def testIntegrity(self):
"""Checks a multi_gpu call on CapsuleModel builds the desired graph.
With the correct inference graph, multi_gpu is able to call inference
multiple times without any increase in number of trainable variables or a
duplication error.
"""
with tf.Graph().as_default():
test_model = capsule_model.CapsuleModel(self.hparams)
toy_image = np.reshape(np.arange(32 * 32), (1, 1, 32, 32))
input_image = tf.constant(toy_image, dtype=tf.float32)
features = {
'height': 32,
'depth': 1,
'images': input_image,
'labels': tf.one_hot([2], 10),
'num_classes': 10,
'num_targets': 1,
}
_, tower_output = test_model.multi_gpu(
[features, features, features], 3)
trainable_vars = tf.get_collection(
tf.GraphKeys.TRAINABLE_VARIABLES)
self.assertEqual(len(trainable_vars), 6)
_, classes = tower_output[0].logits.get_shape()
self.assertEqual(10, classes.value)
def testInferenceWithRemake(self):
"""Checks the correct shape of remakes and total number of variables.
The reconstruction should have same shape as input. Each remake network
should declare 6 sets of variables (weight and bias) and different targets
should share the variables.
"""
with tf.Graph().as_default():
self.hparams.parse('remake=True,verbose=True')
test_model = capsule_model.CapsuleModel(self.hparams)
toy_image = np.reshape(np.arange(32 * 32), (1, 1, 32, 32))
input_image = tf.constant(toy_image, dtype=tf.float32)
features = {
'height': 32,
'depth': 1,
'images': input_image,
'recons_image': input_image,
'spare_image': input_image,
'recons_label': tf.constant([2]),
'spare_label': tf.constant([2]),
'num_targets': 2,
'num_classes': 10,
}
output = test_model.inference(features)
trainable_vars = tf.get_collection(
tf.GraphKeys.TRAINABLE_VARIABLES)
self.assertEqual(len(trainable_vars), 12)
remake_1, remake_2 = output.remakes
self.assertEqual(32 * 32, remake_1.get_shape()[1].value)
self.assertEqual(32 * 32, remake_2.get_shape()[1].value)
def testInference(self):
"""
Checks the correct shape of capsule output and total number of
variables.
The output logit shape should be [batch, 10]. Also each layer should
declare 2 sets of variables (weight and bias), therefore single call to
inference declares 6 variables for a total of 3 layers.
"""
with tf.Graph().as_default():
test_model = capsule_model.CapsuleModel(self.hparams)
toy_image = np.reshape(np.arange(32 * 32), (1, 1, 32, 32))
input_image = tf.constant(toy_image, dtype=tf.float32)
features = {
'height': 32,
'depth': 1,
'num_classes': 10,
'images': input_image
}
output = test_model.inference(features)
trainable_vars = tf.get_collection(
tf.GraphKeys.TRAINABLE_VARIABLES)
self.assertEqual(len(trainable_vars), 6)
_, classes = output.logits.get_shape()
self.assertEqual(10, classes.value)
def testBuildCapsule(self):
"""Checks the correct shape of capsule output and total number of variables.
The output shape should be [batch, 10, 16]. Also each capsule layer should
declare 2 sets of variables (weight and bias), therefore single call to
_build_capsule declares 4 variables for a total of 2 capsule layers.
"""
with tf.Graph().as_default():
test_model = capsule_model.CapsuleModel(self.hparams)
toy_input = np.reshape(np.arange(256 * 14 * 14),
(1, 1, 256, 14, 14))
input_tensor = tf.constant(toy_input, dtype=tf.float32)
output = test_model._build_capsule(input_tensor, 10)
trainable_vars = tf.get_collection(
tf.GraphKeys.TRAINABLE_VARIABLES)
self.assertEqual(len(trainable_vars), 4)
_, capsules, atoms = output.get_shape()
self.assertListEqual([10, 16], [capsules.value, atoms.value])