def testShape(self):
kernel_shape = [5, 5]
shape = kernel_shape + [2, 2] # The input has to be 4-dimensional.
gradients = tf.random_normal(shape)
net = networks.KernelDeepLSTM(layers=(1, 1), kernel_shape=kernel_shape)
state = net.initial_state_for_inputs(gradients)
update, _ = net(gradients, state)
self.assertEqual(update.get_shape().as_list(), shape)
def testTrainable(self):
"""Tests the network contains trainable variables."""
kernel_shape = [5, 5]
shape = kernel_shape + [2, 2] # The input has to be 4-dimensional.
gradients = tf.random_normal(shape)
net = networks.KernelDeepLSTM(layers=(1, ), kernel_shape=kernel_shape)
state = net.initial_state_for_inputs(gradients)
net(gradients, state)
# Weights and biases for two layers.
variables = nn.get_variables_in_module(net)
self.assertEqual(len(variables), 4)
def testResults(self, initializer):
"""Tests zero updates when last layer is initialized to zero."""
kernel_shape = [5, 5]
shape = kernel_shape + [2, 2] # The input has to be 4-dimensional.
gradients = tf.random_normal(shape)
net = networks.KernelDeepLSTM(layers=(1, 1),
kernel_shape=kernel_shape,
initializer=initializer)
state = net.initial_state_for_inputs(gradients)
update, _ = net(gradients, state)
with self.test_session() as sess:
sess.run(tf.global_variables_initializer())
update_np = sess.run(update)
self.assertAllEqual(update_np, np.zeros(shape))
