# Copyright 2020 Huawei Technologies Co., Ltd
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ============================================================================
"""FTRL"""
from mindspore.ops import functional as F, composite as C, operations as P
from mindspore.common import Tensor
import mindspore.common.dtype as mstype
from mindspore._checkparam import Validator as validator
from mindspore._checkparam import Rel
from .optimizer import Optimizer, _apply_decay, _grad_scale
_ftrl_opt = C.MultitypeFuncGraph("ftrl_opt")
@_ftrl_opt.register("Function", "Function", "Function", "Function", "Number", "Number", "Number", "Tensor", "Tensor",
"RowTensor", "Tensor", "Tensor", "Bool", "Bool")
def _tensor_run_opt_with_sparse(opt, spars_opt, push, pull, l1, l2, lr_power, learning_rate, linear,
gradient, weight, moment, ps_parameter, cache_enable):
"""Apply sparse ftrl optimizer to the weight parameter when the gradient is sparse."""
success = True
indices = gradient.indices
values = gradient.values
if ps_parameter and not cache_enable:
op_shape = P.Shape()
shapes = (op_shape(weight), op_shape(moment), op_shape(linear), op_shape(values), op_shape(indices))
success = F.depend(success, pull(push((values, indices), shapes), weight))
else:
success = F.depend(success, spars_opt(weight, moment, linear, values, indices))
return success
@_ftrl_opt.register("Function", "Function", "Function", "Function", "Number", "Number", "Number", "Tensor", "Tensor",
"Tensor", "Tensor", "Tensor", "Bool", "Bool")
def _tensor_run_opt(opt, spars_opt, push, pull, l1, l2, lr_power, learning_rate, linear,
gradient, weight, moment, ps_parameter, cache_enable):
"""Apply ftrl optimizer to the weight parameter."""
success = True
if ps_parameter and not cache_enable:
op_shape = P.Shape()
success = F.depend(success, pull(push((gradient, learning_rate, l1, l2, lr_power),
(op_shape(weight), op_shape(moment), op_shape(linear))), weight))
else:
success = F.depend(success, opt(weight, moment, linear, gradient, learning_rate, l1, l2, lr_power))
return success
def _check_param(initial_accum, lr_power, l1, l2, use_locking, prim_name=None):
"""Check param."""
validator.check_value_type("initial_accum", initial_accum, [float], prim_name)
validator.check_number("initial_accum", initial_accum, 0.0, Rel.GE, prim_name)
validator.check_value_type("lr_power", lr_power, [float], prim_name)
validator.check_number("lr_power", lr_power, 0.0, Rel.LE, prim_name)
validator.check_value_type("l1", l1, [float], prim_name)
validator.check_number("l1", l1, 0.0, Rel.GE, prim_name)
validator.check_value_type("l2", l2, [float], prim_name)
validator.check_number("l2", l2, 0.0, Rel.GE, prim_name)
validator.check_value_type("use_locking", use_locking, [bool], prim_name)
[docs]class FTRL(Optimizer):
r"""
Implements the FTRL algorithm with ApplyFtrl Operator.
FTRL is an online convex optimization algorithm that adaptively chooses its regularization function
based on the loss functions. Refer to paper `Adaptive Bound Optimization for Online Convex Optimization
<https://arxiv.org/abs/1002.4908>`_. Refer to paper `Ad Click Prediction: a View from the Trenches
<https://www.eecs.tufts.edu/~dsculley/papers/ad-click-prediction.pdf>`_ for engineering document.
The updating formulas are as follows,
.. math::
\begin{array}{ll} \\
m_{t+1} = m_{t} + g^2 \\
u_{t+1} = u_{t} + g - \frac{m_{t+1}^\text{-p} - m_{t}^\text{-p}}{\alpha } * \omega_{t} \\
\omega_{t+1} =
\begin{cases}
\frac{(sign(u_{t+1}) * l1 - u_{t+1})}{\frac{m_{t+1}^\text{-p}}{\alpha } + 2 * l2 }
& \text{ if } |u_{t+1}| > l1 \\
0.0
& \text{ otherwise }
\end{cases}\\
\end{array}
:math:`m` represents `accum`, :math:`g` represents `grads`, :math:`t` represents updating step,
:math:`u` represents `linear`, :math:`p` represents `lr_power`, :math:`\alpha` represents `learning_rate`,
:math:`\omega` represents `params`.
Note:
When separating parameter groups, the weight decay in each group will be applied on the parameters if the
weight decay is positive. When not separating parameter groups, the `weight_decay` in the API will be applied
on all of the parameters.
When separating parameter groups, if you want to centralize the gradient, set grad_centralization to True,
but the gradient centralization can only be applied to the parameters of the convolution layer.
If the parameters of the non convolution layer are set to True, an error will be reported.
To improve parameter groups performance, the customized order of parameters can be supported.
The sparse strategy is applied while the SparseGatherV2 operator being used for forward network.
The sparse feature is under continuous development. If the sparse strategy wants to be executed on the host,
set the target to the CPU.
Args:
params (Union[list[Parameter], list[dict]]): When the `params` is a list of `Parameter` which will be updated,
the element in `params` must be class `Parameter`. When the `params` is a list of `dict`, the "params",
"lr", "weight_decay" and "order_params" are the keys can be parsed.
- params: Required. The value must be a list of `Parameter`.
- lr: Using different learning rate by separating parameters is currently not supported.
- weight_decay: Optional. If "weight_decay" in the keys, the value of corresponding weight decay
will be used. If not, the `weight_decay` in the API will be used.
- order_params: Optional. If "order_params" in the keys, the value must be the order of parameters and
the order will be followed in optimizer. There are no other keys in the `dict` and the parameters which
in the value of 'order_params' must be in one of group parameters.
- grad_centralization: Optional. The data type of "grad_centralization" is Bool. If "grad_centralization"
is in the keys, the set value will be used. If not, the `grad_centralization` is False by default.
This parameter only works on the convolution layer.
initial_accum (float): The starting value for accumulators, must be zero or positive values. Default: 0.1.
learning_rate (float): The learning rate value, must be zero or positive, dynamic learning rate is currently
not supported. Default: 0.001.
lr_power (float): Learning rate power controls how the learning rate decreases during training, must be less
than or equal to zero. Use fixed learning rate if lr_power is zero. Default: -0.5.
l1 (float): l1 regularization strength, must be greater than or equal to zero. Default: 0.0.
l2 (float): l2 regularization strength, must be greater than or equal to zero. Default: 0.0.
use_locking (bool): If true, use locks for updating operation. Default: False.
loss_scale (float): Value for the loss scale. It must be greater than 0.0. In general, use the default value.
Only when `FixedLossScaleManager` is used for training and the `drop_overflow_update` in
`FixedLossScaleManager` is set to False, then this value needs to be the same as the `loss_scale` in
`FixedLossScaleManager`. Refer to class :class:`mindspore.FixedLossScaleManager` for more details.
Default: 1.0.
weight_decay (Union[float, int]): Weight decay value to multiply weight, must be zero or positive value.
Default: 0.0.
Inputs:
- **grads** (tuple[Tensor]) - The gradients of `params` in the optimizer, the shape is the same as the `params`
in optimizer.
Outputs:
tuple[Parameter], the updated parameters, the shape is the same as `params`.
Raises:
TypeError: If `initial_accum`, `learning_rate`, `lr_power`, `l1`, `l2` or `loss_scale` is not a float.
TypeError: If element of `parameters` is neither Parameter nor dict.
TypeError: If `weight_decay` is neither float nor int.
TypeError: If `use_nesterov` is not a bool.
ValueError: If `lr_power` is greater than 0.
ValueError: If `loss_scale` is less than or equal to 0.
ValueError: If `initial_accum`, `l1` or `l2` is less than 0.
Supported Platforms:
``Ascend`` ``GPU``
Examples:
>>> net = Net()
>>> #1) All parameters use the same learning rate and weight decay
>>> optim = nn.FTRL(params=net.trainable_params())
>>>
>>> #2) Use parameter groups and set different values
>>> conv_params = list(filter(lambda x: 'conv' in x.name, net.trainable_params()))
>>> no_conv_params = list(filter(lambda x: 'conv' not in x.name, net.trainable_params()))
>>> group_params = [{'params': conv_params, 'weight_decay': 0.01, 'grad_centralization':True},
... {'params': no_conv_params},
... {'order_params': net.trainable_params()}]
>>> optim = nn.FTRL(group_params, learning_rate=0.1, weight_decay=0.0)
>>> # The conv_params's parameters will use default learning rate of 0.1 and weight decay of 0.01 and grad
>>> # centralization of True.
>>> # The no_conv_params's parameters will use default weight decay of 0.0 and grad centralization of False.
>>> # The final parameters order in which the optimizer will be followed is the value of 'order_params'.
>>>
>>> loss = nn.SoftmaxCrossEntropyWithLogits()
>>> model = Model(net, loss_fn=loss, optimizer=optim)
"""
def __init__(self, params, initial_accum=0.1, learning_rate=0.001, lr_power=-0.5, l1=0.0, l2=0.0,
use_locking=False, loss_scale=1.0, weight_decay=0.0):
super(FTRL, self).__init__(learning_rate, params, weight_decay, loss_scale=loss_scale)
if self.dynamic_lr or self.is_group_lr:
raise ValueError('Dynamic learning rate or group learning rate is currently not supported.')
_check_param(initial_accum, lr_power, l1, l2, use_locking, self.cls_name)
self.moments = self.parameters.clone(prefix="moments", init=initial_accum)
self.linear = self.parameters.clone(prefix="linear", init='zeros')
self.l1 = l1
self.l2 = l2
self.lr = learning_rate
self.lr_power = lr_power
if not self.is_group:
self.decay_flags = tuple((lambda: True)() for x in self.parameters)
self.hyper_map = C.HyperMap()
self.opt = P.ApplyFtrl(use_locking=use_locking)
self.use_locking = use_locking
self.sparse_opt = P.SparseApplyFtrl(learning_rate, l1, l2, lr_power, use_locking=use_locking)
self._ps_pull = P.Pull()
self._ps_push = P.Push("Ftrl", [0, 1, 2])
self._ps_push.add_prim_attr("init_accum", initial_accum)
self._ps_push.add_prim_attr("lr", learning_rate)
self._ps_push.add_prim_attr("l1", l1)
self._ps_push.add_prim_attr("l2", l2)
self._ps_push.add_prim_attr("lr_power", lr_power)
def construct(self, grads):
params = self.parameters
moments = self.moments
linear = self.linear
grads = self.decay_weight(grads)
grads = self.scale_grad(grads)
grads = self._grad_sparse_indices_deduplicate(grads)
grads = self.gradients_centralization(grads)
lr = self.get_lr()
success = self.map_(F.partial(_ftrl_opt, self.opt, self.sparse_opt, self._ps_push, self._ps_pull,
self.l1, self.l2, self.lr_power, lr),
linear, grads, params, moments, self.ps_parameters, self.cache_enable)
return success
@Optimizer.target.setter
def target(self, value):
"""
If the input value is set to "CPU", the parameters will be updated on the host using the Fused
optimizer operation."""
if not isinstance(value, str):
raise TypeError("The value must be str type, but got value type is {}".format(type(value)))
if value not in ('CPU', 'Ascend', 'GPU'):
raise ValueError("The value must be 'CPU', 'Ascend' or 'GPU', but got value {}".format(value))
if value == 'CPU':
self.sparse_opt = P.FusedSparseFtrl(self.lr, self.l1, self.l2, self.lr_power, self.use_locking)
self.sparse_opt.add_prim_attr("primitive_target", "CPU")
else:
self.sparse_opt = P.SparseApplyFtrl(self.lr, self.l1, self.l2, self.lr_power, self.use_locking)
self._target = value