def testActivation(self):
fc = core_layers.FullyConnected(2, activation=tf.nn.relu, name='fc1')
inputs = tf.random_uniform((5, 3), seed=1)
outputs = fc(inputs)
self.assertEqual(outputs.op.name, 'fc1/Relu')
fc = core_layers.FullyConnected(2, name='fc2')
inputs = tf.random_uniform((5, 3), seed=1)
outputs = fc(inputs)
self.assertEqual(outputs.op.name, 'fc2/BiasAdd')
def testFCProperties(self):
fc = core_layers.FullyConnected(2, activation=tf.nn.relu, name='fc')
self.assertEqual(fc.output_dim, 2)
self.assertEqual(fc.activation, tf.nn.relu)
self.assertEqual(fc.w_regularizer, None)
self.assertEqual(fc.bias_regularizer, None)
self.assertEqual(fc.activity_regularizer, None)
self.assertEqual(fc.use_bias, True)
self.assertEqual(fc.name, 'fc')
# Test auto-naming
fc = core_layers.FullyConnected(2, activation=tf.nn.relu)
self.assertEqual(fc.name, 'fully_connected')
fc = core_layers.FullyConnected(2, activation=tf.nn.relu)
self.assertEqual(fc.name, 'fully_connected_1')
def testBiasRegularizer(self):
regularizer = lambda x: tf.reduce_sum(x) * 1e-3
fc = core_layers.FullyConnected(2,
name='fc',
bias_regularizer=regularizer)
inputs = tf.random_uniform((5, 3), seed=1)
_ = fc(inputs)
loss_keys = tf.get_collection(tf.GraphKeys.REGULARIZATION_LOSSES)
self.assertEqual(len(loss_keys), 1)
self.assertListEqual(fc.losses, loss_keys)
def testNonTrainable(self):
fc = core_layers.FullyConnected(2, trainable=False, name='fc')
inputs = tf.random_uniform((5, 2), seed=1)
_ = fc(inputs)
self.assertListEqual(fc.weights, [fc.w, fc.bias])
self.assertListEqual(fc.non_trainable_weights, [fc.w, fc.bias])
self.assertListEqual(fc.trainable_weights, [])
self.assertListEqual(fc._trainable_weights, [fc.w, fc.bias])
self.assertListEqual(fc._non_trainable_weights, [])
self.assertEqual(
len(tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES)), 0)
def testNoBias(self):
fc = core_layers.FullyConnected(2, use_bias=False, name='fc')
inputs = tf.random_uniform((5, 2), seed=1)
_ = fc(inputs)
self.assertListEqual(fc.weights, [fc.w])
self.assertListEqual(fc.trainable_weights, [fc.w])
self.assertListEqual(fc.non_trainable_weights, [])
self.assertEqual(
len(tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES)), 1)
self.assertEqual(fc.w.name, 'fc/weights:0')
self.assertEqual(fc.bias, None)
def testOutputShape(self):
fc = core_layers.FullyConnected(7, activation=tf.nn.relu, name='fc')
inputs = tf.random_uniform((5, 3), seed=1)
outputs = fc.apply(inputs)
self.assertEqual(outputs.get_shape().as_list(), [5, 7])
inputs = tf.random_uniform((5, 2, 3), seed=1)
outputs = fc(inputs)
self.assertEqual(outputs.get_shape().as_list(), [5, 2, 7])
inputs = tf.random_uniform((1, 2, 4, 3), seed=1)
outputs = fc.apply(inputs)
self.assertEqual(outputs.get_shape().as_list(), [1, 2, 4, 7])
def testCall(self):
fc = core_layers.FullyConnected(2, activation=tf.nn.relu, name='fc')
inputs = tf.random_uniform((5, 2), seed=1)
_ = fc(inputs)
self.assertListEqual(fc.weights, [fc.w, fc.bias])
self.assertListEqual(fc.trainable_weights, [fc.w, fc.bias])
self.assertListEqual(fc.non_trainable_weights, [])
self.assertListEqual(fc._trainable_weights, [fc.w, fc.bias])
self.assertListEqual(fc._non_trainable_weights, [])
self.assertEqual(
len(tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES)), 2)
self.assertEqual(fc.w.name, 'fc/weights:0')
self.assertEqual(fc.bias.name, 'fc/biases:0')
def testCallOnPlaceHolder(self):
inputs = tf.placeholder(dtype=tf.float32)
fc = core_layers.FullyConnected(4, name='fc')
with self.assertRaises(ValueError):
fc(inputs)
inputs = tf.placeholder(dtype=tf.float32, shape=[None, None])
fc = core_layers.FullyConnected(4, name='fc')
with self.assertRaises(ValueError):
fc(inputs)
inputs = tf.placeholder(dtype=tf.float32, shape=[None, None, None])
fc = core_layers.FullyConnected(4, name='fc')
with self.assertRaises(ValueError):
fc(inputs)
inputs = tf.placeholder(dtype=tf.float32, shape=[None, 3])
fc = core_layers.FullyConnected(4, name='fc')
fc(inputs)
inputs = tf.placeholder(dtype=tf.float32, shape=[None, None, 3])
fc = core_layers.FullyConnected(4, name='fc')
fc(inputs)