def reconstruction(self):
"""Reconstruct the scale input."""
quan_x = ops.soft_quantile_normalization(
self.x,
self.quantiles_prime,
target_weights=tf.nn.softmax(self._b_prime, axis=-1),
**self._kwargs)
self._nmf_factorizer(quan_x, random_seed=0)
self.inner_kl.append(self._nmf_factorizer.losses)
return ops.soft_quantile_normalization(self._nmf_factorizer.reconstruction,
self.quantiles,
target_weights=self.target_weights,
**self._kwargs)
def sliced_reconstruction(self, rows=None, cols=None):
"""Reconstructs the scale input."""
return ops.soft_quantile_normalization(
self.sliced_uv(rows, cols),
self.sliced_quantiles(rows),
target_weights=self.sliced_target_weights(rows),
**self._kwargs)
def test_soft_quantile_normalization(self):
x = tf.constant([1.2, 1.3, 1.5, -4.0, 1.8, 2.4, -1.0])
target = tf.cumsum(tf.ones(x.shape[0]))
xn = ops.soft_quantile_normalization(x, target)
# Make sure that the order of x and xn are identical
self.assertAllEqual(tf.argsort(x), tf.argsort(xn))
# Make sure that the values of xn and target are close.
self.assertAllClose(tf.sort(target), tf.sort(xn), atol=1e-1)
def call(self, inputs):
target_cdf = tf.math.cumsum(tf.math.softmax(self.w), axis=self._axis)
return ops.soft_quantile_normalization(inputs,
target_cdf,
axis=self._axis,
**self._kwargs)
