def test_variable_creation(self):
weight = np.ones((3, 3))
var1 = tf_utils.create_variable(weight, name='var1', shape=(3, 3))
self.assertShapesEqual(var1.shape, (3, 3))
var2 = tf_utils.create_variable(5, name='var2', shape=(4, 3))
self.assertShapesEqual(var2.shape, (4, 3))
np.testing.assert_array_almost_equal(self.eval(var2), 5 * np.ones(
(4, 3)))
initializer = init.Normal()
var3 = tf_utils.create_variable(initializer, name='var3', shape=(4, 7))
self.assertShapesEqual(var3.shape, (4, 7))
weight = tf.Variable(np.ones((3, 3)), dtype=tf.float32)
var4 = tf_utils.create_variable(weight, name='var4', shape=(3, 3))
self.assertShapesEqual(var4.shape, (3, 3))
self.assertIs(var4, weight)
weight = np.ones((3, 4))
with self.assertRaisesRegexp(ValueError, "Cannot create variable"):
tf_utils.create_variable(weight, name='var5', shape=(3, 3))
weight = tf.Variable(np.ones((4, 3)), dtype=tf.float32)
with self.assertRaisesRegexp(ValueError, "Cannot create variable"):
tf_utils.create_variable(weight, name='var6', shape=(3, 3))
def variable(self, value, name, shape=None, trainable=True):
layer_name = 'layer/{layer_name}/{parameter_name}'.format(
layer_name=self.name, parameter_name=name.replace('_', '-'))
self.variables[name] = tf_utils.create_variable(
value, layer_name, shape, trainable)
return self.variables[name]
def init_functions(self):
with tf.variable_scope('rbm'):
self.weight = tf_utils.create_variable(
value=self.weight,
name='weight',
shape=(self.n_visible, self.n_hidden),
trainable=False,
)
self.hidden_bias = tf_utils.create_variable(
value=self.hidden_bias,
name='hidden-bias',
shape=(self.n_hidden, ),
trainable=False,
)
self.visible_bias = tf_utils.create_variable(
value=self.visible_bias,
name='visible-bias',
shape=(self.n_visible, ),
trainable=False,
)
self.h_samples = tf.Variable(
tf.zeros([self.batch_size, self.n_hidden]),
name="hidden-samples",
dtype=tf.float32,
trainable=False,
)
self.network_input = tf.placeholder(
tf.float32,
(None, self.n_visible),
name="network-input",
)
self.network_hidden_input = tf.placeholder(
tf.float32,
(None, self.n_hidden),
name="network-hidden-input",
)