# Copyright 2020-2021 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.
# ============================================================================
"""Loss scale manager abstract class."""
from __future__ import absolute_import
from .._checkparam import Validator as validator
from .. import nn
[文档]class LossScaleManager:
"""
Loss scale (Magnification factor of gradients when mix precision is used) manager abstract class when using
mixed precision.
Derived class needs to implement all of its methods. `get_loss_scale` is used to get current loss scale value.
`update_loss_scale` is used to update loss scale value, `update_loss_scale` will be called during the training.
`get_update_cell` is used to get the instance of :class:`mindspore.nn.Cell` that is used to update the loss scale,
the instance will be called during the training. Currently, the `get_update_cell` is mostly used.
For example, :class:`mindspore.amp.FixedLossScaleManager` and :class:`mindspore.amp.DynamicLossScaleManager`.
"""
[文档] def get_loss_scale(self):
"""Get the value of loss scale, which is the amplification factor of the gradients."""
[文档] def update_loss_scale(self, overflow):
"""
Update the loss scale value according to the status of `overflow`.
Args:
overflow (bool): Whether the overflow occurs during the training.
"""
[文档] def get_update_cell(self):
"""Get the instance of :class:`mindspore.nn.Cell` that is used to update the loss scale."""
[文档]class FixedLossScaleManager(LossScaleManager):
"""
Loss scale(Magnification factor of gradients when mix precision is used) manager with a fixed loss scale value,
inherits from :class:`mindspore.amp.LossScaleManager`.
Args:
loss_scale (float): Magnification factor of gradients. Note that if `drop_overflow_update` is set to False,
the value of `loss_scale` in optimizer should be set to the same as here. Default: 128.0.
drop_overflow_update (bool): Whether to execute optimizer if there is an overflow. If True, the optimizer will
not executed when overflow occurs. Default: True.
Examples:
>>> import mindspore as ms
>>> from mindspore import nn
>>>
>>> net = Net()
>>> #1) Drop the parameter update if there is an overflow
>>> loss_scale_manager = ms.FixedLossScaleManager()
>>> optim = nn.Momentum(params=net.trainable_params(), learning_rate=0.1, momentum=0.9)
>>> model = ms.Model(net, loss_scale_manager=loss_scale_manager, optimizer=optim)
>>>
>>> #2) Execute parameter update even if overflow occurs
>>> loss_scale = 1024.0
>>> loss_scale_manager = ms.FixedLossScaleManager(loss_scale, False)
>>> optim = nn.Momentum(params=net.trainable_params(), learning_rate=0.1, momentum=0.9, loss_scale=loss_scale)
>>> model = ms.Model(net, loss_scale_manager=loss_scale_manager, optimizer=optim)
"""
def __init__(self, loss_scale=128.0, drop_overflow_update=True):
if loss_scale < 1:
raise ValueError("The argument 'loss_scale' must be >= 1, "
"but got {}".format(loss_scale))
self._loss_scale = loss_scale
self._drop_overflow_update = drop_overflow_update
[文档] def get_loss_scale(self):
"""
Get loss scale value.
Returns:
bool, `loss_scale` value.
"""
return self._loss_scale
[文档] def get_drop_overflow_update(self):
"""
Get `drop_overflow_update`, whether to drop optimizer update for current step when there is an overflow.
Returns:
bool, `drop_overflow_update` value.
"""
return self._drop_overflow_update
[文档] def update_loss_scale(self, overflow):
"""
Update loss scale value. The interface at `FixedLossScaleManager` will do nothing.
Args:
overflow (bool): Whether it overflows.
"""
[文档] def get_update_cell(self):
"""
Returns the instance of :class:`mindspore.nn.Cell` that used to update the loss scale which will be called at
:class:`mindspore.nn.TrainOneStepWithLossScaleCell`. As the loss scale is fixed in this class, the instance
will do nothing.
Returns:
None or :class:`mindspore.nn.FixedLossScaleUpdateCell`. Instance of
:class:`mindspore.nn.FixedLossScaleUpdateCell` when `drop_overflow_update` is True. None when
`drop_overflow_update` is False.
"""
if not self._drop_overflow_update:
return None
return nn.FixedLossScaleUpdateCell(self._loss_scale)
[文档]class DynamicLossScaleManager(LossScaleManager):
"""
Loss scale(Magnification factor of gradients when mix precision is used) manager with loss scale dynamically
adjusted, inherits from :class:`mindspore.amp.LossScaleManager`.
Args:
init_loss_scale (float): Initialize loss scale. Default: 2**24.
scale_factor (int): Coefficient of increase and decrease. Default: 2.
scale_window (int): Maximum continuous normal steps when there is no overflow. Default: 2000.
Examples:
>>> import mindspore as ms
>>> from mindspore import nn
>>>
>>> net = Net()
>>> loss_scale_manager = ms.DynamicLossScaleManager()
>>> optim = nn.Momentum(params=net.trainable_params(), learning_rate=0.1, momentum=0.9)
>>> model = ms.Model(net, loss_scale_manager=loss_scale_manager, optimizer=optim)
"""
def __init__(self,
init_loss_scale=2 ** 24,
scale_factor=2,
scale_window=2000):
if init_loss_scale < 1.0:
raise ValueError("The argument 'init_loss_scale' must be > 1, but got {}".format(init_loss_scale))
self.loss_scale = init_loss_scale
validator.check_positive_int(scale_window, "scale_window", self.__class__.__name__)
self.scale_window = scale_window
if scale_factor <= 0:
raise ValueError("The argument 'scale_factor' must be > 0, but got {}".format(scale_factor))
self.scale_factor = scale_factor
self.increase_ratio = scale_factor
self.decrease_ratio = 1 / scale_factor
self.cur_iter = 1
self.last_overflow_iter = 0
self.bad_step_max = 1000
self.bad_step = 0
[文档] def get_loss_scale(self):
"""
Get the current loss scale value.
Returns:
float, `loss_scale` value.
"""
return self.loss_scale
[文档] def update_loss_scale(self, overflow):
"""
Update the loss scale value according to the status of `overflow`. If overflow occurs, decrease loss scale per
`scale_window`, otherwise, increase the loss scale.
Args:
overflow (bool): Whether it overflows.
"""
if overflow:
self.loss_scale = max(self.loss_scale * self.decrease_ratio, 1)
self.last_overflow_iter = self.cur_iter
self.bad_step += 1
else:
if (self.cur_iter - self.last_overflow_iter) % self.scale_window == 0:
self.loss_scale *= self.increase_ratio
self.bad_step = 0
if self.bad_step > self.bad_step_max:
raise RuntimeError("Dynamic loss scale Continuous overflow ", self.bad_step,
" times, has exceeded maximum threshold.")
self.cur_iter += 1
[文档] def get_drop_overflow_update(self):
"""
Whether to drop optimizer update for current step when there is an overflow.
Returns:
bool, always True.
"""
return True
[文档] def get_update_cell(self):
"""
Returns the instance of :class:`mindspore.nn.Cell` that is used to update the loss scale which will be called at
:class:`mindspore.nn.TrainOneStepWithLossScaleCell`.
Returns:
:class:`mindspore.nn.DynamicLossScaleUpdateCell`.
"""
return nn.DynamicLossScaleUpdateCell(self.loss_scale, self.scale_factor, self.scale_window)