def test_softsort(self):
# Tests that the values are sorted (epsilon being small enough)
x = tf.constant([3, 4, 1, 5, 2], dtype=tf.float32)
eps = 1e-3
sinkhorn_threshold = 1e-3
values = ops.softsort(x,
direction='ASCENDING',
epsilon=eps,
sinkhorn_threshold=sinkhorn_threshold)
self.assertEqual(values.shape, x.shape)
self.assertAllGreater(np.diff(values), 0.0)
# Since epsilon is not very small, we cannot expect to retrieve the sorted
# values with high precision.
tolerance = 1e-1
self.assertAllClose(tf.sort(x), values, tolerance, tolerance)
# Test descending sort.
direction = 'DESCENDING'
values = ops.softsort(x,
direction=direction,
epsilon=eps,
sinkhorn_threshold=sinkhorn_threshold)
self.assertEqual(values.shape, x.shape)
self.assertAllLess(np.diff(values), 0.0)
self.assertAllClose(tf.sort(x, direction=direction), values, tolerance,
tolerance)
def test_softsort(self, topk):
# Tests that the values are sorted (epsilon being small enough)
x = tf.constant([3, 4, 1, 5, 2, 9, 12, 11, 8, 15], dtype=tf.float32)
eps = 1e-3
sinkhorn_threshold = 1e-3
values = ops.softsort(x,
direction='ASCENDING',
topk=topk,
epsilon=eps,
threshold=sinkhorn_threshold)
expect_shape = x.shape if topk is None else (topk, )
self.assertEqual(values.shape, expect_shape)
self.assertAllGreater(np.diff(values), 0.0)
# Since epsilon is not very small, we cannot expect to retrieve the sorted
# values with high precision.
tolerance = 1e-1
expected_values = tf.sort(x)
if topk is not None:
expected_values = expected_values[:topk]
self.assertAllClose(expected_values, values, tolerance, tolerance)
# Test descending sort.
direction = 'DESCENDING'
values = ops.softsort(x,
direction=direction,
topk=topk,
epsilon=eps,
threshold=sinkhorn_threshold)
expected_values = tf.sort(x, direction=direction)
if topk is not None:
expected_values = expected_values[:topk]
self.assertEqual(values.shape, expect_shape)
self.assertAllLess(np.diff(values), 0.0)
self.assertAllClose(expected_values, values, tolerance, tolerance)
def call(self, y_true, y_pred):
error = tf.pow(tf.abs(tf.squeeze(y_pred) - y_true), self._power)
target_weights, target_index = self._get_target_weights_and_indices()
quantiles = ops.softsort(error,
axis=0,
target_weights=target_weights,
**self._kwargs)
return tf.gather(quantiles, target_index, axis=0)
def call(self, y_true, y_pred):
error = tf.pow(tf.abs(tf.squeeze(y_pred) - y_true), self._power)
target_weights = [
self._start_quantile,
self._end_quantile - self._start_quantile,
1.0 - self._end_quantile
]
quantiles = ops.softsort(
error, axis=0, target_weights=target_weights, **self._kwargs)
return quantiles[1]
def call(self, inputs):
outputs = ops.softsort(
inputs, axis=self._axis, topk=self._topk, **self._kwargs)
if self._topk is not None:
return outputs
# For some reason, when doing a full sort, tf has a hard time computing the
# shape of the output tensor. To specify the shape to tf we use the trick
# to call tf.reshape.
return tf.reshape(outputs, tf.shape(inputs))