def get_config(self):
serialized_optimizer = optimizers.serialize(self._optimizer)
serialized_loss_scale = keras_loss_scale_module.serialize(self._loss_scale)
return {
'optimizer': serialized_optimizer,
'loss_scale': serialized_loss_scale,
}
def testSerializationWithBuiltInOptimizer(self, use_v1):
opt = gradient_descent.SGD(2., momentum=0.5)
if use_v1:
loss_scale = tf_loss_scale_module.DynamicLossScale(
initial_loss_scale=2., increment_period=3.)
opt = loss_scale_optimizer.LossScaleOptimizerV1(opt, loss_scale)
else:
opt = loss_scale_optimizer.LossScaleOptimizer(
opt, initial_scale=2., dynamic_growth_steps=3.)
config = optimizers.serialize(opt)
opt = optimizers.deserialize(config)
# Force hyperparameters to be created
opt.lr # pylint: disable=pointless-statement
self.evaluate(variables.global_variables_initializer())
self.assertEqual(self.evaluate(opt.lr), 2.)
self.assertEqual(self.evaluate(opt._optimizer.momentum), 0.5)
self.assertEqual(self.evaluate(opt.loss_scale), 2.)
self.assertEqual(opt.dynamic_growth_steps, 3.)
self.assertTrue(opt.dynamic, 4.)
# Deserializing a LossScaleOptimizer always always results in a V2
# LossScaleOptimizer, even if serialized with a LossScaleOptimizerV1.
self.assertAllEqual(type(opt), loss_scale_optimizer.LossScaleOptimizer)
# Ensure the optimizer can be used
var = variables.Variable([5.0])
run_op = self._run_fn_with_grad_check(
distribution_strategy_context.get_strategy(), var, opt, 2)()
self.evaluate(variables.global_variables_initializer())
self._run_if_in_graph_mode(run_op)
self.assertEqual(self.evaluate(var), [3.])
self.assertEqual(self.evaluate(opt.dynamic_counter), 1)
def get_config(self):
config = {
'optimizer': optimizers.serialize(self._optimizer),
'lr_multipliers': self._lr_multipliers
}
base_config = super(LearningRateMultiplier, self).get_config()
return dict(list(base_config.items()) + list(config.items()))
def testSerializationWithCustomOptimizer(self):
class MySGD(gradient_descent.SGD):
def __init__(self, *args, **kwargs):
super(MySGD, self).__init__(*args, **kwargs)
self.my_attribute = 123
opt = MySGD(2., momentum=0.5)
loss_scale = loss_scale_module.DynamicLossScale(
initial_loss_scale=2., increment_period=3.,
multiplier=4.)
opt = loss_scale_optimizer.LossScaleOptimizer(opt, loss_scale)
config = optimizers.serialize(opt)
custom_objects = {'MySGD': MySGD}
opt = optimizers.deserialize(config, custom_objects=custom_objects)
# Force hyperparameters to be created
opt.lr # pylint: disable=pointless-statement
self.evaluate(variables.global_variables_initializer())
self.assertEqual(self.evaluate(opt.lr), 2.)
self.assertEqual(self.evaluate(opt._optimizer.momentum), 0.5)
self.assertEqual(self.evaluate(opt.loss_scale()), 2.)
self.assertEqual(opt.loss_scale.increment_period, 3.)
self.assertEqual(opt.loss_scale.multiplier, 4.)
self.assertEqual(opt._optimizer.my_attribute, 123)
def valid_optimizer(optimizer):
if optimizer and isinstance(optimizer, dict):
class_name = optimizer.get('class_name')
optimizer = get(class_name).from_config(optimizer.get('config', {}))
optimizer = serialize(optimizer)
elif isinstance(optimizer, DataFrame):
optimizer = json.loads(optimizer.first().optimizer)
return optimizer
def get_config(self):
serialized_optimizer = optimizers.serialize(self._optimizer)
return {
'optimizer': serialized_optimizer,
'dynamic': self.dynamic,
'initial_scale': self.initial_scale,
'dynamic_growth_steps': self.dynamic_growth_steps,
}
def testSerializationWithBuiltInOptimizer(self):
opt = gradient_descent.SGD(2., momentum=0.5)
loss_scale = loss_scale_module.DynamicLossScale(initial_loss_scale=2.,
increment_period=3.,
multiplier=4.)
opt = loss_scale_optimizer.LossScaleOptimizer(opt, loss_scale)
config = optimizers.serialize(opt)
opt = optimizers.deserialize(config)
# Force hyperparameters to be created
opt.lr # pylint: disable=pointless-statement
self.evaluate(variables.global_variables_initializer())
self.assertEqual(self.evaluate(opt.lr), 2.)
self.assertEqual(self.evaluate(opt._optimizer.momentum), 0.5)
self.assertEqual(self.evaluate(opt.loss_scale()), 2.)
self.assertEqual(opt.loss_scale.increment_period, 3.)
self.assertEqual(opt.loss_scale.multiplier, 4.)
def __add_compile_layer(self, layer):
return self._add_or_create_column('optimizer',
json.dumps(serialize(layer)))