def testFunctions(self):
# Create a sequential network with one function.
net = network.Sequential([nn_ops.relu])
two = constant_op.constant(2.0)
self.assertEqual(2.0, net(two).numpy())
self.assertEqual(0.0, net(-two).numpy())
# Add a second function.
net.add(math_ops.negative)
self.assertEqual(-2.0, net(two).numpy())
def testTrainingLayer(self):
net = network.Sequential([core.Dropout(0.99999)])
two = constant_op.constant(2.0)
self.assertEqual(2.0, net(two).numpy())
self.assertEqual(2.0, net(two, training=False).numpy())
for _ in range(20):
with_dropout = net(two, training=True).numpy()
self.assertIn(with_dropout, [0.0, 2.0])
if with_dropout == 0.0:
return
# Should only fail spuriously 1 in 10^100 runs.
self.fail("Didn't see dropout happen after 20 tries.")
def testTrainingFunction(self):
# Output depends on value of "training".
def add_training(input_value, training=None):
if training is None:
return input_value
elif training:
return input_value + 1
return input_value - 1
# Passing a "training" argument to double would cause an error.
def double(input_value):
return 2 * input_value
net = network.Sequential([add_training, double])
two = constant_op.constant(2)
self.assertEqual(4, net(two).numpy())
self.assertEqual(2, net(two, training=False).numpy())
self.assertEqual(6, net(two, training=True).numpy())
def testTwoLayers(self):
# Create a sequential network with one layer.
net = network.Sequential([core.Dense(1, use_bias=False)])
# Set that layer's weights so it multiplies by 3
l1 = net.get_layer(index=0)
net(constant_op.constant([[2.0]])) # Create l1's variables
self.assertEqual(1, len(l1.trainable_variables))
l1.trainable_variables[0].assign([[3.0]])
self.assertEqual(21.0, net(constant_op.constant([[7.0]])).numpy())
# Add a second layer to the network.
l2 = core.Dense(1, use_bias=False)
net.add(l2)
# Set the second layer's weights so it multiplies by 11
net(constant_op.constant([[2.0]])) # Create l2's variables
self.assertEqual(1, len(l2.trainable_variables))
l2.trainable_variables[0].assign([[11.0]])
self.assertEqual(231.0, net(constant_op.constant([[7.0]])).numpy())
