def make_kernel_bias_prior_spike_and_slab(
kernel_shape,
bias_shape,
kernel_initializer=None, # pylint: disable=unused-argument
bias_initializer=None, # pylint: disable=unused-argument
kernel_batch_ndims=0, # pylint: disable=unused-argument
bias_batch_ndims=0, # pylint: disable=unused-argument
dtype=tf.float32,
kernel_name='prior_kernel',
bias_name='prior_bias'):
"""Create prior for Variational layers with kernel and bias.
Note: Distribution scale is inversely related to regularization strength.
Consider a "Normal" prior; bigger scale corresponds to less L2 regularization.
I.e.,
```python
scale = (2. * l2weight)**-0.5
l2weight = scale**-2. / 2.
```
have a similar regularizing effect.
The std. deviation of each of the component distributions returned by this
function is approximately `1415` (or approximately `l2weight = 25e-6`). In
other words this prior is extremely "weak".
Args:
kernel_shape: ...
bias_shape: ...
kernel_initializer: Ignored.
Default value: `None` (i.e., `tf.initializers.glorot_uniform()`).
bias_initializer: Ignored.
Default value: `None` (i.e., `tf.zeros`).
kernel_batch_ndims: ...
Default value: `0`.
bias_batch_ndims: ...
Default value: `0`.
dtype: ...
Default value: `tf.float32`.
kernel_name: ...
Default value: `"prior_kernel"`.
bias_name: ...
Default value: `"prior_bias"`.
Returns:
kernel_and_bias_distribution: ...
"""
w = MixtureSameFamily(mixture_distribution=Categorical(probs=[0.5, 0.5]),
components_distribution=Normal(loc=0.,
scale=tf.constant(
[1., 2000.],
dtype=dtype)))
return JointDistributionSequential([
Sample(w, kernel_shape, name=kernel_name),
Sample(w, bias_shape, name=bias_name),
])
def compute_log_prob_uncompressed(
alignment: Alignment,
transition_probs_tree: TensorflowUnrootedTree,
frequencies,
):
sequences_encoded = alignment.get_encoded_sequence_tensor(
transition_probs_tree.taxon_set)
dist = Sample(
LeafCTMC(transition_probs_tree, frequencies),
sample_shape=(alignment.site_count, ),
)
return dist.log_prob(sequences_encoded)
def make_kernel_bias_prior_spike_and_slab(kernel_shape,
bias_shape,
dtype=tf.float32,
kernel_initializer=None,
bias_initializer=None):
"""Create prior for Variational layers with kernel and bias."""
del kernel_initializer, bias_initializer
w = MixtureSameFamily(mixture_distribution=Categorical(probs=[0.5, 0.5]),
components_distribution=Normal(loc=0.,
scale=tf.constant(
[1., 2000.],
dtype=dtype)))
return JointDistributionSequential([
Sample(w, kernel_shape, name='prior_kernel'),
Sample(w, bias_shape, name='prior_bias'),
])