# Copyright 2023 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.
# ============================================================================
"""Configuration of parameters for strategy-searching algorithm in auto_parallel"""
from __future__ import absolute_import
from __future__ import division
import threading
from mindspore._c_expression import CostModelContext
from mindspore._checkparam import args_type_check
__all__ = ["get_algo_parameters", "reset_algo_parameters", "set_algo_parameters"]
_PARAMETER_CONFIG = None
class _AlgoParameterConfig:
"""
_AlgoParameterConfig is the configuration of setting parameters used in th algorithm.
Note:
Creating a config through instantiating _AlgoParameterConfig object is not recommended.
Use algo_parameter_config() to get the configuration since _AlgoParameterConfig is singleton.
"""
_instance = None
_instance_lock = threading.Lock()
def __init__(self):
self._config_handle = CostModelContext.get_instance()
def check_config_handle(self):
"""
Check config handle.
Raises:
ValueError: If the config handle is none.
"""
if self._config_handle is None:
raise ValueError("Config handle is none!!!")
def set_fully_use_devices(self, not_fully):
"""
Set the flag of whether only generating strategies that fully use all available devices.
Default: ``True``
Args:
not_fully (bool): The flag.
"""
self.check_config_handle()
self._config_handle.set_fully_use_devices(not_fully)
def get_fully_use_devices(self):
"""
Get the flag of whether only generating strategies that fully use all available devices.
Return:
The flag.
"""
self.check_config_handle()
return self._config_handle.get_fully_use_devices()
def set_elementwise_op_strategy_follow(self, element_strategy_follow):
"""
Set the flag of whether the elementwise operator has the same strategies as its subsequent operators.
Default: False
Args:
element_strategy_follow (bool): The flag.
"""
self.check_config_handle()
self._config_handle.set_elementwise_op_strategy_follow(element_strategy_follow)
def get_elementwise_op_strategy_follow(self):
"""
Get the flag of whether the elementwise operator has the same strategies as its subsequent operators.
Returns:
The flag.
"""
self.check_config_handle()
return self._config_handle.get_elementwise_op_strategy_follow()
def set_tensor_slice_align_enable(self, align_enable):
"""
Set the flag of whether to check the shape of tensor slice of MatMul.
Default: False
Args:
align_enable (bool): The flag.
"""
self.check_config_handle()
self._config_handle.set_tensor_slice_align_enable(align_enable)
def get_tensor_slice_align_enable(self):
"""
Get the flag of whether to check the shape of tensor slice of MatMul.
Returns:
The flag.
"""
self.check_config_handle()
return self._config_handle.get_tensor_slice_align_enable()
def set_tensor_slice_align_size(self, align_size):
"""
Set tensor slice align size.
Args:
align_size (int): The minimum tensor slice shape.
Raises:
ValueError: If align_size is not in [1, 1024].
"""
self.check_config_handle()
if align_size < 1 or align_size > 1024:
raise ValueError('Align_size must be in [1, 1024], but got {}'.format(align_size))
self._config_handle.set_tensor_slice_align_size(align_size)
def get_tensor_slice_align_size(self):
"""
Get the tensor slice align size.
Returns:
The size.
"""
self.check_config_handle()
return self._config_handle.get_tensor_slice_align_size()
def set_dp_algo_enable_approxi(self, enable_flag):
"""
Set the flag of whether to enable the approximation in the DP algorithms.
Default: ``False``.
Args:
enable_flag (bool): The flag.
"""
self.check_config_handle()
self._config_handle.set_dp_algo_enable_approxi(enable_flag)
def get_dp_algo_enable_approxi(self):
"""
Get the flag of whether to enable the approximation in the DP algorithms.
Returns:
The flag.
"""
self.check_config_handle()
return self._config_handle.get_dp_algo_enable_approxi()
def set_dp_algo_approxi_epsilon(self, epsilon):
"""
Set the epsilon value used in the approximation DP algorithm.
Default: 0.1.
Args:
epsilon (float): The epsilon value, should in the range dp_(0, 1].
"""
self.check_config_handle()
self._config_handle.set_dp_algo_approxi_epsilon(epsilon)
def get_dp_algo_approxi_epsilon(self):
"""
Get the epsilon value used in the approximation DP algorithm.
Returns:
The epsilon value.
"""
self.check_config_handle()
return self._config_handle.get_dp_algo_approxi_epsilon()
def reset_algo_parameters(self):
"""
Reset algorithm parameter attributes.
"""
self.check_config_handle()
self._config_handle.reset_algo_parameters()
_G_ALGO_PARAMETER_CONFIG = None
def _algo_parameter_config():
"""
Get the global _G_ALGO_PARAMETER_CONFIG. If it is not created, create a new one.
Returns:
The global _G_ALGO_PARAMETER_CONFIG.
"""
global _G_ALGO_PARAMETER_CONFIG
if _G_ALGO_PARAMETER_CONFIG is None:
_G_ALGO_PARAMETER_CONFIG = _AlgoParameterConfig()
return _G_ALGO_PARAMETER_CONFIG
set_algo_parameters_config_func_map = {
"fully_use_devices": _algo_parameter_config().set_fully_use_devices,
"elementwise_op_strategy_follow": _algo_parameter_config().set_elementwise_op_strategy_follow,
"tensor_slice_align_enable": _algo_parameter_config().set_tensor_slice_align_enable,
"tensor_slice_align_size": _algo_parameter_config().set_tensor_slice_align_size,
"enable_algo_approxi": _algo_parameter_config().set_dp_algo_enable_approxi,
"algo_approxi_epsilon": _algo_parameter_config().set_dp_algo_approxi_epsilon}
get_algo_parameters_config_func_map = {
"fully_use_devices": _algo_parameter_config().get_fully_use_devices,
"elementwise_op_strategy_follow": _algo_parameter_config().get_elementwise_op_strategy_follow,
"tensor_slice_align_enable": _algo_parameter_config().get_tensor_slice_align_enable,
"tensor_slice_align_size": _algo_parameter_config().get_tensor_slice_align_size,
"enable_algo_approxi": _algo_parameter_config().get_dp_algo_enable_approxi,
"algo_approxi_epsilon": _algo_parameter_config().get_dp_algo_approxi_epsilon}
[docs]@args_type_check(tensor_slice_align_enable=bool, tensor_slice_align_size=int,
fully_use_devices=bool, elementwise_op_strategy_follow=bool,
enable_algo_approxi=bool, algo_approxi_epsilon=float)
def set_algo_parameters(**kwargs):
"""
Set parameters in the algorithm for parallel strategy searching. See a typical use in
`test_auto_parallel_resnet.py
<https://gitee.com/mindspore/mindspore/blob/r2.3.q1/tests/ut/python/parallel/test_auto_parallel_resnet.py>`_.
Note:
The attribute name is required. This interface works ONLY in AUTO_PARALLEL mode.
Args:
fully_use_devices (bool): Whether ONLY searching strategies that fully use all available devices.
Default: ``True`` . For example with 8 devices available, if set ``True`` , strategy (4, 1) will not be
included in ReLU's candidate strategies, because strategy (4, 1) only utilizes 4 devices.
elementwise_op_strategy_follow (bool): Whether the elementwise operator has the consistent strategies as its
subsequent operators. Elementwise operators refer to operators that operate on input element by element,
such as Add, ReLU, etc. Default: ``False`` . For the example of ReLU followed by Add, if this flag is set
``True`` , then the searched strategy by the algorithm guarantees that strategies of these two operators
are consistent, e.g., ReLU's strategy (8, 1) and Add's strategy ((8, 1), (8, 1)).
enable_algo_approxi (bool): Whether to enable the approximation in the algorithms. Default: ``False`` . Due to
large solution space in searching parallel strategy for large DNN model, the algorithm takes fairly long
time in this case. To mitigate it, if this flag is set ``True`` , an approximation is made to discard some
candidate strategies, so that the solution space is shrunken.
algo_approxi_epsilon (float): The epsilon value used in the approximation algorithm. Default: ``0.1`` . This
value describes the extent of approximation. For example, the number of candidate strategies of an operator
is S, if 'enable_algo_approxi' is ``True`` , then the remaining strategies is of size: min{S, 1/epsilon}.
tensor_slice_align_enable (bool): Whether to check the shape of tensor slice of MatMul. Default: ``False`` .
Due to properties of some hardware, MatMul kernel only with large shapes can show advantages. If this flag
is ``True`` , then the slice shape of MatMul is checked to prevent irregular shapes.
tensor_slice_align_size (int): The minimum tensor slice shape of MatMul, the value must be in [1, 1024].
Default: ``16`` . If 'tensor_slice_align_enable' is set ``True`` , then the slice size of last dimension of
MatMul tensors should be multiple of this value.
Raises:
ValueError: If context keyword is not recognized.
Examples:
.. note::
Before running the following examples, you need to configure the communication environment variables.
For the Ascend devices, users need to prepare the rank table, set rank_id and device_id.
Please see the `rank table startup
<https://www.mindspore.cn/tutorials/experts/en/r2.3.0rc1/parallel/rank_table.html>`_
for more details.
For the GPU devices, users need to prepare the host file and mpi, please see the `mpirun startup
<https://www.mindspore.cn/tutorials/experts/en/r2.3.0rc1/parallel/mpirun.html>`_ .
For the CPU device, users need to write a dynamic cluster startup script, please see the `Dynamic Cluster
Startup <https://www.mindspore.cn/tutorials/experts/en/r2.3.0rc1/parallel/dynamic_cluster.html>`_ .
>>> import numpy as np
>>> import mindspore as ms
>>> import mindspore.dataset as ds
>>> from mindspore import nn, ops, train
>>> from mindspore.communication import init
>>> from mindspore.common.initializer import initializer
>>>
>>> ms.set_context(mode=ms.GRAPH_MODE)
>>> ms.set_auto_parallel_context(parallel_mode=ms.ParallelMode.AUTO_PARALLEL,
... search_mode="sharding_propagation")
>>> init()
>>> ms.set_algo_parameters(fully_use_devices=True)
>>> ms.set_algo_parameters(elementwise_op_strategy_follow=True)
>>> ms.set_algo_parameters(enable_algo_approxi=True)
>>> ms.set_algo_parameters(algo_approxi_epsilon=0.2)
>>> ms.set_algo_parameters(tensor_slice_align_enable=True)
>>> ms.set_algo_parameters(tensor_slice_align_size=8)
>>>
>>> # Define the network structure.
>>> class Dense(nn.Cell):
... def __init__(self, in_channels, out_channels):
... super().__init__()
... self.weight = ms.Parameter(initializer("normal", [in_channels, out_channels], ms.float32))
... self.bias = ms.Parameter(initializer("normal", [out_channels], ms.float32))
... self.matmul = ops.MatMul()
... self.add = ops.Add()
...
... def construct(self, x):
... x = self.matmul(x, self.weight)
... x = self.add(x, self.bias)
... return x
>>>
>>> class FFN(nn.Cell):
... def __init__(self):
... super().__init__()
... self.flatten = ops.Flatten()
... self.dense1 = Dense(28*28, 64)
... self.relu = ops.ReLU()
... self.dense2 = Dense(64, 10)
...
... def construct(self, x):
... x = self.flatten(x)
... x = self.dense1(x)
... x = self.relu(x)
... x = self.dense2(x)
... return x
>>> net = FFN()
>>> net.dense1.matmul.shard(((2, 1), (1, 2)))
>>>
>>> # Create dataset.
>>> step_per_epoch = 16
>>> def get_dataset(*inputs):
... def generate():
... for _ in range(step_per_epoch):
... yield inputs
... return generate
>>>
>>> input_data = np.random.rand(1, 28, 28).astype(np.float32)
>>> label_data = np.random.rand(1).astype(np.int32)
>>> fake_dataset = get_dataset(input_data, label_data)
>>> dataset = ds.GeneratorDataset(fake_dataset, ["input", "label"])
>>> # Train network.
>>> optimizer = nn.Momentum(net.trainable_params(), 1e-3, 0.1)
>>> loss_fn = nn.CrossEntropyLoss()
>>> loss_cb = train.LossMonitor()
>>> model = ms.Model(network=net, loss_fn=loss_fn, optimizer=optimizer)
>>> model.train(epoch=2, train_dataset=dataset, callbacks=[loss_cb])
"""
for key, value in kwargs.items():
if key not in set_algo_parameters_config_func_map:
raise ValueError("Set context keyword %s is not recognized!" % key)
set_func = set_algo_parameters_config_func_map[key]
set_func(value)
[docs]def get_algo_parameters(attr_key):
"""
Get the algorithm parameter config attributes.
Note:
The attribute name is required. This interface works ONLY in AUTO_PARALLEL mode.
Args:
attr_key (str): The key of the attribute. The keys include: "fully_use_devices",
"elementwise_op_strategy_follow", "enable_algo_approxi", "algo_approxi_epsilon",
"tensor_slice_align_enable","tensor_slice_align_size".
See :func:`mindspore.set_algo_parameters` for more details about the meaning of the attributes.
Returns:
Return attribute value according to the key.
Raises:
ValueError: If context keyword is not recognized.
Examples:
>>> import mindspore as ms
>>> ms.get_algo_parameters("fully_use_devices")
True
"""
if attr_key not in get_algo_parameters_config_func_map:
raise ValueError("Get context keyword %s is not recognized!" % attr_key)
get_func = get_algo_parameters_config_func_map[attr_key]
return get_func()
[docs]def reset_algo_parameters():
"""Reset the algorithm parameter attributes.
Note:
This interface works ONLY in AUTO_PARALLEL mode.
After reset, the values of the attributes are:
- fully_use_devices: True.
- elementwise_op_strategy_follow: False.
- enable_algo_approxi: False.
- algo_approxi_epsilon: 0.1.
- tensor_slice_align_enable: False.
- tensor_slice_align_size: 16.
Examples:
>>> import mindspore as ms
>>> ms.reset_algo_parameters()
"""
_algo_parameter_config().reset_algo_parameters()