def test_process_whole_image(self, data_format, shape):
keras.backend.set_image_data_format(data_format)
num_crops = 2
receptive_field = 3
features = 3
images = np.ones(shape)
input_shape = running.get_cropped_input_shape(
images, num_crops,
receptive_field=receptive_field,
data_format=data_format)
for padding in ['reflect', 'zero']:
with self.test_session():
inputs = keras.layers.Input(shape=input_shape)
outputs = layers.TensorProduct(features)(inputs)
model = keras.models.Model(inputs=inputs,
outputs=[outputs, outputs])
output = running.process_whole_image(
model, images,
num_crops=num_crops,
receptive_field=receptive_field,
padding=padding)
if data_format == 'channels_first':
expected_shape = tuple([images.shape[0], features] +
list(images.shape[2:]))
else:
expected_shape = tuple([images.shape[0]] +
list(images.shape[1:-1]) +
[features])
self.assertEqual(output.shape, expected_shape)
with self.assertRaises(ValueError):
inputs = keras.layers.Input(shape=(3, 4, 5))
outputs = layers.TensorProduct(features)(inputs)
model = keras.models.Model(inputs=inputs, outputs=outputs)
output = running.process_whole_image(
model, images,
num_crops=num_crops,
receptive_field=receptive_field,
padding='reflect')
with self.assertRaises(ValueError):
inputs = keras.layers.Input(shape=input_shape)
outputs = layers.TensorProduct(features)(inputs)
model = keras.models.Model(inputs=inputs, outputs=outputs)
output = running.process_whole_image(
model, images,
num_crops=num_crops,
receptive_field=receptive_field,
padding=None)
def test_tensorproduct_constraints(self):
k_constraint = tf.keras.constraints.max_norm(0.01)
b_constraint = tf.keras.constraints.max_norm(0.01)
layer = layers.TensorProduct(3,
kernel_constraint=k_constraint,
bias_constraint=b_constraint)
layer(tf.keras.backend.variable(np.ones((2, 4))))
self.assertEqual(layer.kernel.constraint, k_constraint)
self.assertEqual(layer.bias.constraint, b_constraint)
def test_tensorproduct_regularization(self):
layer = layers.TensorProduct(
3,
kernel_regularizer=tf.keras.regularizers.l1(0.01),
bias_regularizer='l1',
activity_regularizer='l2',
name='tensorproduct_reg')
layer(tf.keras.backend.variable(np.ones((2, 4))))
self.assertEqual(3, len(layer.losses))