def testGeneratorInitializersOrthogonal(self):
gin.bind_parameter("weights.initializer", consts.ORTHOGONAL_INIT)
valid_initalizer = [
"kernel/Initializer/mul_1",
"bias/Initializer/Const",
"kernel/Initializer/mul_1",
"bias/Initializer/Const",
"beta/Initializer/zeros",
"gamma/Initializer/ones",
]
valid_op_names = "/({}):0$".format("|".join(valid_initalizer))
with tf.Graph().as_default():
z = tf.zeros((2, 128))
fake_image = resnet5.Generator(image_shape=(128, 128, 3))(
z, y=None, is_training=True)
resnet5.Discriminator()(fake_image, y=None, is_training=True)
for var in tf.trainable_variables():
op_name = var.initializer.inputs[1].name
self.assertRegex(op_name, valid_op_names)
def testInitializersOldDefault(self):
valid_initalizer = [
"kernel/Initializer/random_normal",
"bias/Initializer/Const",
# truncated_normal is the old default for conv2d.
"kernel/Initializer/truncated_normal",
"bias/Initializer/Const",
"beta/Initializer/zeros",
"gamma/Initializer/ones",
]
valid_op_names = "/({}):0$".format("|".join(valid_initalizer))
with tf.Graph().as_default():
z = tf.zeros((2, 128))
fake_image = resnet5.Generator(image_shape=(128, 128, 3))(
z, y=None, is_training=True)
resnet5.Discriminator()(fake_image, y=None, is_training=True)
for var in tf.trainable_variables():
op_name = var.initializer.inputs[1].name
self.assertRegex(op_name, valid_op_names)
def testResNet5(self, image_shape):
self.assertArchitectureBuilds(
gen=resnet5.Generator(image_shape=image_shape),
disc=resnet5.Discriminator(),
image_shape=image_shape)