def _parameter_properties(cls, dtype, num_classes=None):
# pylint: disable=g-long-lambda
return dict(total_count=parameter_properties.ParameterProperties(
default_constraining_bijector_fn=parameter_properties.
BIJECTOR_NOT_IMPLEMENTED),
concentration1=parameter_properties.ParameterProperties(
default_constraining_bijector_fn=(
lambda: softplus_bijector.Softplus(low=dtype_util.
eps(dtype)))),
concentration0=parameter_properties.ParameterProperties(
default_constraining_bijector_fn=(
lambda: softplus_bijector.Softplus(low=dtype_util.
eps(dtype)))))
def _parameter_properties(cls, dtype, num_classes=None):
# pylint: disable=g-long-lambda
return dict(
distribution=parameter_properties.BatchedComponentProperties(),
shift=parameter_properties.ParameterProperties(),
scale=parameter_properties.ParameterProperties(
default_constraining_bijector_fn=(
lambda: softplus_bijector.Softplus(low=dtype_util.eps(dtype
)))),
tailweight=parameter_properties.ParameterProperties(
default_constraining_bijector_fn=(
lambda: softplus_bijector.Softplus(low=dtype_util.eps(dtype
)))))
def _parameter_properties(cls, dtype, num_classes=None):
# pylint: disable=g-long-lambda
return dict(
loc=parameter_properties.ParameterProperties(),
scale=parameter_properties.ParameterProperties(
default_constraining_bijector_fn=(
lambda: softplus_bijector.Softplus(low=dtype_util.eps(dtype
)))),
low=parameter_properties.ParameterProperties(),
# TODO(b/169874884): Support decoupled parameterization.
high=parameter_properties.ParameterProperties(
default_constraining_bijector_fn=parameter_properties.
BIJECTOR_NOT_IMPLEMENTED, ))
def _parameter_properties(cls, dtype, num_classes=None):
# pylint: disable=g-long-lambda
return dict(
temperature=parameter_properties.ParameterProperties(
shape_fn=lambda sample_shape: sample_shape[:-1],
default_constraining_bijector_fn=(
lambda: softplus_bijector.Softplus(low=dtype_util.eps(dtype)))),
logits=parameter_properties.ParameterProperties(event_ndims=1),
probs=parameter_properties.ParameterProperties(
event_ndims=1,
default_constraining_bijector_fn=softmax_centered_bijector
.SoftmaxCentered,
is_preferred=False))
def _parameter_properties(cls, dtype, num_classes=None):
return dict(
loc=parameter_properties.ParameterProperties(event_ndims=1),
atol=parameter_properties.ParameterProperties(
event_ndims=1,
default_constraining_bijector_fn=parameter_properties.
BIJECTOR_NOT_IMPLEMENTED,
is_preferred=False),
rtol=parameter_properties.ParameterProperties(
event_ndims=1,
default_constraining_bijector_fn=parameter_properties.
BIJECTOR_NOT_IMPLEMENTED,
is_preferred=False))
def _parameter_properties(cls, dtype):
from tensorflow_probability.python.bijectors import softplus # pylint:disable=g-import-not-at-top
return dict(
base_kernel=parameter_properties.BatchedComponentProperties(),
fixed_inputs=parameter_properties.ParameterProperties(
event_ndims=lambda self: self.base_kernel.feature_ndims + 1),
fixed_inputs_mask=parameter_properties.ParameterProperties(
event_ndims=1),
diag_shift=parameter_properties.ParameterProperties(
default_constraining_bijector_fn=(
lambda: softplus.Softplus(low=dtype_util.eps(dtype)))),
_precomputed_divisor_matrix_cholesky=(
parameter_properties.ParameterProperties(event_ndims=2)))
def _parameter_properties(cls, dtype):
from tensorflow_probability.python.bijectors import softplus # pylint:disable=g-import-not-at-top
return dict(
bias_variance=parameter_properties.ParameterProperties(
default_constraining_bijector_fn=(
lambda: softplus.Softplus(low=dtype_util.eps(dtype)))),
exponent=parameter_properties.ParameterProperties(
default_constraining_bijector_fn=(
lambda: softplus.Softplus(low=dtype_util.eps(dtype)))),
slope_variance=parameter_properties.ParameterProperties(
default_constraining_bijector_fn=(
lambda: softplus.Softplus(low=dtype_util.eps(dtype)))),
shift=parameter_properties.ParameterProperties())
def _parameter_properties(cls, dtype, num_classes=None):
return dict(
rate1=parameter_properties.
ParameterProperties(default_constraining_bijector_fn=(
lambda: softplus_bijector.Softplus(low=dtype_util.eps(dtype))),
is_preferred=False),
rate2=parameter_properties.
ParameterProperties(default_constraining_bijector_fn=(
lambda: softplus_bijector.Softplus(low=dtype_util.eps(dtype))),
is_preferred=False),
log_rate1=parameter_properties.ParameterProperties(),
log_rate2=parameter_properties.ParameterProperties(),
)
def _parameter_properties(cls, dtype, num_classes=None):
# pylint: disable=g-long-lambda
return dict(
loc=parameter_properties.ParameterProperties(event_ndims=1),
scale_diag=parameter_properties.ParameterProperties(
event_ndims=1,
default_constraining_bijector_fn=(
lambda: softplus_bijector.Softplus(low=dtype_util.eps(dtype
)))),
scale_identity_multiplier=parameter_properties.
ParameterProperties(default_constraining_bijector_fn=(
lambda: softplus_bijector.Softplus(low=dtype_util.eps(dtype))),
is_preferred=False))
def _parameter_properties(cls, dtype, num_classes=None):
return dict(concentration=parameter_properties.ParameterProperties(
default_constraining_bijector_fn=(
lambda: softplus_bijector.Softplus(low=dtype_util.eps(dtype))
)),
scale=parameter_properties.ParameterProperties(
default_constraining_bijector_fn=(
lambda: softplus_bijector.Softplus(low=dtype_util.
eps(dtype)))),
upper_bound=parameter_properties.ParameterProperties(
default_constraining_bijector_fn=(
lambda: softplus_bijector.Softplus(low=dtype_util.
eps(dtype)))))
def _parameter_properties(cls, dtype, num_classes=None):
# pylint: disable=g-long-lambda
return dict(
logits=parameter_properties.ParameterProperties(
event_ndims=1,
shape_fn=lambda sample_shape: ps.concat(
[sample_shape, [num_classes]], axis=0)),
probs=parameter_properties.ParameterProperties(
event_ndims=1,
shape_fn=lambda sample_shape: ps.concat(
[sample_shape, [num_classes]], axis=0),
default_constraining_bijector_fn=softmax_centered_bijector.
SoftmaxCentered,
is_preferred=False))
def _parameter_properties(cls, dtype, num_classes=None):
from tensorflow_probability.python.bijectors import ascending # pylint:disable=g-import-not-at-top
from tensorflow_probability.python.bijectors import softplus # pylint:disable=g-import-not-at-top
return dict(
kernels=parameter_properties.BatchedComponentProperties(
event_ndims=lambda self: [0 for _ in self.kernels]),
locs=parameter_properties.ParameterProperties(
event_ndims=1,
default_constraining_bijector_fn=lambda: ascending.Ascending(
)), # pylint:disable=unnecessary-lambda
slopes=parameter_properties.ParameterProperties(
event_ndims=1,
default_constraining_bijector_fn=(
lambda: softplus.Softplus(low=dtype_util.eps(dtype)))))
def _parameter_properties(cls, dtype):
from tensorflow_probability.python.bijectors import softplus # pylint:disable=g-import-not-at-top
return dict(
amplitude=parameter_properties.ParameterProperties(
default_constraining_bijector_fn=(
lambda: softplus.Softplus(low=dtype_util.eps(dtype)))),
length_scale=parameter_properties.ParameterProperties(
default_constraining_bijector_fn=(
lambda: softplus.Softplus(low=dtype_util.eps(dtype)))),
inverse_length_scale=parameter_properties.ParameterProperties(
default_constraining_bijector_fn=softplus.Softplus),
scale_mixture_rate=parameter_properties.ParameterProperties(
default_constraining_bijector_fn=(
lambda: softplus.Softplus(low=dtype_util.eps(dtype)))))
def _parameter_properties(cls, dtype, num_classes=None):
# pylint: disable=g-long-lambda
return dict(
loc=parameter_properties.ParameterProperties(event_ndims=1),
covariance_matrix=parameter_properties.ParameterProperties(
event_ndims=2,
shape_fn=lambda sample_shape: ps.concat(
[sample_shape, sample_shape[-1:]], axis=0),
default_constraining_bijector_fn=(
lambda: chain_bijector.Chain([
cholesky_outer_product_bijector.CholeskyOuterProduct(),
fill_scale_tril_bijector.FillScaleTriL(
diag_shift=dtype_util.eps(dtype))
]))))
def _parameter_properties(cls, dtype, num_classes=None):
return dict(
df=parameter_properties.ParameterProperties(
default_constraining_bijector_fn=(
lambda: softplus_bijector.Softplus( # pylint: disable=g-long-lambda
low=dtype_util.as_numpy_dtype(dtype)(2.)))),
index_points=parameter_properties.ParameterProperties(
event_ndims=lambda self: self.kernel.feature_ndims + 1,
shape_fn=parameter_properties.SHAPE_FN_NOT_IMPLEMENTED),
kernel=parameter_properties.BatchedComponentProperties(),
observation_noise_variance=parameter_properties.ParameterProperties(
default_constraining_bijector_fn=(
lambda: softplus_bijector.Softplus(low=dtype_util.eps(dtype))),
shape_fn=parameter_properties.SHAPE_FN_NOT_IMPLEMENTED))
def _parameter_properties(cls, dtype, num_classes=None):
# pylint: disable=g-long-lambda
return dict(
concentration=parameter_properties.ParameterProperties(
default_constraining_bijector_fn=(
lambda: softplus_bijector.Softplus(low=dtype_util.eps(dtype
)))),
mixing_concentration=parameter_properties.ParameterProperties(
default_constraining_bijector_fn=(
lambda: softplus_bijector.Softplus(low=dtype_util.eps(dtype
)))),
mixing_rate=parameter_properties.ParameterProperties(
default_constraining_bijector_fn=(
lambda: softplus_bijector.Softplus(low=dtype_util.eps(dtype
)))))
def _parameter_properties(cls, dtype, num_classes=None):
return dict(
loc=parameter_properties.ParameterProperties(event_ndims=1),
precision_factor=parameter_properties.BatchedComponentProperties(),
precision=parameter_properties.BatchedComponentProperties(),
nonzeros=parameter_properties.BatchedComponentProperties(
event_ndims=1))
def _parameter_properties(cls, dtype, num_classes=None):
# pylint: disable=g-long-lambda
return dict(df=parameter_properties.ParameterProperties(
shape_fn=lambda sample_shape: sample_shape[:-2],
default_constraining_bijector_fn=parameter_properties.
BIJECTOR_NOT_IMPLEMENTED),
scale=parameter_properties.BatchedComponentProperties())
def _parameter_properties(cls, dtype, num_classes=None):
# pylint: disable=g-long-lambda
return dict(concentration=parameter_properties.ParameterProperties(
shape_fn=lambda sample_shape: sample_shape[:-2],
default_constraining_bijector_fn=(
lambda: softplus_bijector.Softplus(low=tf.convert_to_tensor(
1. + dtype_util.eps(dtype), dtype=dtype)))))
def _parameter_properties(cls, dtype, num_classes=None):
# pylint: disable=g-long-lambda
return dict(
scores=parameter_properties.ParameterProperties(
event_ndims=1,
default_constraining_bijector_fn=(
lambda: softplus_bijector.Softplus(low=dtype_util.eps(dtype)))))
def _parameter_properties(cls, dtype, num_classes=None):
# pylint: disable=g-long-lambda
return dict(
power=parameter_properties.ParameterProperties(
default_constraining_bijector_fn=(
lambda: softplus_bijector.Softplus(
low=tf.convert_to_tensor(
1. + dtype_util.eps(dtype), dtype=dtype)))))
def _parameter_properties(cls, dtype):
from tensorflow_probability.python.bijectors import softplus # pylint:disable=g-import-not-at-top
return dict(
kernel=parameter_properties.BatchedComponentProperties(),
scale_diag=parameter_properties.ParameterProperties(
event_ndims=lambda self: self.kernel.feature_ndims,
default_constraining_bijector_fn=(
lambda: softplus.Softplus(low=dtype_util.eps(dtype)))))
def _parameter_properties(cls, dtype, num_classes=None):
return dict(
loc=parameter_properties.ParameterProperties(
event_ndims=1),
scale_diag=parameter_properties.ParameterProperties(
event_ndims=1,
default_constraining_bijector_fn=(
lambda: softplus_bijector.Softplus(low=dtype_util.eps(dtype)))),
scale_perturb_factor=parameter_properties.ParameterProperties(
event_ndims=2,
shape_fn=parameter_properties.SHAPE_FN_NOT_IMPLEMENTED,
is_preferred=False),
scale_perturb_diag=parameter_properties.ParameterProperties(
event_ndims=1,
is_preferred=False,
default_constraining_bijector_fn=(
lambda: softplus_bijector.Softplus(low=dtype_util.eps(dtype)))))
def _parameter_properties(cls, dtype):
return {
'scale_diag':
parameter_properties.ParameterProperties(
event_ndims=1,
default_constraining_bijector_fn=(
lambda: softplus.Softplus(low=dtype_util.eps(dtype))))
}
def _parameter_properties(cls, dtype):
# pylint: disable=g-long-lambda
return dict(scale_tril=parameter_properties.ParameterProperties(
event_ndims=2,
shape_fn=lambda sample_shape: ps.concat(
[sample_shape, sample_shape[-1:]], axis=0),
default_constraining_bijector_fn=fill_triangular_bijector.
FillTriangular))
def _parameter_properties(cls, dtype, num_classes=None):
from tensorflow_probability.python.bijectors import softplus as softplus_bijector # pylint:disable=g-import-not-at-top
return dict(
amplitudes=parameter_properties.ParameterProperties(
event_ndims=1,
default_constraining_bijector_fn=(
softplus_bijector.Softplus(low=dtype_util.eps(dtype)))),
kernel=parameter_properties.BatchedComponentProperties(event_ndims=1))
def _parameter_properties(cls, dtype, num_classes=None):
return dict(
total_count=parameter_properties.ParameterProperties(
default_constraining_bijector_fn=parameter_properties.
BIJECTOR_NOT_IMPLEMENTED,
# The method `_sample_bates` currently constructs intermediate
# samples with a shape that depends on `total_count`, so, although
# `total_count` is not *inherently* a shape parameter, we annotate
# it as one in the current implementation (making it the rare case
# of a shape parameter that also has batch semantics). This could
# be removed if a different sampling method (eg, rejection sampling)
# were used.
specifies_shape=True),
low=parameter_properties.ParameterProperties(),
# TODO(b/169874884): Support decoupled parameterization.
high=parameter_properties.ParameterProperties(
default_constraining_bijector_fn=parameter_properties.
BIJECTOR_NOT_IMPLEMENTED))
def _parameter_properties(cls, dtype):
return dict(
bin_widths=parameter_properties.ParameterProperties(
event_ndims=1,
shape_fn=parameter_properties.SHAPE_FN_NOT_IMPLEMENTED,
default_constraining_bijector_fn=parameter_properties
.BIJECTOR_NOT_IMPLEMENTED),
bin_heights=parameter_properties.ParameterProperties(
event_ndims=1,
shape_fn=parameter_properties.SHAPE_FN_NOT_IMPLEMENTED,
default_constraining_bijector_fn=parameter_properties
.BIJECTOR_NOT_IMPLEMENTED),
knot_slopes=parameter_properties.ParameterProperties(
event_ndims=1,
shape_fn=parameter_properties.SHAPE_FN_NOT_IMPLEMENTED,
default_constraining_bijector_fn=(
lambda: softplus_bijector.Softplus(low=dtype_util.eps(dtype)))),
range_min=parameter_properties.ParameterProperties(
shape_fn=parameter_properties.SHAPE_FN_NOT_IMPLEMENTED,))
def _parameter_properties(cls, dtype):
return dict(rightmost_transposed_ndims=parameter_properties.
ParameterProperties(
shape_fn=lambda sample_shape: [],
default_constraining_bijector_fn=parameter_properties.
BIJECTOR_NOT_IMPLEMENTED,
is_preferred=False),
perm=parameter_properties.ParameterProperties(
event_ndims=1,
shape_fn=parameter_properties.SHAPE_FN_NOT_IMPLEMENTED,
default_constraining_bijector_fn=parameter_properties.
BIJECTOR_NOT_IMPLEMENTED))
def _parameter_properties(cls, dtype, num_classes=None):
return dict(index_points=parameter_properties.ParameterProperties(
event_ndims=lambda self: self.kernel.feature_ndims + 1,
shape_fn=parameter_properties.SHAPE_FN_NOT_IMPLEMENTED,
),
kernel=parameter_properties.BatchedComponentProperties(),
observation_noise_variance=parameter_properties.
ParameterProperties(
event_ndims=0,
shape_fn=lambda sample_shape: sample_shape[:-1],
default_constraining_bijector_fn=(
lambda: softplus_bijector.Softplus(low=dtype_util.
eps(dtype)))))
