def test_kl_divergence_raises_on_event_shape(self):
base_dist1 = tfd.MultivariateNormalDiag([0.1, 0.5, 0.9], [0.1, 1.1, 2.5])
base_dist2 = tfd.Normal(-0.1, 1.5)
bij1 = block.Block(lambda_bijector.Lambda(lambda x: x), ndims=1)
bij2 = lambda_bijector.Lambda(lambda x: x)
distrax_dist1 = transformed.Transformed(base_dist1, bij1)
distrax_dist2 = transformed.Transformed(base_dist2, bij2)
with self.assertRaises(ValueError):
distrax_dist1.kl_divergence(distrax_dist2)
def as_bijector(obj: BijectorLike) -> bijector.BijectorT:
"""Converts a bijector-like object to a Distrax bijector.
Bijector-like objects are: Distrax bijectors, TFP bijectors, and callables.
Distrax bijectors are returned unchanged. TFP bijectors are converted to a
Distrax equivalent. Callables are wrapped by `distrax.Lambda`, with a few
exceptions where an explicit implementation already exists and is returned.
Args:
obj: The bijector-like object to be converted.
Returns:
A Distrax bijector.
"""
if isinstance(obj, bijector.Bijector):
return obj
elif isinstance(obj, tfb.Bijector):
return bijector_from_tfp.BijectorFromTFP(obj)
elif obj is jax.nn.sigmoid:
return sigmoid.Sigmoid()
elif obj is jnp.tanh:
return tanh.Tanh()
elif callable(obj):
return lambda_bijector.Lambda(obj)
else:
raise TypeError(
f"A bijector-like object can be a `distrax.Bijector`, a `tfb.Bijector`,"
f" or a callable. Got type `{type(obj)}`.")
def test_sample_and_log_prob(self, mu, sigma, sample_shape, base_dist):
base = base_dist(mu, sigma)
bijector = lambda_bijector.Lambda(lambda x: 10 * jnp.tanh(0.1 * x))
dist = transformed.Transformed(base, bijector)
def sample_and_log_prob_fn(seed, sample_shape):
return dist.sample_and_log_prob(seed=seed,
sample_shape=sample_shape)
samples, log_prob = self.variant(sample_and_log_prob_fn,
ignore_argnums=(1, ),
static_argnums=(1, ))(self.seed,
sample_shape)
expected_samples = bijector.forward(
base.sample(seed=self.seed, sample_shape=sample_shape))
tfp_bijector = tfb.Chain([tfb.Scale(10), tfb.Tanh(), tfb.Scale(0.1)])
tfp_dist = tfd.TransformedDistribution(conversion.to_tfp(base),
tfp_bijector)
tfp_samples = tfp_dist.sample(seed=self.seed,
sample_shape=sample_shape)
tfp_log_prob = tfp_dist.log_prob(samples)
chex.assert_equal_shape([samples, tfp_samples])
np.testing.assert_allclose(log_prob, tfp_log_prob, rtol=RTOL)
np.testing.assert_allclose(samples, expected_samples, rtol=RTOL)
def test_kl_divergence(self, mode_string):
base_dist1 = tfd.Normal([0.1, 0.5, 0.9], [0.1, 1.1, 2.5])
base_dist2 = tfd.Normal(-0.1, 1.5)
bij_tfp1 = tfb.Identity()
bij_tfp2 = tfb.Identity()
bij_distrax1 = bij_tfp1
bij_distrax2 = lambda_bijector.Lambda(lambda x: x)
tfp_dist1 = tfd.TransformedDistribution(base_dist1, bij_tfp1)
tfp_dist2 = tfd.TransformedDistribution(base_dist2, bij_tfp2)
distrax_dist1 = transformed.Transformed(base_dist1, bij_distrax1)
distrax_dist2 = transformed.Transformed(base_dist2, bij_distrax2)
expected_result_fwd = base_dist1.kl_divergence(base_dist2)
expected_result_inv = base_dist2.kl_divergence(base_dist1)
distrax_fn1 = self.variant(distrax_dist1.kl_divergence)
distrax_fn2 = self.variant(distrax_dist2.kl_divergence)
if mode_string == 'distrax_to_distrax':
result_fwd = distrax_fn1(distrax_dist2)
result_inv = distrax_fn2(distrax_dist1)
elif mode_string == 'distrax_to_tfp':
result_fwd = distrax_fn1(tfp_dist2)
result_inv = distrax_fn2(tfp_dist1)
elif mode_string == 'tfp_to_distrax':
result_fwd = tfp_dist1.kl_divergence(distrax_dist2)
result_inv = tfp_dist2.kl_divergence(distrax_dist1)
np.testing.assert_allclose(result_fwd, expected_result_fwd, rtol=RTOL)
np.testing.assert_allclose(result_inv, expected_result_inv, rtol=RTOL)
def test_jittable(self):
@jax.jit
def f(x, b):
return b.forward(x)
bijector = lambda_bijector.Lambda(lambda x: x)
x = np.zeros(())
f(x, bijector)
def test_kl_divergence_raises_on_different_bijectors(self):
base_dist1 = tfd.Normal([0.1, 0.5, 0.9], [0.1, 1.1, 2.5])
base_dist2 = tfd.Normal(-0.1, 1.5)
bij1 = lambda_bijector.Lambda(lambda x: x)
bij2 = sigmoid.Sigmoid()
distrax_dist1 = transformed.Transformed(base_dist1, bij1)
distrax_dist2 = transformed.Transformed(base_dist2, bij2)
with self.assertRaises(NotImplementedError):
distrax_dist1.kl_divergence(distrax_dist2)
def test_event_shape(self, mu, sigma, base_dist):
base = base_dist(mu, sigma)
bijector = lambda_bijector.Lambda(jnp.tanh)
dist = transformed.Transformed(base, bijector)
tfp_bijector = tfb.Tanh()
tfp_dist = tfd.TransformedDistribution(conversion.to_tfp(base),
tfp_bijector)
assert dist.event_shape == tfp_dist.event_shape
def test_prob(self, mu, sigma, value, base_dist):
base = base_dist(mu, sigma)
bijector = lambda_bijector.Lambda(jnp.tanh)
dist = transformed.Transformed(base, bijector)
actual = self.variant(dist.prob)(value)
tfp_bijector = tfb.Tanh()
tfp_dist = tfd.TransformedDistribution(conversion.to_tfp(base),
tfp_bijector)
expected = tfp_dist.prob(value)
np.testing.assert_allclose(actual, expected, rtol=RTOL)
def test_log_dets(self, lambda_bjct, tfp_bijector_fn):
bijector = lambda_bijector.Lambda(lambda_bjct)
tfp_bijector = tfp_bijector_fn()
x = np.array([0.05, 0.3, 0.45], dtype=np.float32)
fldj = tfp_bijector.forward_log_det_jacobian(x, event_ndims=0)
fldj_ = self.variant(bijector.forward_log_det_jacobian)(x)
np.testing.assert_allclose(fldj_, fldj, rtol=RTOL)
y = bijector.forward(x)
ildj = tfp_bijector.inverse_log_det_jacobian(y, event_ndims=0)
ildj_ = self.variant(bijector.inverse_log_det_jacobian)(y)
np.testing.assert_allclose(ildj_, ildj, rtol=RTOL)
def test_method(self, function_string, mu, sigma, base_dist):
base = base_dist(mu, sigma)
bijector = lambda_bijector.Lambda(lambda x: x + 3)
dist = transformed.Transformed(base, bijector)
tfp_bijector = tfb.Shift(3)
tfp_dist = tfd.TransformedDistribution(conversion.to_tfp(base),
tfp_bijector)
np.testing.assert_allclose(self.variant(getattr(dist,
function_string))(),
getattr(tfp_dist, function_string)(),
rtol=RTOL)
def test_raises_on_invalid_input_shape(self):
bij = lambda_bijector.Lambda(
forward=lambda x: x,
inverse=lambda y: y,
forward_log_det_jacobian=lambda x: jnp.zeros_like(x[:-1]),
inverse_log_det_jacobian=lambda y: jnp.zeros_like(y[:-1]),
event_ndims_in=1)
for fn in [
bij.forward, bij.inverse, bij.forward_log_det_jacobian,
bij.inverse_log_det_jacobian, bij.forward_and_log_det,
bij.inverse_and_log_det
]:
with self.assertRaises(ValueError):
fn(jnp.array(0))
def test_against_tfp_bijectors(self, lambda_bjct, tfp_bijector, base_dist):
mu = np.array([-1.0, 0.0, 1.0], dtype=np.float32)
sigma = np.array([0.5, 1.0, 2.5], dtype=np.float32)
base = base_dist(mu, sigma)
bijector = lambda_bijector.Lambda(lambda_bjct)
dist = transformed.Transformed(base, bijector)
tfp_dist = tfd.TransformedDistribution(conversion.to_tfp(base),
tfp_bijector())
y = np.array([0.05, 0.3, 0.95], dtype=np.float32)
lp_y = tfp_dist.log_prob(y)
lp_y_ = self.variant(dist.log_prob)(y)
np.testing.assert_allclose(lp_y_, lp_y, rtol=RTOL)
p_y = tfp_dist.prob(y)
p_y_ = self.variant(dist.prob)(y)
np.testing.assert_allclose(p_y_, p_y, rtol=RTOL)
def test_sample_shape(self, mu, sigma, sample_shape, base_dist):
base = base_dist(mu, sigma)
bijector = lambda_bijector.Lambda(jnp.tanh)
dist = transformed.Transformed(base, bijector)
def sample_fn(seed, sample_shape):
return dist.sample(seed=seed, sample_shape=sample_shape)
samples = self.variant(sample_fn,
ignore_argnums=(1, ),
static_argnums=1)(self.seed, sample_shape)
tfp_bijector = tfb.Tanh()
tfp_dist = tfd.TransformedDistribution(conversion.to_tfp(base),
tfp_bijector)
tfp_samples = tfp_dist.sample(sample_shape=sample_shape,
seed=self.seed)
chex.assert_equal_shape([samples, tfp_samples])
def test_raises_on_event_ndims_without_log_det(self, ndims_in, ndims_out):
with self.assertRaises(ValueError):
lambda_bijector.Lambda(forward=lambda x: x,
event_ndims_in=ndims_in,
event_ndims_out=ndims_out)
def test_raises_on_log_det_without_event_ndims(self):
with self.assertRaises(ValueError):
lambda_bijector.Lambda(
forward=lambda x: x,
forward_log_det_jacobian=lambda x: jnp.zeros_like(x[:-1]),
event_ndims_in=None)
def test_raises_on_both_none(self):
with self.assertRaises(ValueError):
lambda_bijector.Lambda(forward=None, inverse=None)
