y)
x0, log_d = self._inner_bijector(params).inverse_and_log_det(y)
x = self._neg_mask * x0 + masked_y
logdet = math.sum_last(self._neg_mask * log_d, self._event_ndims)
return x, logdet
def forward_and_log_det(self, x: Array) -> Tuple[Array, Array]: """Computes y = f(x) and log|det J(f)(x)|.""" self._check_forward_input_shape(x) masked_x = self._mask * x params = self._conditioner(masked_x) y0, log_d = self._inner_bijector(params).forward_and_log_det(x) y = self._neg_mask * y0 + masked_x logdet = math.sum_last(self._neg_mask * log_d, self._event_ndims) return y, logdet
def inverse_and_log_det(self, y: Array) -> Tuple[Array, Array]:
"""Computes x = f^{-1}(y) and log|det J(f^{-1})(y)|."""
self._check_inverse_input_shape(y)
masked_y = jnp.where(self._event_mask, y, 0.)
params = self._conditioner(masked_y)
x0, log_d = self._inner_bijector(params).inverse_and_log_det(y)
x = jnp.where(self._event_mask, y, x0)
logdet = math.sum_last(jnp.where(self._mask, 0., log_d),
self._event_ndims - self._inner_event_ndims)
return x, logdet
def forward_and_log_det(self, x: Array) -> Tuple[Array, Array]:
"""Computes y = f(x) and log|det J(f)(x)|."""
self._check_forward_input_shape(x)
masked_x = jnp.where(self._event_mask, x, 0.)
params = self._conditioner(masked_x)
y0, log_d = self._inner_bijector(params).forward_and_log_det(x)
y = jnp.where(self._event_mask, x, y0)
logdet = math.sum_last(jnp.where(self._mask, 0., log_d),
self._event_ndims - self._inner_event_ndims)
return y, logdet
def inverse_log_det_jacobian(self, y: Array) -> Array:
"""Computes log|det J(f^{-1})(y)|."""
self._check_inverse_input_shape(y)
log_det = self._bijector.inverse_log_det_jacobian(y)
return math.sum_last(log_det, self._ndims)
def forward_log_det_jacobian(self, x: Array) -> Array:
"""Computes log|det J(f)(x)|."""
self._check_forward_input_shape(x)
log_det = self._bijector.forward_log_det_jacobian(x)
return math.sum_last(log_det, self._ndims)
def inverse_and_log_det(self, y: Array) -> Tuple[Array, Array]:
"""Computes x = f^{-1}(y) and log|det J(f^{-1})(y)|."""
self._check_inverse_input_shape(y)
x, log_det = self._bijector.inverse_and_log_det(y)
return x, math.sum_last(log_det, self._ndims)
def forward_and_log_det(self, x: Array) -> Tuple[Array, Array]:
"""Computes y = f(x) and log|det J(f)(x)|."""
self._check_forward_input_shape(x)
y, log_det = self._bijector.forward_and_log_det(x)
return y, math.sum_last(log_det, self._ndims)
def test_sum_last(self):
x = jax.random.normal(jax.random.PRNGKey(42), (2, 3, 4))
np.testing.assert_array_equal(math.sum_last(x, 0), x)
np.testing.assert_array_equal(math.sum_last(x, 1), x.sum(-1))
np.testing.assert_array_equal(math.sum_last(x, 2), x.sum((-2, -1)))
np.testing.assert_array_equal(math.sum_last(x, 3), x.sum())
