def RAdam(parameters, lr, momentum=0.9, unit_gain=C.default_unit_gain_value(),
beta2=0.999, l1_regularization_weight=0.0, l2_regularization_weight=0.0,
gaussian_noise_injection_std_dev=0.0, gradient_clipping_threshold_per_sample=np.inf,
gradient_clipping_with_truncation=True, use_mean_gradient=None, epsilon=1e-8, adamax=False,
minibatch_size=None, epoch_size=None):
""" RAdam like implementation using Adam with exponential warmup schedule. No tuning of
warmup schedule required, unlike Adam.
This is a simple untuned warmup of Adam with 'rule-of-thumb' warmup schedule that performs
more-or-less identically to RAdam in typical practical settings based on
'On the adequacy of untuned warmup for adaptive optimization' by Jerry Ma and Denis Yarats.
For more details, paper can be found here 'https://arxiv.org/abs/1910.04209'
Args:
... please look at original documentation in cntk.learner.adam
epoch_size (optional, int): number of samples as a scheduling unit for learning rate, momentum and variance_momentum. See also: :func:`learning_parameter_schedule`
Returns:
:class:`~cntk.learners.Learner`: learner instance that can be passed to
the :class:`~cntk.train.trainer.Trainer`
See also:
[1] D. Kingma, J. Ba. `Adam: A Method for Stochastic Optimization
<https://arxiv.org/abs/1412.6980>`_. International Conference for
Learning Representations, 2015.
"""
if epoch_size is None:
raise ValueError("epoch size should be set to the number of samples per minibatch "
"(i.e. number of samples trained in every training update) so that "
"learning rate factor can be updated after every training update")
lr = adam_exponential_warmup_schedule(lr, beta2) # rule-of-thumb exponential warmup schedule
lr, minibatch_size = _infer_learning_rate_schedule_and_ref_minibatch_size(use_mean_gradient, minibatch_size, lr, epoch_size)
momentum = _infer_learning_parameter_schedule(momentum, minibatch_size, epoch_size)
_verify_momentum_type(momentum)
variance_momentum = _infer_learning_parameter_schedule(beta2, minibatch_size, epoch_size)
_verify_momentum_type(variance_momentum)
gaussian_noise_injection_std_dev = C.training_parameter_schedule(gaussian_noise_injection_std_dev)
additional_options = cntk_py.AdditionalLearningOptions()
additional_options.l1_regularization_weight = l1_regularization_weight
additional_options.l2_regularization_weight = l2_regularization_weight
additional_options.gaussian_noise_injection_std_dev = gaussian_noise_injection_std_dev
additional_options.gradient_clipping_threshold_per_sample = gradient_clipping_threshold_per_sample
additional_options.gradient_clipping_with_truncation = gradient_clipping_with_truncation
if minibatch_size is not None:
additional_options.dict_options[cntk_py.Learner._MINIBATCH_SIZE] = cntk_py.SizeTWrapper(minibatch_size) # need this to make proper typed DictionaryValue
opt = cntk_py.adam_learner(parameters, lr, momentum, unit_gain, variance_momentum, epsilon, adamax, additional_options)
opt.is_minibatch_size_explicitly_specified = minibatch_size is not None
return opt
def test_learner_init():
i = C.input_variable(shape=(1,), needs_gradient=True, name='a')
w = parameter(shape=(1,))
res = i * w
learner = sgd(res.parameters, lr=learning_rate_schedule(0.1, UnitType.sample))
assert learner.learning_rate() == 0.1
learner.reset_learning_rate(learning_rate_schedule([1,2,3], UnitType.minibatch));
assert learner.learning_rate() == 1.0
learner_parameter = learner.parameters
from cntk.variables import Parameter
param = learner_parameter[0]
assert isinstance(param, Parameter)
unit_gain_value = C.default_unit_gain_value()
assert unit_gain_value
momentum_time_constant = C.momentum_as_time_constant_schedule(1100)
lr_per_sample = learning_rate_schedule(0.1, UnitType.sample)
C.momentum_sgd(res.parameters, lr_per_sample, momentum_time_constant)
C.momentum_sgd(res.parameters, lr_per_sample, momentum_time_constant, unit_gain_value)
C.momentum_sgd(res.parameters, lr_per_sample, momentum_time_constant, unit_gain=unit_gain_value)
C.set_default_unit_gain_value(False)
unit_gain_value = C.default_unit_gain_value()
assert not unit_gain_value
lr_per_sample = learning_rate_schedule([0.1, 0.2], UnitType.sample)
C.nesterov(res.parameters, lr=lr_per_sample, momentum=momentum_time_constant)
C.nesterov(res.parameters, lr_per_sample, momentum_time_constant, unit_gain_value)
C.nesterov(res.parameters, lr=lr_per_sample, momentum=momentum_time_constant, unit_gain=unit_gain_value)
lr_per_sample = learning_rate_schedule([0.1]*3 +[0.2]*2 +[0.3], UnitType.sample)
C.adagrad(res.parameters, lr=lr_per_sample, need_ave_multiplier=True)
C.set_default_unit_gain_value(True)
unit_gain_value = C.default_unit_gain_value()
assert unit_gain_value
lr_per_sample = learning_rate_schedule([(3,0.1), (2, 0.2), (1, 0.3)], UnitType.sample)
C.fsadagrad(res.parameters, lr=lr_per_sample, momentum=momentum_time_constant)
C.fsadagrad(res.parameters, lr_per_sample, momentum_time_constant, unit_gain_value)
C.fsadagrad(res.parameters, lr=lr_per_sample, momentum=momentum_time_constant, unit_gain=unit_gain_value)
gamma, inc, dec, max, min = [0.1]*5
lr_per_sample = learning_rate_schedule([0.1, 0.2], UnitType.sample, 100)
C.rmsprop(res.parameters, lr_per_sample, gamma, inc, dec, max, min, True)
C.set_default_use_mean_gradient_value(False)
use_mean_gradient_value = C.default_use_mean_gradient_value()
assert not use_mean_gradient_value
C.adadelta(res.parameters, lr_per_sample)
C.set_default_use_mean_gradient_value(True)
use_mean_gradient_value = C.default_use_mean_gradient_value()
assert use_mean_gradient_value
C.adadelta(res.parameters, lr_per_sample)
def test_learner_init():
i = C.input_variable(shape=(1, ), needs_gradient=True, name='a')
w = parameter(shape=(1, ))
res = i * w
#test new API: learning_parameter_schedule
#explictly specify reference minibatch size and learning rate is in number:
learner = sgd(res.parameters, lr=0.1, minibatch_size=25)
assert learner.is_compatible_mode() == False
assert learner.minibatch_size == 25 #the learner's reference minibatch
#with direct learner learning rate number specification, the learning rate schedule get the reference minibatch size from the learner parameters:
assert learner._learning_rate_schedule.minibatch_size == 25
assert learner.learning_rate() == 0.1
#no explictly specification of reference minibatch size and learning rate is in number:
learner = sgd(res.parameters, lr=learning_parameter_schedule(0.1))
assert learner.is_compatible_mode() == False
assert learner.minibatch_size == C.learners.IGNORE #the learner's reference minibatch
#with direct learner learning rate number specification, the learning rate schedule get the reference minibatch size from the learner parameters:
assert learner._learning_rate_schedule.minibatch_size == C.learners.IGNORE
assert learner.learning_rate() == 0.1
learner = sgd(res.parameters,
lr=learning_parameter_schedule(0.1, 20),
minibatch_size=25)
assert learner.is_compatible_mode() == False
assert learner.minibatch_size == 25 #the learner's reference minibatch
#with direct learner learning rate number specification, the learning rate schedule get the reference minibatch size from the learner parameters:
assert learner._learning_rate_schedule.minibatch_size == 20
assert learner.learning_rate() == 0.1
learner = sgd(res.parameters, lr=learning_parameter_schedule(0.1, 20))
assert learner.is_compatible_mode() == False
#with direct learner learning rate number specification, the learning rate schedule get the reference minibatch size from the learner parameters:
assert learner._learning_rate_schedule.minibatch_size == 20
assert learner.learning_rate() == 0.1
#no explictly specification of reference minibatch size and learning rate is in number:
learner = sgd(res.parameters, lr=learning_parameter_schedule(0.1))
assert learner.is_compatible_mode() == False
assert learner.minibatch_size == C.learners.IGNORE #the learner's reference minibatch
#with direct learner learning rate number specification, the learning rate schedule get the reference minibatch size from the learner parameters:
assert learner._learning_rate_schedule.minibatch_size == C.learners.IGNORE
assert learner.learning_rate() == 0.1
#no explictly specification of reference minibatch size and learning rate is in number:
learner = sgd(res.parameters,
lr=learning_parameter_schedule(0.1),
minibatch_size=C.learners.IGNORE)
assert learner.is_compatible_mode() == True
assert learner.minibatch_size == C.learners.IGNORE #the learner's reference minibatch
#with direct learner learning rate number specification, the learning rate schedule get the reference minibatch size from the learner parameters:
assert learner._learning_rate_schedule.minibatch_size == C.learners.IGNORE
assert learner.learning_rate() == 0.1
learner = sgd(res.parameters,
lr=learning_parameter_schedule(0.1, 20),
minibatch_size=C.learners.IGNORE)
assert learner.is_compatible_mode() == True
assert learner.minibatch_size == C.learners.IGNORE #the learner's reference minibatch
#with direct learner learning rate number specification, the learning rate schedule get the reference minibatch size from the learner parameters:
assert learner._learning_rate_schedule.minibatch_size == 20
assert learner.learning_rate() == 0.1
#no explictly specification of reference minibatch size and learning rate is in number:
learner = sgd(res.parameters,
lr=learning_parameter_schedule(0.1),
minibatch_size=C.learners.IGNORE)
assert learner.is_compatible_mode() == True
assert learner.minibatch_size == C.learners.IGNORE #the learner's reference minibatch
#with direct learner learning rate number specification, the learning rate schedule get the reference minibatch size from the learner parameters:
assert learner._learning_rate_schedule.minibatch_size == C.learners.IGNORE
assert learner.learning_rate() == 0.1
mysgd = C.sgd(parameters=res.parameters, lr=0.4, minibatch_size=32)
assert mysgd.minibatch_size == 32
assert mysgd._learning_rate_schedule.minibatch_size == 32
assert mysgd.learning_rate() == 0.4
mymomentum = C.momentum_sgd(parameters=res.parameters,
lr=0.4,
momentum=0.9,
minibatch_size=32)
assert mymomentum.minibatch_size == 32
assert mymomentum._learning_rate_schedule.minibatch_size == 32
assert mymomentum.learning_rate() == 0.4
myadadelta = C.adadelta(parameters=res.parameters,
lr=0.4,
minibatch_size=32)
assert myadadelta.minibatch_size == 32
assert myadadelta._learning_rate_schedule.minibatch_size == 32
assert myadadelta.learning_rate() == 0.4
myadam = C.adam(parameters=res.parameters,
lr=0.4,
momentum=0.9,
variance_momentum=0.9,
minibatch_size=32)
assert myadam.minibatch_size == 32
assert myadam._learning_rate_schedule.minibatch_size == 32
assert myadam.learning_rate() == 0.4
myadagrad = C.adagrad(parameters=res.parameters, lr=0.4, minibatch_size=32)
assert myadagrad.minibatch_size == 32
assert myadagrad._learning_rate_schedule.minibatch_size == 32
assert myadagrad.learning_rate() == 0.4
myfsadagrad = C.fsadagrad(parameters=res.parameters,
lr=0.4,
momentum=0.9,
variance_momentum=0.9,
minibatch_size=32)
assert myfsadagrad.minibatch_size == 32
assert myfsadagrad._learning_rate_schedule.minibatch_size == 32
assert myfsadagrad.learning_rate() == 0.4
mynesterov = C.nesterov(parameters=res.parameters,
lr=0.4,
momentum=0.9,
minibatch_size=32)
assert mynesterov.minibatch_size == 32
assert mynesterov._learning_rate_schedule.minibatch_size == 32
assert mynesterov.learning_rate() == 0.4
myrmsrop = C.rmsprop(parameters=res.parameters,
lr=0.4,
gamma=0.5,
inc=1.2,
dec=0.7,
max=10,
min=1e-8,
minibatch_size=32)
assert myrmsrop.minibatch_size == 32
assert myrmsrop._learning_rate_schedule.minibatch_size == 32
assert myrmsrop.learning_rate() == 0.4
mysgd = C.sgd(parameters=res.parameters,
lr=[0.4, 0.1, 0.001],
minibatch_size=32,
epoch_size=512)
assert mysgd.minibatch_size == 32
assert mysgd._learning_rate_schedule.minibatch_size == 32
assert mysgd._learning_rate_schedule[0] == 0.4
assert mysgd._learning_rate_schedule[512] == 0.1
assert mysgd._learning_rate_schedule[512 * 2] == 0.001
mymomentum = C.momentum_sgd(parameters=res.parameters,
lr=[0.4, 0.1, 0.001],
momentum=[0.9],
minibatch_size=32,
epoch_size=512)
assert mymomentum.minibatch_size == 32
assert mymomentum._learning_rate_schedule.minibatch_size == 32
assert mymomentum._learning_rate_schedule[0] == 0.4
assert mymomentum._learning_rate_schedule[512] == 0.1
assert mymomentum._learning_rate_schedule[512 * 2] == 0.001
myadadelta = C.adadelta(parameters=res.parameters,
lr=[0.4, 0.1, 0.001],
minibatch_size=32,
epoch_size=512)
assert myadadelta.minibatch_size == 32
assert myadadelta._learning_rate_schedule.minibatch_size == 32
assert myadadelta._learning_rate_schedule[0] == 0.4
assert myadadelta._learning_rate_schedule[512] == 0.1
assert myadadelta._learning_rate_schedule[512 * 2] == 0.001
myadam = C.adam(parameters=res.parameters,
lr=[0.4, 0.1, 0.001],
momentum=[0.9, 0.1, 0.001],
variance_momentum=[0.9],
minibatch_size=32,
epoch_size=512)
assert myadam.minibatch_size == 32
assert myadam._learning_rate_schedule.minibatch_size == 32
assert myadam._learning_rate_schedule[0] == 0.4
assert myadam._learning_rate_schedule[512] == 0.1
assert myadam._learning_rate_schedule[512 * 2] == 0.001
myadagrad = C.adagrad(parameters=res.parameters,
lr=[0.4, 0.1, 0.001],
minibatch_size=32,
epoch_size=512)
assert myadagrad.minibatch_size == 32
assert myadagrad._learning_rate_schedule.minibatch_size == 32
assert myadagrad._learning_rate_schedule[0] == 0.4
assert myadagrad._learning_rate_schedule[512] == 0.1
assert myadagrad._learning_rate_schedule[512 * 2] == 0.001
myfsadagrad = C.fsadagrad(parameters=res.parameters,
lr=[0.4, 0.1, 0.001],
momentum=[0.9],
variance_momentum=[0.9],
minibatch_size=32,
epoch_size=512)
assert myadagrad.minibatch_size == 32
assert myadagrad._learning_rate_schedule.minibatch_size == 32
assert myadagrad._learning_rate_schedule[0] == 0.4
assert myadagrad._learning_rate_schedule[512] == 0.1
assert myadagrad._learning_rate_schedule[512 * 2] == 0.001
mynesterov = C.nesterov(parameters=res.parameters,
lr=[0.4, 0.1, 0.001],
momentum=[0.9],
minibatch_size=32,
epoch_size=512)
assert mynesterov.minibatch_size == 32
assert mynesterov._learning_rate_schedule.minibatch_size == 32
assert mynesterov._learning_rate_schedule[0] == 0.4
assert mynesterov._learning_rate_schedule[512] == 0.1
assert mynesterov._learning_rate_schedule[512 * 2] == 0.001
myrmsrop = C.rmsprop(parameters=res.parameters,
lr=[0.4, 0.1, 0.001],
gamma=0.5,
inc=1.2,
dec=0.7,
max=10,
min=1e-8,
minibatch_size=32,
epoch_size=512)
assert myrmsrop.minibatch_size == 32
assert myrmsrop._learning_rate_schedule.minibatch_size == 32
assert myrmsrop._learning_rate_schedule[0] == 0.4
assert myrmsrop._learning_rate_schedule[512] == 0.1
assert myrmsrop._learning_rate_schedule[512 * 2] == 0.001
learner_parameter = learner.parameters
from cntk.variables import Parameter
param = learner_parameter[0]
assert isinstance(param, Parameter)
unit_gain_value = C.default_unit_gain_value()
assert unit_gain_value
momentum = C.momentum_schedule(0.999, minibatch_size=1)
lr_per_sample = learning_parameter_schedule(0.1, minibatch_size=1)
C.momentum_sgd(res.parameters, lr_per_sample, momentum)
C.momentum_sgd(res.parameters, lr_per_sample, momentum, unit_gain_value)
C.momentum_sgd(res.parameters,
lr_per_sample,
momentum,
unit_gain=unit_gain_value)
C.set_default_unit_gain_value(False)
unit_gain_value = C.default_unit_gain_value()
assert not unit_gain_value
lr_per_sample = learning_parameter_schedule([0.1, 0.2], minibatch_size=1)
C.nesterov(res.parameters, lr=lr_per_sample, momentum=momentum)
C.nesterov(res.parameters, lr_per_sample, momentum, unit_gain_value)
C.nesterov(res.parameters,
lr=lr_per_sample,
momentum=momentum,
unit_gain=unit_gain_value)
lr_per_sample = learning_parameter_schedule([0.1] * 3 + [0.2] * 2 + [0.3],
minibatch_size=1)
C.adagrad(res.parameters, lr=lr_per_sample, need_ave_multiplier=True)
C.set_default_unit_gain_value(True)
unit_gain_value = C.default_unit_gain_value()
assert unit_gain_value
lr_per_sample = learning_parameter_schedule([(3, 0.1), (2, 0.2), (1, 0.3)],
minibatch_size=1)
C.fsadagrad(res.parameters, lr=lr_per_sample, momentum=momentum)
C.fsadagrad(res.parameters, lr_per_sample, momentum, unit_gain_value)
C.fsadagrad(res.parameters,
lr=lr_per_sample,
momentum=momentum,
unit_gain=unit_gain_value)
gamma, inc, dec, max, min = [0.5, 1.2, 0.7, 10, 1e-8]
lr_per_sample = learning_parameter_schedule([0.1, 0.2],
minibatch_size=1,
epoch_size=100)
C.rmsprop(res.parameters, lr_per_sample, gamma, inc, dec, max, min, True)
C.adadelta(res.parameters, lr_per_sample)
def test_learner_init_legacy():
i = C.input_variable(shape=(1, ), needs_gradient=True, name='a')
w = parameter(shape=(1, ))
res = i * w
# for backcompatibility test
# this will be deprecated in future version
learner = sgd(res.parameters,
lr=learning_rate_schedule(0.1, UnitType.sample))
assert learner._learning_rate_schedule.minibatch_size == 1 # the deprecated per sample schedule should not use compatible mode
assert learner.learning_rate() == 0.1
# for backcompatibility test
# this will be deprecated in future version
# The UnitType will provide per minibatch instruction for the learner
# this will be deprecated in future version
learner = sgd(res.parameters,
lr=learning_rate_schedule(0.1, UnitType.minibatch))
assert learner.is_compatible_mode() == False
assert learner.learning_rate() == 0.1
assert learner.minibatch_size == C.learners.IGNORE
assert learner._learning_rate_schedule.minibatch_size == 0
# for backcompatibility test, in reset learning rate, the learner won't receive the reference minibatch size from the schedule
# user will need to specify the reference minibatch size explicitly
# this will be deprecated in future version
learner = sgd(res.parameters, lr=0.1)
learner.reset_learning_rate(
learning_rate_schedule([1, 2, 3], UnitType.minibatch))
assert learner.learning_rate() == 1.0
learner.minibatch_size = C.learners.IGNORE # reset to be per minibatch
assert learner.minibatch_size == C.learners.IGNORE
assert learner._learning_rate_schedule.minibatch_size == C.learners.IGNORE
assert learner.is_compatible_mode() == True
# for backcompatibility test
# this will be deprecated in future version
learner = sgd(res.parameters,
lr=learning_rate_schedule(0.1, UnitType.sample),
minibatch_size=C.learners.IGNORE)
assert learner.is_compatible_mode() == True
assert learner.learning_rate() == 0.1
assert learner.minibatch_size == C.learners.IGNORE # the learner's reference minibatch size is still 0
# this will be deprecated in future version: This is logical invalid combination but it was the only way to use mean gradient and set learning rate in the past.
learner = sgd(res.parameters,
lr=learning_rate_schedule(0.1, UnitType.sample),
use_mean_gradient=True)
assert learner.is_compatible_mode() == True
assert learner.learning_rate() == 0.1
#test the override in the new version
assert learner._learning_rate_schedule.minibatch_size == C.learners.IGNORE
assert learner.minibatch_size == C.learners.IGNORE # the learner's reference minibatch size is still 0
# for backcompatibility test
# this will be deprecated in future version
# The UnitType will provide per minibatch instruction for the learner
# this will be deprecated in future version
learner = sgd(res.parameters,
lr=learning_rate_schedule(0.1, UnitType.minibatch),
minibatch_size=C.learners.IGNORE)
assert learner.is_compatible_mode() == True
assert learner.learning_rate() == 0.1
assert learner.minibatch_size == C.learners.IGNORE
assert learner._learning_rate_schedule.minibatch_size == C.learners.IGNORE
# for backcompatibility test, in reset learning rate, the learner won't receive the reference minibatch size from the schedule
# user will need to specify the reference minibatch size explicitly
# this will be deprecated in future version
learner = sgd(res.parameters, lr=0.1)
learner.reset_learning_rate(
learning_rate_schedule([1, 2, 3], UnitType.minibatch))
assert learner.learning_rate() == 1.0
learner.minibatch_size = C.learners.IGNORE # reset to be per minibatch
assert learner.minibatch_size == C.learners.IGNORE
assert learner._learning_rate_schedule.minibatch_size == C.learners.IGNORE
assert learner.is_compatible_mode() == True
learner_parameter = learner.parameters
from cntk.variables import Parameter
param = learner_parameter[0]
assert isinstance(param, Parameter)
unit_gain_value = C.default_unit_gain_value()
assert unit_gain_value
# back compatible API test
momentum_time_constant = C.momentum_as_time_constant_schedule(1100)
lr_per_sample = learning_parameter_schedule(0.1, minibatch_size=1)
C.momentum_sgd(res.parameters, lr_per_sample, momentum_time_constant)
C.momentum_sgd(res.parameters, lr_per_sample, momentum_time_constant,
unit_gain_value)
C.momentum_sgd(res.parameters,
lr_per_sample,
momentum_time_constant,
unit_gain=unit_gain_value)
C.set_default_unit_gain_value(False)
unit_gain_value = C.default_unit_gain_value()
assert not unit_gain_value
C.set_default_unit_gain_value(True)
unit_gain_value = C.default_unit_gain_value()
assert unit_gain_value
lr_per_sample = learning_rate_schedule([(3, 0.1), (2, 0.2), (1, 0.3)],
unit=UnitType.sample)
C.fsadagrad(res.parameters,
lr=lr_per_sample,
momentum=momentum_time_constant)
C.fsadagrad(res.parameters, lr_per_sample, momentum_time_constant,
unit_gain_value)
C.fsadagrad(res.parameters,
lr=lr_per_sample,
momentum=momentum_time_constant,
unit_gain=unit_gain_value)
gamma, inc, dec, max, min = [0.5, 1.2, 0.7, 10, 1e-8]
lr_per_sample = learning_rate_schedule([0.1, 0.2],
unit=UnitType.sample,
epoch_size=100)
C.rmsprop(res.parameters, lr_per_sample, gamma, inc, dec, max, min, True)
C.adadelta(res.parameters, lr_per_sample, use_mean_gradient=True)
def test_learner_init():
i = C.input_variable(shape=(1,), needs_gradient=True, name='a')
w = parameter(shape=(1,))
res = i * w
#test new API: learning_parameter_schedule
#explicitly specify reference minibatch size and learning rate is in number:
learner = sgd(res.parameters, lr=0.1, minibatch_size = 25)
assert learner.is_compatible_mode() == False
assert learner.minibatch_size == 25 #the learner's reference minibatch
#with direct learner learning rate number specification, the learning rate schedule get the reference minibatch size from the learner parameters:
assert learner._learning_rate_schedule.minibatch_size == 25
assert learner.learning_rate() == 0.1
#no explicitly specification of reference minibatch size and learning rate is in number:
learner = sgd(res.parameters, lr=learning_parameter_schedule(0.1))
assert learner.is_compatible_mode() == False
assert learner.minibatch_size == C.learners.IGNORE #the learner's reference minibatch
#with direct learner learning rate number specification, the learning rate schedule get the reference minibatch size from the learner parameters:
assert learner._learning_rate_schedule.minibatch_size == C.learners.IGNORE
assert learner.learning_rate() == 0.1
learner = sgd(res.parameters, lr=learning_parameter_schedule(0.1, 20), minibatch_size = 25)
assert learner.is_compatible_mode() == False
assert learner.minibatch_size == 25 #the learner's reference minibatch
#with direct learner learning rate number specification, the learning rate schedule get the reference minibatch size from the learner parameters:
assert learner._learning_rate_schedule.minibatch_size == 20
assert learner.learning_rate() == 0.1
learner = sgd(res.parameters, lr=learning_parameter_schedule(0.1, 20))
assert learner.is_compatible_mode() == False
#with direct learner learning rate number specification, the learning rate schedule get the reference minibatch size from the learner parameters:
assert learner._learning_rate_schedule.minibatch_size == 20
assert learner.learning_rate() == 0.1
#no explicitly specification of reference minibatch size and learning rate is in number:
learner = sgd(res.parameters, lr=learning_parameter_schedule(0.1))
assert learner.is_compatible_mode() == False
assert learner.minibatch_size == C.learners.IGNORE #the learner's reference minibatch
#with direct learner learning rate number specification, the learning rate schedule get the reference minibatch size from the learner parameters:
assert learner._learning_rate_schedule.minibatch_size == C.learners.IGNORE
assert learner.learning_rate() == 0.1
#no explicitly specification of reference minibatch size and learning rate is in number:
learner = sgd(res.parameters, lr=learning_parameter_schedule(0.1), minibatch_size=C.learners.IGNORE)
assert learner.is_compatible_mode() == True
assert learner.minibatch_size == C.learners.IGNORE #the learner's reference minibatch
#with direct learner learning rate number specification, the learning rate schedule get the reference minibatch size from the learner parameters:
assert learner._learning_rate_schedule.minibatch_size == C.learners.IGNORE
assert learner.learning_rate() == 0.1
learner = sgd(res.parameters, lr=learning_parameter_schedule(0.1, 20), minibatch_size=C.learners.IGNORE)
assert learner.is_compatible_mode() == True
assert learner.minibatch_size == C.learners.IGNORE #the learner's reference minibatch
#with direct learner learning rate number specification, the learning rate schedule get the reference minibatch size from the learner parameters:
assert learner._learning_rate_schedule.minibatch_size == 20
assert learner.learning_rate() == 0.1
#no explicitly specification of reference minibatch size and learning rate is in number:
learner = sgd(res.parameters, lr=learning_parameter_schedule(0.1), minibatch_size=C.learners.IGNORE)
assert learner.is_compatible_mode() == True
assert learner.minibatch_size == C.learners.IGNORE #the learner's reference minibatch
#with direct learner learning rate number specification, the learning rate schedule get the reference minibatch size from the learner parameters:
assert learner._learning_rate_schedule.minibatch_size == C.learners.IGNORE
assert learner.learning_rate() == 0.1
mysgd = C.sgd(parameters=res.parameters, lr=0.4, minibatch_size=32)
assert mysgd.minibatch_size == 32
assert mysgd._learning_rate_schedule.minibatch_size == 32
assert mysgd.learning_rate() == 0.4
mymomentum = C.momentum_sgd(parameters=res.parameters, lr=0.4, momentum=0.9, minibatch_size=32)
assert mymomentum.minibatch_size == 32
assert mymomentum._learning_rate_schedule.minibatch_size == 32
assert mymomentum.learning_rate() == 0.4
myadadelta = C.adadelta(parameters=res.parameters, lr=0.4, minibatch_size=32)
assert myadadelta.minibatch_size == 32
assert myadadelta._learning_rate_schedule.minibatch_size == 32
assert myadadelta.learning_rate() == 0.4
myadam = C.adam(parameters=res.parameters, lr=0.4, momentum=0.9, variance_momentum=0.9, minibatch_size=32)
assert myadam.minibatch_size == 32
assert myadam._learning_rate_schedule.minibatch_size == 32
assert myadam.learning_rate() == 0.4
myadagrad = C.adagrad(parameters=res.parameters, lr=0.4, minibatch_size=32)
assert myadagrad.minibatch_size == 32
assert myadagrad._learning_rate_schedule.minibatch_size == 32
assert myadagrad.learning_rate() == 0.4
myfsadagrad = C.fsadagrad(parameters=res.parameters, lr=0.4, momentum=0.9, variance_momentum=0.9,
minibatch_size=32)
assert myfsadagrad.minibatch_size == 32
assert myfsadagrad._learning_rate_schedule.minibatch_size == 32
assert myfsadagrad.learning_rate() == 0.4
mynesterov = C.nesterov(parameters=res.parameters, lr=0.4, momentum=0.9, minibatch_size=32)
assert mynesterov.minibatch_size == 32
assert mynesterov._learning_rate_schedule.minibatch_size == 32
assert mynesterov.learning_rate() == 0.4
myrmsrop = C.rmsprop(parameters=res.parameters, lr=0.4, gamma=0.5, inc=1.2, dec=0.7, max=10, min=1e-8,
minibatch_size=32)
assert myrmsrop.minibatch_size == 32
assert myrmsrop._learning_rate_schedule.minibatch_size == 32
assert myrmsrop.learning_rate() == 0.4
mysgd = C.sgd(parameters=res.parameters, lr=[0.4, 0.1, 0.001], minibatch_size=32, epoch_size=512)
assert mysgd.minibatch_size == 32
assert mysgd._learning_rate_schedule.minibatch_size == 32
assert mysgd._learning_rate_schedule[0] == 0.4
assert mysgd._learning_rate_schedule[512] == 0.1
assert mysgd._learning_rate_schedule[512 * 2] == 0.001
mymomentum = C.momentum_sgd(parameters=res.parameters, lr=[0.4, 0.1, 0.001], momentum=[0.9],
minibatch_size=32, epoch_size=512)
assert mymomentum.minibatch_size == 32
assert mymomentum._learning_rate_schedule.minibatch_size == 32
assert mymomentum._learning_rate_schedule[0] == 0.4
assert mymomentum._learning_rate_schedule[512] == 0.1
assert mymomentum._learning_rate_schedule[512 * 2] == 0.001
myadadelta = C.adadelta(parameters=res.parameters, lr=[0.4, 0.1, 0.001],
minibatch_size=32, epoch_size=512)
assert myadadelta.minibatch_size == 32
assert myadadelta._learning_rate_schedule.minibatch_size == 32
assert myadadelta._learning_rate_schedule[0] == 0.4
assert myadadelta._learning_rate_schedule[512] == 0.1
assert myadadelta._learning_rate_schedule[512 * 2] == 0.001
myadam = C.adam(parameters=res.parameters, lr=[0.4, 0.1, 0.001], momentum=[0.9, 0.1, 0.001], variance_momentum=[0.9],
minibatch_size=32, epoch_size=512)
assert myadam.minibatch_size == 32
assert myadam._learning_rate_schedule.minibatch_size == 32
assert myadam._learning_rate_schedule[0] == 0.4
assert myadam._learning_rate_schedule[512] == 0.1
assert myadam._learning_rate_schedule[512 * 2] == 0.001
myadagrad = C.adagrad(parameters=res.parameters, lr=[0.4, 0.1, 0.001], minibatch_size=32, epoch_size=512)
assert myadagrad.minibatch_size == 32
assert myadagrad._learning_rate_schedule.minibatch_size == 32
assert myadagrad._learning_rate_schedule[0] == 0.4
assert myadagrad._learning_rate_schedule[512] == 0.1
assert myadagrad._learning_rate_schedule[512 * 2] == 0.001
myfsadagrad = C.fsadagrad(parameters=res.parameters, lr=[0.4, 0.1, 0.001], momentum=[0.9],
variance_momentum=[0.9],
minibatch_size=32, epoch_size=512)
assert myadagrad.minibatch_size == 32
assert myadagrad._learning_rate_schedule.minibatch_size == 32
assert myadagrad._learning_rate_schedule[0] == 0.4
assert myadagrad._learning_rate_schedule[512] == 0.1
assert myadagrad._learning_rate_schedule[512 * 2] == 0.001
mynesterov = C.nesterov(parameters=res.parameters, lr=[0.4, 0.1, 0.001], momentum=[0.9],
minibatch_size=32, epoch_size=512)
assert mynesterov.minibatch_size == 32
assert mynesterov._learning_rate_schedule.minibatch_size == 32
assert mynesterov._learning_rate_schedule[0] == 0.4
assert mynesterov._learning_rate_schedule[512] == 0.1
assert mynesterov._learning_rate_schedule[512 * 2] == 0.001
myrmsrop = C.rmsprop(parameters=res.parameters, lr=[0.4, 0.1, 0.001], gamma=0.5, inc=1.2, dec=0.7, max=10,
min=1e-8,
minibatch_size=32, epoch_size=512)
assert myrmsrop.minibatch_size == 32
assert myrmsrop._learning_rate_schedule.minibatch_size == 32
assert myrmsrop._learning_rate_schedule[0] == 0.4
assert myrmsrop._learning_rate_schedule[512] == 0.1
assert myrmsrop._learning_rate_schedule[512 * 2] == 0.001
learner_parameter = learner.parameters
from cntk.variables import Parameter
param = learner_parameter[0]
assert isinstance(param, Parameter)
unit_gain_value = C.default_unit_gain_value()
assert unit_gain_value
momentum = C.momentum_schedule(0.999, minibatch_size=1)
lr_per_sample = learning_parameter_schedule(0.1, minibatch_size = 1)
C.momentum_sgd(res.parameters, lr_per_sample, momentum)
C.momentum_sgd(res.parameters, lr_per_sample, momentum, unit_gain_value)
C.momentum_sgd(res.parameters, lr_per_sample, momentum, unit_gain=unit_gain_value)
C.set_default_unit_gain_value(False)
unit_gain_value = C.default_unit_gain_value()
assert not unit_gain_value
lr_per_sample = learning_parameter_schedule([0.1, 0.2], minibatch_size = 1)
C.nesterov(res.parameters, lr=lr_per_sample, momentum=momentum)
C.nesterov(res.parameters, lr_per_sample, momentum, unit_gain_value)
C.nesterov(res.parameters, lr=lr_per_sample, momentum=momentum, unit_gain=unit_gain_value)
lr_per_sample = learning_parameter_schedule([0.1]*3 +[0.2]*2 +[0.3], minibatch_size=1)
C.adagrad(res.parameters, lr=lr_per_sample, need_ave_multiplier=True)
C.set_default_unit_gain_value(True)
unit_gain_value = C.default_unit_gain_value()
assert unit_gain_value
lr_per_sample = learning_parameter_schedule([(3,0.1), (2, 0.2), (1, 0.3)], minibatch_size=1)
C.fsadagrad(res.parameters, lr=lr_per_sample, momentum=momentum)
C.fsadagrad(res.parameters, lr_per_sample, momentum, unit_gain_value)
C.fsadagrad(res.parameters, lr=lr_per_sample, momentum=momentum, unit_gain=unit_gain_value)
gamma, inc, dec, max, min = [0.5, 1.2, 0.7, 10, 1e-8]
lr_per_sample = learning_parameter_schedule([0.1, 0.2], minibatch_size = 1, epoch_size = 100)
C.rmsprop(res.parameters, lr_per_sample, gamma, inc, dec, max, min, True)
C.adadelta(res.parameters, lr_per_sample)
def test_learner_init_legacy():
i = C.input_variable(shape=(1,), needs_gradient=True, name='a')
w = parameter(shape=(1,))
res = i * w
# for backcompatibility test
# this will be deprecated in future version
learner = sgd(res.parameters, lr=learning_rate_schedule(0.1, UnitType.sample))
assert learner._learning_rate_schedule.minibatch_size == 1 # the deprecated per sample schedule should not use compatible mode
assert learner.learning_rate() == 0.1
# for backcompatibility test
# this will be deprecated in future version
# The UnitType will provide per minibatch instruction for the learner
# this will be deprecated in future version
learner = sgd(res.parameters, lr=learning_rate_schedule(0.1, UnitType.minibatch))
assert learner.is_compatible_mode() == False
assert learner.learning_rate() == 0.1
assert learner.minibatch_size == C.learners.IGNORE
assert learner._learning_rate_schedule.minibatch_size == 0
# for backcompatibility test, in reset learning rate, the learner won't receive the reference minibatch size from the schedule
# user will need to specify the reference minibatch size explicitly
# this will be deprecated in future version
learner = sgd(res.parameters, lr=0.1)
learner.reset_learning_rate(learning_rate_schedule([1, 2, 3], UnitType.minibatch))
assert learner.learning_rate() == 1.0
learner.minibatch_size = C.learners.IGNORE # reset to be per minibatch
assert learner.minibatch_size == C.learners.IGNORE
assert learner._learning_rate_schedule.minibatch_size == C.learners.IGNORE
assert learner.is_compatible_mode() == True
# for backcompatibility test
# this will be deprecated in future version
learner = sgd(res.parameters, lr=learning_rate_schedule(0.1, UnitType.sample), minibatch_size=C.learners.IGNORE)
assert learner.is_compatible_mode() == True
assert learner.learning_rate() == 0.1
assert learner.minibatch_size == C.learners.IGNORE # the learner's reference minibatch size is still 0
# this will be deprecated in future version: This is logical invalid combination but it was the only way to use mean gradient and set learning rate in the past.
learner = sgd(res.parameters, lr=learning_rate_schedule(0.1, UnitType.sample), use_mean_gradient=True)
assert learner.is_compatible_mode() == True
assert learner.learning_rate() == 0.1
#test the override in the new version
assert learner._learning_rate_schedule.minibatch_size == C.learners.IGNORE
assert learner.minibatch_size == C.learners.IGNORE # the learner's reference minibatch size is still 0
# for backcompatibility test
# this will be deprecated in future version
# The UnitType will provide per minibatch instruction for the learner
# this will be deprecated in future version
learner = sgd(res.parameters, lr=learning_rate_schedule(0.1, UnitType.minibatch), minibatch_size=C.learners.IGNORE)
assert learner.is_compatible_mode() == True
assert learner.learning_rate() == 0.1
assert learner.minibatch_size == C.learners.IGNORE
assert learner._learning_rate_schedule.minibatch_size == C.learners.IGNORE
# for backcompatibility test, in reset learning rate, the learner won't receive the reference minibatch size from the schedule
# user will need to specify the reference minibatch size explicitly
# this will be deprecated in future version
learner = sgd(res.parameters, lr=0.1)
learner.reset_learning_rate(learning_rate_schedule([1, 2, 3], UnitType.minibatch))
assert learner.learning_rate() == 1.0
learner.minibatch_size = C.learners.IGNORE # reset to be per minibatch
assert learner.minibatch_size == C.learners.IGNORE
assert learner._learning_rate_schedule.minibatch_size == C.learners.IGNORE
assert learner.is_compatible_mode() == True
learner_parameter = learner.parameters
from cntk.variables import Parameter
param = learner_parameter[0]
assert isinstance(param, Parameter)
unit_gain_value = C.default_unit_gain_value()
assert unit_gain_value
# back compatible API test
momentum_time_constant = C.momentum_as_time_constant_schedule(1100)
lr_per_sample = learning_parameter_schedule(0.1, minibatch_size=1)
C.momentum_sgd(res.parameters, lr_per_sample, momentum_time_constant)
C.momentum_sgd(res.parameters, lr_per_sample, momentum_time_constant, unit_gain_value)
C.momentum_sgd(res.parameters, lr_per_sample, momentum_time_constant, unit_gain=unit_gain_value)
C.set_default_unit_gain_value(False)
unit_gain_value = C.default_unit_gain_value()
assert not unit_gain_value
C.set_default_unit_gain_value(True)
unit_gain_value = C.default_unit_gain_value()
assert unit_gain_value
lr_per_sample = learning_rate_schedule([(3, 0.1), (2, 0.2), (1, 0.3)], unit=UnitType.sample)
C.fsadagrad(res.parameters, lr=lr_per_sample, momentum=momentum_time_constant)
C.fsadagrad(res.parameters, lr_per_sample, momentum_time_constant, unit_gain_value)
C.fsadagrad(res.parameters, lr=lr_per_sample, momentum=momentum_time_constant, unit_gain=unit_gain_value)
gamma, inc, dec, max, min = [0.5, 1.2, 0.7, 10, 1e-8]
lr_per_sample = learning_rate_schedule([0.1, 0.2], unit=UnitType.sample, epoch_size=100)
C.rmsprop(res.parameters, lr_per_sample, gamma, inc, dec, max, min, True)
C.adadelta(res.parameters, lr_per_sample, use_mean_gradient=True)
