def test_clone(self):
reg = prior.SpikeAndSlabRegularizer(weight=1.0)
model1 = tf.keras.models.Sequential([
tf.keras.layers.Dense(10, activation='relu',
kernel_regularizer=reg,
bias_regularizer=reg),
tf.keras.layers.Dense(10)])
ndata = 256
data = tf.random.normal((ndata, 10))
pred1 = model1(data)
model2 = bnnmodel.clone_model_and_weights(model1, (None, 10))
model1 = None
pred2 = model2(data)
self.assertAllClose(pred1, pred2, msg='Model cloning failed.')
def test_tfd_prior_spike_and_slab(self): self._check_regularizer(prior.SpikeAndSlabRegularizer())
import scipy.integrate as integrate
import tensorflow.compat.v1 as tf
import tensorflow_probability as tfp
from cold_posterior_bnn.core import prior
tfd = tfp.distributions
REGULARIZER_INSTANCES_LOGPDF = [
prior.NormalRegularizer(stddev=0.1, weight=0.5),
prior.ShiftedNormalRegularizer(mean=0.2, stddev=0.1, weight=0.5),
prior.StretchedNormalRegularizer(offset=0.2, scale=1.2, weight=0.1),
prior.LaplaceRegularizer(stddev=0.1),
prior.CauchyRegularizer(scale=0.2),
prior.SpikeAndSlabRegularizer(scale_spike=0.01, scale_slab=0.3,
mass_spike=0.6, weight=0.8),
]
REGULARIZER_INSTANCES_NOLOGPDF = [
prior.HeNormalRegularizer(scale=0.8, weight=0.7),
prior.GlorotNormalRegularizer(scale=0.8, weight=0.7),
prior.EmpiricalBayesNormal(ig_shape=2.5, ig_scale=0.25, weight=0.3),
prior.HeNormalEBRegularizer(scale=0.25, weight=0.1),
]
REGULARIZER_INSTANCES = REGULARIZER_INSTANCES_LOGPDF + \
REGULARIZER_INSTANCES_NOLOGPDF
# Regularizers with scale parameters, scale=1
REGULARIZER_INSTANCES_SCALE1 = [
('stddev', prior.NormalRegularizer(stddev=1.0, weight=0.5)),
('stddev', prior.ShiftedNormalRegularizer(mean=0., stddev=1.0,
weight=0.5)),