def testSample(self):
"""Ensure samples can be drawn."""
with tf.Graph().as_default(), self.test_session() as sess:
logits = np.asarray([
[0., 0., 0.], #
[1., -1., 0.]
]).astype(np.float32)
loss = loss_functions.CategoricalLogitsNegativeLogProbLoss(
tf.constant(logits))
sample = loss.sample(42)
sample = sess.run(sample)
self.assertEqual(sample.shape, (2,))
def testEvaluateOnSample(self):
"""Ensure log probability of a sample can be drawn."""
with tf.Graph().as_default(), self.test_session() as sess:
logits = np.asarray([
[0., 0., 0.], #
[1., -1., 0.]
]).astype(np.float32)
loss = loss_functions.CategoricalLogitsNegativeLogProbLoss(
tf.constant(logits))
neg_log_prob = loss.evaluate_on_sample(42)
# Simply ensure this doesn't crash. As the output is random, it's
# difficult to say if the output is correct or not...
neg_log_prob = sess.run(neg_log_prob)
def testMultiplyFisherBatch(self):
with tf.Graph().as_default(), self.test_session() as sess:
logits = np.array([[1., 2., 3.], [4., 6., 8.]])
loss = loss_functions.CategoricalLogitsNegativeLogProbLoss(logits)
vector = np.array([[1., 2., 3.], [5., 3., 1.]])
na = np.newaxis
probs = np.exp(logits - np.logaddexp.reduce(logits, axis=-1,
keepdims=True))
fishers = probs[..., na] * np.eye(3) - probs[..., na] * probs[..., na, :]
result = loss.multiply_fisher(vector)
expected_result = np.matmul(vector[..., na, :], fishers)[..., 0, :]
self.assertEqual(sess.run(result).shape, logits.shape)
self.assertAllClose(expected_result, sess.run(result))
def testMultiplyFisherSingleVector(self):
with tf.Graph().as_default(), self.test_session() as sess:
logits = np.array([1., 2., 3.])
loss = loss_functions.CategoricalLogitsNegativeLogProbLoss(logits)
# the LossFunction.multiply_fisher docstring only says it supports the
# case where the vector is the same shape as the input natural parameters
# (i.e. the logits here), but here we also test leading dimensions
vector = np.array([1., 2., 3.])
vectors = [vector, vector.reshape(1, -1), np.stack([vector] * 4)]
probs = np.exp(logits - np.logaddexp.reduce(logits))
fisher = np.diag(probs) - np.outer(probs, probs)
for vector in vectors:
result = loss.multiply_fisher(vector)
expected_result = np.dot(vector, fisher)
self.assertAllClose(expected_result, sess.run(result))
def testEvaluateOnTargets(self):
"""Ensure log probability can be evaluated correctly."""
with tf.Graph().as_default(), self.test_session() as sess:
logits = np.asarray([
[0., 0., 0.], #
[1., -1., 0.]
]).astype(np.float32)
targets = np.asarray([2, 1]).astype(np.int32)
loss = loss_functions.CategoricalLogitsNegativeLogProbLoss(
tf.constant(logits), targets=tf.constant(targets))
neg_log_prob = loss.evaluate()
neg_log_prob = sess.run(neg_log_prob)
# Calculate explicit log probability of targets.
probs = np.exp(logits) / np.sum(np.exp(logits), axis=1, keepdims=True)
log_probs = np.log([
probs[0, targets[0]], #
probs[1, targets[1]]
])
expected_log_prob = np.sum(log_probs)
self.assertAllClose(neg_log_prob, -expected_log_prob)