def test_single_example_correct(self):
# The correct label should have low cross entropy.
correct_xent = util.mnist_cross_entropy(real_digit(), one_hot_real())
# The incorrect label should have high cross entropy.
wrong_xent = util.mnist_cross_entropy(real_digit(), one_hot1())
# A random digit should have medium cross entropy for any label.
fake_xent1 = util.mnist_cross_entropy(fake_digit(), one_hot_real())
fake_xent6 = util.mnist_cross_entropy(fake_digit(), one_hot1())
with self.cached_session() as sess:
self.assertNear(0.00996, sess.run(correct_xent), 1e-5)
self.assertNear(18.63073, sess.run(wrong_xent), 1e-5)
self.assertNear(2.2, sess.run(fake_xent1), 1e-1)
self.assertNear(2.2, sess.run(fake_xent6), 1e-1)
def _disabled_test_minibatch_correct(self):
"""Tests correctness of the mnist_cross_entropy function with batches."""
# Disabled since it requires loading the tfhub MNIST module.
# Reorded minibatches should have the same value.
xent1 = util.mnist_cross_entropy(
tf.concat([real_digit(), real_digit(), fake_digit()], 0),
tf.concat([one_hot_real(), one_hot1(), one_hot1()], 0))
xent2 = util.mnist_cross_entropy(
tf.concat([real_digit(), fake_digit(), real_digit()], 0),
tf.concat([one_hot_real(), one_hot1(), one_hot1()], 0))
with self.cached_session() as sess:
self.assertNear(6.972539, sess.run(xent1), 1e-5)
self.assertNear(sess.run(xent1), sess.run(xent2), 1e-5)
def test_minibatch_correct(self):
if tf.executing_eagerly():
# `run_image_classifier` doesn't work in eager.
return
# Reorded minibatches should have the same value.
xent1 = util.mnist_cross_entropy(
tf.concat([real_digit(), real_digit(), fake_digit()], 0),
tf.concat([one_hot_real(), one_hot1(), one_hot1()], 0))
xent2 = util.mnist_cross_entropy(
tf.concat([real_digit(), fake_digit(), real_digit()], 0),
tf.concat([one_hot_real(), one_hot1(), one_hot1()], 0))
with self.cached_session() as sess:
self.assertNear(6.972539, sess.run(xent1), 1e-5)
self.assertNear(sess.run(xent1), sess.run(xent2), 1e-5)
def _disabled_test_single_example_correct(self):
"""Tests correctness of the mnist_cross_entropy function."""
# Disabled since it requires loading the tfhub MNIST module.
# The correct label should have low cross entropy.
correct_xent = util.mnist_cross_entropy(real_digit(), one_hot_real())
# The incorrect label should have high cross entropy.
wrong_xent = util.mnist_cross_entropy(real_digit(), one_hot1())
# A random digit should have medium cross entropy for any label.
fake_xent1 = util.mnist_cross_entropy(fake_digit(), one_hot_real())
fake_xent6 = util.mnist_cross_entropy(fake_digit(), one_hot1())
with self.cached_session() as sess:
self.assertNear(0.00996, sess.run(correct_xent), 1e-5)
self.assertNear(18.63073, sess.run(wrong_xent), 1e-5)
self.assertNear(2.2, sess.run(fake_xent1), 1e-1)
self.assertNear(2.2, sess.run(fake_xent6), 1e-1)
def test_minibatch_correct(self):
# Reorded minibatches should have the same value.
xent1 = util.mnist_cross_entropy(
tf.concat([real_digit(), real_digit(),
fake_digit()], 0),
tf.concat([one_hot_real(), one_hot1(),
one_hot1()], 0))
xent2 = util.mnist_cross_entropy(
tf.concat([real_digit(), fake_digit(),
real_digit()], 0),
tf.concat([one_hot_real(), one_hot1(),
one_hot1()], 0))
with self.cached_session() as sess:
self.assertNear(6.972539, sess.run(xent1), 1e-5)
self.assertNear(sess.run(xent1), sess.run(xent2), 1e-5)
def test_deterministic(self):
xent = util.mnist_cross_entropy(real_digit(), one_hot_real())
with self.cached_session() as sess:
ent1 = sess.run(xent)
ent2 = sess.run(xent)
self.assertEqual(ent1, ent2)
with self.cached_session() as sess:
ent3 = sess.run(xent)
self.assertEqual(ent1, ent3)
def test_deterministic(self, mock_tfhub_load):
mock_tfhub_load.return_value = fake_logit_fn
xent = util.mnist_cross_entropy(real_digit(), one_hot_real())
with self.cached_session() as sess:
ent1 = sess.run(xent)
ent2 = sess.run(xent)
self.assertEqual(ent1, ent2)
with self.cached_session() as sess:
ent3 = sess.run(xent)
self.assertEqual(ent1, ent3)
def test_deterministic(self):
if tf.executing_eagerly():
# `run_image_classifier` doesn't work in eager.
return
xent = util.mnist_cross_entropy(real_digit(), one_hot_real())
with self.cached_session() as sess:
ent1 = sess.run(xent)
ent2 = sess.run(xent)
self.assertEqual(ent1, ent2)
with self.cached_session() as sess:
ent3 = sess.run(xent)
self.assertEqual(ent1, ent3)
def evaluate(hparams, run_eval_loop=True):
"""Runs an evaluation loop.
Args:
hparams: An HParams instance containing the eval hyperparameters.
run_eval_loop: Whether to run the full eval loop. Set to False for testing.
"""
with tf.compat.v1.name_scope('inputs'):
noise, one_hot_labels = _get_generator_inputs(
hparams.num_images_per_class, NUM_CLASSES, hparams.noise_dims)
# Generate images.
with tf.compat.v1.variable_scope(
'Generator'): # Same scope as in train job.
images = networks.conditional_generator((noise, one_hot_labels),
is_training=False)
# Visualize images.
reshaped_img = tfgan.eval.image_reshaper(
images, num_cols=hparams.num_images_per_class)
tf.compat.v1.summary.image('generated_images', reshaped_img, max_outputs=1)
# Calculate evaluation metrics.
tf.compat.v1.summary.scalar(
'MNIST_Classifier_score',
util.mnist_score(images, hparams.classifier_filename))
tf.compat.v1.summary.scalar(
'MNIST_Cross_entropy',
util.mnist_cross_entropy(images, one_hot_labels,
hparams.classifier_filename))
# Write images to disk.
image_write_ops = None
if hparams.write_to_disk:
image_write_ops = tf.io.write_file(
'%s/%s' % (hparams.eval_dir, 'conditional_gan.png'),
tf.image.encode_png(
data_provider.float_image_to_uint8(reshaped_img[0])))
# For unit testing, use `run_eval_loop=False`.
if not run_eval_loop:
return
evaluation.evaluate_repeatedly(
hparams.checkpoint_dir,
hooks=[
evaluation.SummaryAtEndHook(hparams.eval_dir),
evaluation.StopAfterNEvalsHook(1)
],
eval_ops=image_write_ops,
max_number_of_evaluations=hparams.max_number_of_evaluations)
def test_any_batch_size(self, mock_tfhub_load):
mock_tfhub_load.return_value = fake_logit_fn
# Create a graph since placeholders don't work in eager execution mode.
with tf.Graph().as_default():
num_classes = 10
one_label = np.array([[1] + [0] * (num_classes - 1)])
inputs = tf.placeholder(tf.float32, shape=[None, 28, 28, 1])
one_hot_label = tf.placeholder(tf.int32, shape=[None, num_classes])
entropy = util.mnist_cross_entropy(inputs, one_hot_label)
for batch_size in [4, 16, 30]:
with self.cached_session() as sess:
sess.run(entropy, feed_dict={
inputs: np.zeros([batch_size, 28, 28, 1]),
one_hot_label: np.concatenate([one_label] * batch_size)})
def test_any_batch_size(self):
if tf.executing_eagerly():
# Placeholders don't work in eager execution mode.
return
num_classes = 10
one_label = np.array([[1] + [0] * (num_classes - 1)])
inputs = tf.compat.v1.placeholder(tf.float32, shape=[None, 28, 28, 1])
one_hot_label = tf.compat.v1.placeholder(
tf.int32, shape=[None, num_classes])
entropy = util.mnist_cross_entropy(inputs, one_hot_label)
for batch_size in [4, 16, 30]:
with self.cached_session() as sess:
sess.run(entropy, feed_dict={
inputs: np.zeros([batch_size, 28, 28, 1]),
one_hot_label: np.concatenate([one_label] * batch_size)})
