def __init__(self,
model_params,
logdir,
tasks,
task_types,
train=True,
verbosity=None):
"""Constructs the computational graph.
Args:
train: whether model is in train mode
model_params: dictionary of model parameters
logdir: Location to save data
This function constructs the computational graph for the model. It relies
subclassed methods (build/cost) to construct specific graphs.
"""
self.graph = tf.Graph()
self.model_params = model_params
self.logdir = logdir
self.tasks = tasks
self.task_types = task_types
self.num_tasks = len(task_types)
self.verbosity = verbosity
# Lazily created by _get_shared_session().
self._shared_session = None
# Guard variable to make sure we don't Restore() this model
# from a disk checkpoint more than once.
self._restored_model = False
# Cache of TensorFlow scopes, to prevent '_1' appended scope names
# when subclass-overridden methods use the same scopes.
self._name_scopes = {}
# Path to save checkpoint files, which matches the
# replicated supervisor's default path.
self._save_path = os.path.join(logdir, 'model.ckpt')
with self.graph.as_default():
model_ops.set_training(train)
self.placeholder_root = 'placeholders'
with tf.name_scope(self.placeholder_root) as scope:
self.placeholder_scope = scope
self.setup()
if train:
self.add_training_cost()
self.merge_updates()
else:
self.add_output_ops() # add softmax heads
def _CheckBatchNormalization(self, features, convolution, mean, variance,
mask=None):
model_ops.set_training(True)
epsilon = 0.001
with self.test_session() as sess:
features_t = tf.constant(features, dtype=tf.float32)
batch_norm_t = model_ops.BatchNormalize(
features_t, convolution=convolution, epsilon=epsilon, mask=mask)
sess.run(tf.initialize_all_variables())
batch_norm, beta, gamma = sess.run(
[batch_norm_t] + tf.trainable_variables())
expected = gamma * (features - mean) / np.sqrt(variance + epsilon) + beta
self.assertAllClose(batch_norm, np.ma.filled(expected, 0),
rtol=1e-5, atol=1e-5)
def __init__(self, model_params, logdir, task_types, train=True,
verbosity=None):
"""Constructs the computational graph.
Args:
train: whether model is in train mode
model_params: dictionary of model parameters
logdir: Location to save data
This function constructs the computational graph for the model. It relies
subclassed methods (build/cost) to construct specific graphs.
"""
self.graph = tf.Graph()
self.model_params = model_params
self.logdir = logdir
self.task_types = task_types
self.num_tasks = len(task_types)
self.verbosity = verbosity
# Lazily created by _get_shared_session().
self._shared_session = None
# Guard variable to make sure we don't Restore() this model
# from a disk checkpoint more than once.
self._restored_model = False
# Cache of TensorFlow scopes, to prevent '_1' appended scope names
# when subclass-overridden methods use the same scopes.
self._name_scopes = {}
# Path to save checkpoint files, which matches the
# replicated supervisor's default path.
self._save_path = os.path.join(logdir, 'model.ckpt')
with self.graph.as_default():
model_ops.set_training(train)
self.placeholder_root = 'placeholders'
with tf.name_scope(self.placeholder_root) as scope:
self.placeholder_scope = scope
self.valid = tf.placeholder(tf.bool,
shape=[model_params["batch_size"]],
name='valid')
self.setup()
if train:
self.add_training_cost()
self.merge_updates()
else:
self.add_output_ops() # add softmax heads
def testBatchNormalizationInference(self):
# create a simple batch-normalized model
model_ops.set_training(True)
epsilon = 0.001
decay = 0.95
checkpoint = os.path.join(self.root, 'my-checkpoint')
with self.test_session() as sess:
features = np.random.random((2, 3, 2, 3))
features_t = tf.constant(features, dtype=tf.float32)
# create variables for beta, gamma, and moving mean and variance
model_ops.BatchNormalize(features_t,
convolution=False,
epsilon=epsilon,
decay=decay)
sess.run(tf.initialize_all_variables())
updates = tf.group(
*tf.get_default_graph().get_collection('updates'))
sess.run(updates) # update moving mean and variance
expected_mean, expected_variance, _, _ = tf.all_variables()
expected_mean = expected_mean.eval()
expected_variance = expected_variance.eval()
# save a checkpoint
saver = tf.train.Saver()
saver.save(sess, checkpoint)
super(ModelOpsTest, self).setUp() # reset the default graph
# check that the moving mean and variance are used for evaluation
# get a new set of features to verify that the correct mean and var are used
model_ops.set_training(False)
with self.test_session() as sess:
new_features = np.random.random((2, 3, 2, 3))
new_features_t = tf.constant(new_features, dtype=tf.float32)
batch_norm_t = model_ops.BatchNormalize(new_features_t,
convolution=False,
epsilon=epsilon,
decay=decay)
saver = tf.train.Saver()
saver.restore(sess, checkpoint)
batch_norm, mean, variance, beta, gamma = sess.run(
[batch_norm_t] + tf.all_variables())
self.assertAllClose(mean, expected_mean)
self.assertAllClose(variance, expected_variance)
expected = (gamma *
(new_features - mean) / np.sqrt(variance + epsilon) +
beta)
self.assertAllClose(batch_norm, expected)
def testBatchNormalizationInference(self):
# create a simple batch-normalized model
model_ops.set_training(True)
epsilon = 0.001
decay = 0.95
checkpoint = os.path.join(self.root, 'my-checkpoint')
with self.test_session() as sess:
features = np.random.random((2, 3, 2, 3))
features_t = tf.constant(features, dtype=tf.float32)
# create variables for beta, gamma, and moving mean and variance
model_ops.BatchNormalize(
features_t, convolution=False, epsilon=epsilon, decay=decay)
sess.run(tf.initialize_all_variables())
updates = tf.group(*tf.get_default_graph().get_collection('updates'))
sess.run(updates) # update moving mean and variance
expected_mean, expected_variance, _, _ = tf.all_variables()
expected_mean = expected_mean.eval()
expected_variance = expected_variance.eval()
# save a checkpoint
saver = tf.train.Saver()
saver.save(sess, checkpoint)
super(ModelOpsTest, self).setUp() # reset the default graph
# check that the moving mean and variance are used for evaluation
# get a new set of features to verify that the correct mean and var are used
model_ops.set_training(False)
with self.test_session() as sess:
new_features = np.random.random((2, 3, 2, 3))
new_features_t = tf.constant(new_features, dtype=tf.float32)
batch_norm_t = model_ops.BatchNormalize(
new_features_t, convolution=False, epsilon=epsilon, decay=decay)
saver = tf.train.Saver()
saver.restore(sess, checkpoint)
batch_norm, mean, variance, beta, gamma = sess.run(
[batch_norm_t] + tf.all_variables())
self.assertAllClose(mean, expected_mean)
self.assertAllClose(variance, expected_variance)
expected = (gamma * (new_features - mean) /
np.sqrt(variance + epsilon) + beta)
self.assertAllClose(batch_norm, expected)
def _CheckBatchNormalization(self,
features,
convolution,
mean,
variance,
mask=None):
model_ops.set_training(True)
epsilon = 0.001
with self.test_session() as sess:
features_t = tf.constant(features, dtype=tf.float32)
batch_norm_t = model_ops.BatchNormalize(features_t,
convolution=convolution,
epsilon=epsilon,
mask=mask)
sess.run(tf.initialize_all_variables())
batch_norm, beta, gamma = sess.run([batch_norm_t] +
tf.trainable_variables())
expected = gamma * (features - mean) / np.sqrt(variance +
epsilon) + beta
self.assertAllClose(batch_norm,
np.ma.filled(expected, 0),
rtol=1e-5,
atol=1e-5)
def GetModel(self, train=True):
model_ops.set_training(train)
# dummy variable for testing Restore
tf.Variable(tf.constant(10.0, shape=[1]), name='v0')
def GetModel(self, train=True):
model_ops.set_training(train)
# dummy variable for testing Restore
tf.Variable(tf.constant(10.0, shape=[1]), name='v0')