def test_logit_0d(self):
with self.test_session():
x = tf.constant(0.01)
self.assertAllClose(logit(x).eval(), -4.5951198501345898)
x = tf.constant(0.25)
self.assertAllClose(logit(x).eval(), -1.0986122886681096)
x = tf.constant(0.5)
self.assertAllEqual(logit(x).eval(), 0.0)
x = tf.constant(0.75)
self.assertAllClose(logit(x).eval(), 1.0986122886681096)
x = tf.constant(0.99)
self.assertAllClose(logit(x).eval(), 4.5951198501345898)
def test_logit_1d(self):
with self.test_session():
x = tf.constant([0.01, 0.01])
self.assertAllClose(logit(x).eval(), [-4.5951198501345898, -4.5951198501345898])
x = tf.constant([0.25, 0.25])
self.assertAllClose(logit(x).eval(), [-1.0986122886681096, -1.0986122886681096])
x = tf.constant([0.5, 0.5])
self.assertAllEqual(logit(x).eval(), [0.0, 0.0])
x = tf.constant([0.75, 0.75])
self.assertAllClose(logit(x).eval(), [1.0986122886681096, 1.0986122886681096])
x = tf.constant([0.99, 0.99])
self.assertAllClose(logit(x).eval(), [4.5951198501345898, 4.5951198501345898])
def test_logit_0d(self):
with self.test_session():
x = tf.constant(0.01)
self.assertAllClose(logit(x).eval(), -4.5951198501345898)
x = tf.constant(0.25)
self.assertAllClose(logit(x).eval(), -1.0986122886681096)
x = tf.constant(0.5)
self.assertAllEqual(logit(x).eval(), 0.0)
x = tf.constant(0.75)
self.assertAllClose(logit(x).eval(), 1.0986122886681096)
x = tf.constant(0.99)
self.assertAllClose(logit(x).eval(), 4.5951198501345898)
def test_logit_1d(self):
with self.test_session():
x = tf.constant([0.01, 0.01])
self.assertAllClose(
logit(x).eval(), [-4.5951198501345898, -4.5951198501345898])
x = tf.constant([0.25, 0.25])
self.assertAllClose(
logit(x).eval(), [-1.0986122886681096, -1.0986122886681096])
x = tf.constant([0.5, 0.5])
self.assertAllEqual(logit(x).eval(), [0.0, 0.0])
x = tf.constant([0.75, 0.75])
self.assertAllClose(
logit(x).eval(), [1.0986122886681096, 1.0986122886681096])
x = tf.constant([0.99, 0.99])
self.assertAllClose(
logit(x).eval(), [4.5951198501345898, 4.5951198501345898])
def sample_n(self, n, seed=None):
if self.n != 1:
logits = logit(tf.ones(self.n, dtype=tf.float32) /
tf.cast(self.n, dtype=tf.float32))
cat = tf.contrib.distributions.Categorical(logits=logits)
indices = cat.sample_n(n, seed)
return tf.gather(self._params, indices)
else:
multiples = tf.concat(0, [tf.expand_dims(n, 0),
[1] * len(self.get_event_shape())])
return tile(self._params, multiples)
def _sample_n(self, n, seed=None):
if self.n != 1:
logits = logit(
tf.ones(self.n, dtype=tf.float32) /
tf.cast(self.n, dtype=tf.float32))
cat = tf.contrib.distributions.Categorical(logits=logits)
indices = cat._sample_n(n, seed)
return tf.gather(self._params, indices)
else:
multiples = tf.concat(
[tf.expand_dims(n, 0), [1] * len(self.get_event_shape())], 0)
return tile(self._params, multiples)
def binary_crossentropy(y_true, y_pred):
"""
Parameters
----------
y_true : tf.Tensor
Tensor of 0s and 1s.
y_pred : tf.Tensor
Tensor of probabilities.
"""
y_true = tf.cast(y_true, tf.float32)
y_pred = logit(tf.cast(y_pred, tf.float32))
return tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(y_pred, y_true))
def binary_crossentropy(y_true, y_pred):
"""
Parameters
----------
y_true : tf.Tensor
Tensor of 0s and 1s.
y_pred : tf.Tensor
Tensor of probabilities.
"""
y_true = tf.cast(y_true, tf.float32)
y_pred = logit(tf.cast(y_pred, tf.float32))
return tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(y_pred, y_true))
def test_contraint_raises(self):
with self.test_session():
x = tf.constant([0.01, -20])
with self.assertRaisesOpError("Condition"):
logit(x).eval()
x = tf.constant([0.01, 20])
with self.assertRaisesOpError("Condition"):
logit(x).eval()
x = tf.constant([0.01, np.inf])
with self.assertRaisesOpError("Condition"):
logit(x).eval()
x = tf.constant([0.01, np.nan])
with self.assertRaisesOpError("Condition"):
logit(x).eval()
def test_contraint_raises(self):
with self.test_session():
x = tf.constant([0.01, -20])
with self.assertRaisesOpError('Condition'):
logit(x).eval()
x = tf.constant([0.01, 20])
with self.assertRaisesOpError('Condition'):
logit(x).eval()
x = tf.constant([0.01, np.inf])
with self.assertRaisesOpError('Condition'):
logit(x).eval()
x = tf.constant([0.01, np.nan])
with self.assertRaisesOpError('Condition'):
logit(x).eval()
def sparse_categorical_crossentropy(y_true, y_pred):
"""Multi-class cross entropy. Label {0, 1, .., K-1} representation
for ``y_true.``
Parameters
----------
y_true : tf.Tensor
Tensor of integers {0, 1, ..., K-1}.
y_pred : tf.Tensor
Tensor of probabilities, with shape ``(y_true.get_shape(), K)``.
The outermost dimension are the categorical probabilities for
that data point.
"""
y_true = tf.cast(y_true, tf.int64)
y_pred = logit(tf.cast(y_pred, tf.float32))
return tf.reduce_mean(tf.nn.sparse_softmax_cross_entropy_with_logits(y_pred, y_true))
def categorical_crossentropy(y_true, y_pred):
"""Multi-class cross entropy. One-hot representation for ``y_true``.
Parameters
----------
y_true : tf.Tensor
Tensor of 0s and 1s, where the outermost dimension of size K
has only one 1 per row.
y_pred : tf.Tensor
Tensor of probabilities, with same shape as y_true.
The outermost dimension denote the categorical probabilities for
that data point per row.
"""
y_true = tf.cast(y_true, tf.float32)
y_pred = logit(tf.cast(y_pred, tf.float32))
return tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(y_pred, y_true))
def categorical_crossentropy(y_true, y_pred):
"""Multi-class cross entropy. One-hot representation for ``y_true``.
Parameters
----------
y_true : tf.Tensor
Tensor of 0s and 1s, where the outermost dimension of size K
has only one 1 per row.
y_pred : tf.Tensor
Tensor of probabilities, with same shape as y_true.
The outermost dimension denote the categorical probabilities for
that data point per row.
"""
y_true = tf.cast(y_true, tf.float32)
y_pred = logit(tf.cast(y_pred, tf.float32))
return tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(y_pred, y_true))
def sparse_categorical_crossentropy(y_true, y_pred):
"""Multi-class cross entropy. Label {0, 1, .., K-1} representation
for ``y_true.``
Parameters
----------
y_true : tf.Tensor
Tensor of integers {0, 1, ..., K-1}.
y_pred : tf.Tensor
Tensor of probabilities, with shape ``(y_true.get_shape(), K)``.
The outermost dimension are the categorical probabilities for
that data point.
"""
y_true = tf.cast(y_true, tf.int64)
y_pred = logit(tf.cast(y_pred, tf.float32))
return tf.reduce_mean(tf.nn.sparse_softmax_cross_entropy_with_logits(
logits=y_pred, labels=y_true))
