# 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.
# ============================================================================
"""primitive"""
import functools
import inspect
import copy
from mindspore.common.api import _wrap_func
from mindspore.log import _LogActionOnce
from mindspore import context, log as logger
from mindspore.parallel._utils import _is_in_auto_parallel_mode, _is_in_data_parallel_mode, _is_in_hybrid_parallel_mode
from mindspore.parallel._ps_context import _is_ps_mode, _is_role_sched
from mindspore.common.parameter import Parameter
from mindspore.common.api import _pynative_executor
from mindspore._c_expression import Primitive_, prim_type
from mindspore._checkparam import Validator
from mindspore.ops import signature as sig
[文档]class Primitive(Primitive_):
"""
Primitive is the base class of operator primitives in python.
Args:
name (str): Name for the current Primitive.
Examples:
>>> from mindspore.ops import prim_attr_register, Primitive
>>> add = Primitive('add')
>>>
>>> # or work with prim_attr_register:
>>> # init a Primitive class with attr1 and attr2
>>> class Add(Primitive):
... @prim_attr_register
... def __init__(self, attr1, attr2):
... '''init for add'''
... # check attr1 and attr2 or do some initializations
... # init a Primitive obj with attr1=1 and attr2=2
>>> add = Add(attr1=1, attr2=2)
"""
_repr_ignore_list = ['input_names', 'output_names']
def __init__(self, name):
self.name = name
self.attrs = {}
self.init_attrs = {"name": name}
self._update_parameter = False
Primitive_.__init__(self, name)
if hasattr(self.__class__, '__mindspore_signature__'):
out = self._fill_signature(self.__class__.__mindspore_signature__)
self.set_signatures(out)
[文档] def add_prim_attr(self, name, value):
"""
Add primitive attribute.
Args:
name (str): Attribute Name.
value (Any): Attribute value.
Examples:
>>> import mindspore.ops as ops
>>> a = ops.Add()
>>> a = a.add_prim_attr("attr",1)
>>> out = a.attrs["attr"]
>>> print(out)
1
"""
self.__dict__[name] = value
self.attrs[name] = value
self.add_attr(name, value)
return self
[文档] def set_device(self, device_target):
"""
Set primitive been executed device.
Args:
device_target (str): The target device to run, support "Ascend", "GPU", and "CPU".
Examples:
>>> import mindspore.ops as ops
>>> a = ops.Add()
>>> a = a.set_device("GPU")
>>> print(a.primitive_target)
GPU
"""
return self.add_prim_attr("primitive_target", device_target)
def _fill_signature(self, signatures):
"""fills signature."""
signatures_new = []
for signature in signatures:
if isinstance(signature, sig.Signature):
signatures_new.append(signature)
elif isinstance(signature, sig.sig_dtype):
signatures_new.append(sig.make_sig(dtype=signature))
else:
if len(signature) < 3:
raise ValueError(f"[Internal Error]Signature for one parameter len must > 3, but {signature}")
signatures_new.append(sig.make_sig(*signature))
return tuple(signatures_new)
def _clone(self):
"""
Deeply clones the primitive object.
Calls the __init__() method with the same arguments. This method is called in parser if the
flag self.__setattr_flag__ is True.
"""
cloned = copy.deepcopy(self)
init_params = list()
if hasattr(cloned.__init__, 'decorated_func'):
init_params = inspect.getfullargspec(cloned.__init__.decorated_func).args[1:]
init_args = self.init_attrs
for name in init_params:
value = self.attrs[name]
init_args[name] = value
# __init__ should be called to construct cpp object.
cloned.__init__(**init_args)
for name in self.attrs:
value = self.attrs[name]
cloned.add_prim_attr(name, value)
if hasattr(self, 'instance_name'):
cloned.set_prim_instance_name(self.instance_name)
return cloned
[文档] def del_prim_attr(self, name):
"""
Delete primitive attribute.
Args:
name (str): Attribute Name.
Examples:
>>> import mindspore.ops as ops
>>> a = ops.Add()
>>> a = a.add_prim_attr("attr",1)
>>> a = a.del_prim_attr("attr")
>>> print(a.attrs)
{'input_names': ['x', 'y'], 'output_names' : ['output']}
"""
if name in self.__dict__ and name in self.attrs:
del self.__dict__[name]
del self.attrs[name]
self.del_attr(name)
return self
[文档] def set_stage(self, stage):
"""
Add stage id to primitive attribute.
Note:
It is valid only in semi auto parallel.
In other parallel modes, please set it to be 0.
Args:
stage (int): The stage id for the current operation.
Examples:
>>> from mindspore import ops
>>> add = ops.Add()
>>> print(add.set_stage(0))
Prim[Add]<stage=0>
"""
self.add_prim_attr("stage", stage)
return self
# The decorator has been deleted.
[文档] def shard(self, in_strategy=None, out_strategy=None):
"""
Add strategies to primitive attribute.
Note:
It is valid only in semi auto parallel or auto parallel mode.
In other parallel modes, strategies set here will be ignored.
Args:
in_strategy (tuple): Describe the split strategy of operator input. Default: None.
out_strategy (tuple): Describe the split strategy of operator output, it is only for certain operators,
such as MatMul. Default: None.
Examples:
>>> from mindspore import ops
>>> add = ops.Add()
>>> print(add.shard(((1, 1), (1, 1))))
Prim[Add]<in_strategy=((1, 1), (1, 1)), out_strategy=None>
"""
mode = context.get_auto_parallel_context("parallel_mode")
if in_strategy is not None:
if not isinstance(in_strategy, tuple):
raise TypeError(f'in_strategy must be tuple type, but got:{type(in_strategy)}')
for in_ele in in_strategy:
if not isinstance(in_ele, tuple):
raise TypeError(f'The element of strategy must be tuple type, but got:{type(in_ele)}')
for in_value in in_ele:
if not isinstance(in_value, int):
raise TypeError(f'The in_strategy: {in_strategy} of {self.name} is not valid,'
f' the value of strategy must be int type, but got:{type(in_value)}')
if out_strategy is not None:
if not isinstance(out_strategy, tuple):
raise TypeError(f'out strategy must be tuple type, but got:{type(out_strategy)}')
for out_ele in out_strategy:
if not isinstance(out_ele, tuple):
raise TypeError(f'The element of strategy must be tuple type, but got:{type(out_ele)}')
for out_value in out_ele:
if not isinstance(out_value, int):
raise TypeError(f'The in_strategy: {out_strategy} of {self.name} is not valid,'
f' the value of strategy must be int type, but got:{type(out_value)}')
if in_strategy is None and out_strategy is not None:
raise ValueError(f'The out_strategy of {self.name} is {out_strategy}, need to set in_strategy,'
f' but got none')
if not _is_in_auto_parallel_mode():
if in_strategy is not None:
logger.warning(f"The in_strategy of the operator in your network will not take effect in {mode} mode. "
f"This means the the shard function called in the network is ignored. \n"
f"If you want to enable it, please use semi auto or auto parallel mode by "
f"context.set_auto_parallel_context(parallel_mode=ParallelMode.SEMI_AUTO_PARALLEL "
f"or context.set_auto_parallel_context(parallel_mode=ParallelMode.AUTO_PARALLEL)")
if out_strategy is not None:
logger.warning(f"The out_strategy of the operator in your network will not take effect in {mode} mode."
f" This means the the shard function called in the network is ignored. \n"
f"If you want to enable it, please use semi auto or auto parallel mode by "
f"context.set_auto_parallel_context(parallel_mode=ParallelMode.SEMI_AUTO_PARALLEL "
f"or context.set_auto_parallel_context(parallel_mode=ParallelMode.AUTO_PARALLEL)")
self.add_prim_attr("in_strategy", in_strategy)
self.add_prim_attr("out_strategy", out_strategy)
return self
[文档] def set_prim_instance_name(self, instance_name):
"""
Set instance name to primitive operator.
Note:
It will be called by default when user defines primitive operator.
Args:
instance_name (str): Instance name of primitive operator set by user.
Examples:
>>> import mindspore.ops as ops
>>> a = ops.Add()
>>> a = a.set_prim_instance_name("add")
>>> print(a.instance_name)
add
"""
self.set_instance_name(instance_name)
self.instance_name = instance_name
return self
def __getattr__(self, item):
if item == 'infer_dynamic_shape':
return None
if item in super().get_attr_dict():
return super().get_attr_dict()[item]
if item in self.attrs:
return self.attrs[item]
err_msg = "'{prim}' object has no attribute '{attr}'".format(prim=self.name, attr=item)
raise AttributeError(err_msg)
[文档] def check_elim(self, *args):
"""
Check if the primitive can be eliminated. Subclass in need should override this method.
Args:
args(Primitive args): Same as arguments of current Primitive.
Returns:
A tuple consisting of two elements.
The first element means if the primitive can be calculated in compiling stage,
the second element is calculated result.
Examples:
>>> import numpy as np
>>> import mindspore
>>> from mindspore import Tensor
>>> from mindspore.ops import prim_attr_register, Primitive
>>> class AddN(Primitive):
... @prim_attr_register
... def __init__(self):
... self.init_prim_io_names(inputs=["inputs"], outputs=["sum"])
... def check_elim(self, inputs):
... if len(inputs) != 1:
... return (False, None)
... if isinstance(inputs[0], Tensor):
... return (True, inputs[0])
...
>>> addn = AddN()
>>> input_x = Tensor(np.array([1, 2, 3]), mindspore.float32)
>>> output = addn.check_elim((input_x,))
>>> print(output)
(True, Tensor(shape=[3], dtype=Float32, value= [ 1.00000000e+00, 2.00000000e+00, 3.00000000e+00]))
"""
return (False, None)
def __call__(self, *args):
should_elim, output = self.check_elim(*args)
for arg in args:
if isinstance(arg, Parameter) and arg.has_init:
arg.init_data()
if should_elim:
return output
return _run_op(self, self.name, args)
def __getstate__(self):
return self.__dict__
def __setstate__(self, d):
self.__dict__.update(d)
def __deepcopy__(self, memo):
return type(self)(**self.init_attrs)
def __repr__(self):
attr = ', '.join([f'{k}={self.attrs.get(k)}' for k in self.attrs if k not in Primitive._repr_ignore_list])
info_str = f'Prim[{self.name}]'
if attr:
info_str += f'<{attr}>'
return info_str
[文档] def init_prim_io_names(self, inputs, outputs):
"""
Initialize the name of inputs and outputs of Tensor or attributes.
Args:
inputs (list[str]): list of inputs names.
outputs (list[str]): list of outputs names.
Examples:
>>> import mindspore.ops as ops
>>> a = ops.Add()
>>> a.init_prim_io_names(["x","y"],["sum"])
>>> print(a.input_names)
['x','y']
>>> print(a.output_names)
['sum']
"""
# for checking para names with kernel implementation
self.add_prim_attr("input_names", inputs)
# for checking output number with kernel implementation
self.add_prim_attr("output_names", outputs)
@property
def update_parameter(self):
"""Return whether the primitive will update the value of parameter."""
return self._update_parameter
[文档] def recompute(self, mode=True):
"""
Set the primitive recomputed. If a primitive set recomputed feeds into some backward nodes
for computing gradient, rather than storing the intermediate activation computed in forward
pass, we will recompute it in backward pass.
Note:
- If the computation involves something like randomization or global variable, the equivalence
is not guaranteed currently.
- Not supported in pynative mode
Args:
mode (bool): Specifies whether the primitive is recomputed. Default: True.
Examples:
>>> import numpy as np
>>> import mindspore as ms
>>> from mindspore import Tensor, ops, nn
>>> class NetRecompute(nn.Cell):
... def __init__(self):
... super(NetRecompute,self).__init__()
... self.relu = ops.ReLU().recompute()
... self.sqrt = ops.Sqrt()
... def construct(self, x):
... out = self.relu(x)
... return self.sqrt(out)
...
>>> class GradNet(nn.Cell):
... def __init__(self, network):
... super(GradNet,self).__init__()
... self.network = network
... self.grad = ops.GradOperation()
... def construct(self, x):
... g_out = self.grad(self.network)(x)
... return g_out
...
>>> x = Tensor(np.array([-1,1]).astype(np.float32))
>>> net = NetRecompute()
>>> grad = GradNet(net)
>>> a = grad(x)
>>> print(a)
[0. 0.5]
"""
if context.get_context("mode") == context.PYNATIVE_MODE:
raise TypeError("Recompute is not supported in pynative mode currently.")
Validator.check_bool(mode)
self.add_prim_attr("recompute", mode)
return self
[文档] def place(self, role, rank_id):
"""
Set the label for this primitive.
This label tells MindSpore compiler on which process this operator should be launched.
And each process's identical label consists of input 'role' and 'rank_id'.
So by setting different operators with different labels,
which will be launched on different processes, users can launch a distributed training job.
Note:
- This method is effective only after
"mindspore.communication.init()" is called for dynamic cluster building.
Args:
role (str): The role of the process on which this operator will be launched.
Only 'MS_WORKER' is supported for now.
rank_id (int): The rank id of the process on which this operator will be launched.
The rank_id is unique in processes with the same role.
Examples:
>>> from mindspore import context
>>> import mindspore.ops as ops
>>> context.set_context(mode=context.GRAPH_MODE)
>>> matmul = ops.MatMul()
>>> matmul.place('MS_WORKER', 0)
"""
if _is_role_sched():
return
Validator.check_non_negative_int(rank_id, "rank_id", "Primitive.place")
Validator.check_string(role, "MS_WORKER", "role", "Primitive.place")
if context.get_context("mode") == context.PYNATIVE_MODE:
raise RuntimeError("You are calling Primitive.place in pynative mode."
"It's only supported in graph mode. Please switch to graph mode.")
# Get the execution context and check whether calling of this 'place' method is valid.
# This is because placing operators to arbitrary processes while other distributed training mode
# is enabled is very unpredictable and may cause fatal error.
# Some of these cases are under development and others should not be supported.
if _is_ps_mode():
raise RuntimeError(
"You are calling Primitive.place mixed with Parameter Server training. "
"This case is not supported yet. "
"Please call Primitive.place without Parameter Server training.")
if _is_in_auto_parallel_mode() or _is_in_data_parallel_mode() or _is_in_hybrid_parallel_mode():
raise RuntimeError(
"You are calling Primitive.place mixed with other parallel features: "
"'auto_parallel', 'data_parallel' and 'hybrid_parallel'. "
"This case is still under development and not supported yet. "
"Please call Primitive.place without these features.")
self.add_prim_attr("ms_role", role)
self.add_prim_attr("rank_id", rank_id)
[文档]class PrimitiveWithCheck(Primitive):
"""
PrimitiveWithCheck is the base class of primitives in python, which defines functions to check the input arguments
of operators, but uses the infer method registered in c++ source codes.
There are three methods can be overridden to define the check logic of the primitive: __check__(), check_shape(),
check_dtype(). If __check__() is defined in primitive, the __check__() has the highest priority to be called.
If __check__() is not defined, check_shape() and check_dtype() can be defined to describe the check logic of
the shape and type. Method infer_value() can also be defined (such as PrimitiveWithInfer) for constant propagation.
Args:
name (str): Name of the current Primitive.
Supported Platforms:
``Ascend`` ``GPU`` ``CPU``
Examples:
>>> from mindspore import dtype as mstype
>>> from mindspore.ops import prim_attr_register, PrimitiveWithCheck
>>> # init a Primitive class with check
>>> class Flatten(PrimitiveWithCheck):
... @prim_attr_register
... def __init__(self):
... pass
... def check_shape(self, input_x):
... validator.check_int(len(input_x), 1, Rel.GE, 'input_x rank', self.name)
...
... def check_dtype(self, input_x):
... validator.check_subclass("input_x", input_x, mstype.tensor, self.name)
...
>>> # init a Primitive obj
>>> add = Flatten()
"""
def __init__(self, name):
Primitive.__init__(self, name)
self.set_prim_type(prim_type.py_infer_check)
def __check__(self, *args):
"""Checking the input shape and the input type of ops is valid """
tracks = ['dtype', 'shape']
for track in tracks:
fn = getattr(self, 'check_' + track)
fn(*(x[track] for x in args))
def _clone(self):
"""
Deeply clones the primitive object.
Calls the __init__() method with the same arguments. This method is called in parser if the
flag self.__setattr_flag__ is True.
"""
cloned_prim = Primitive._clone(self)
return cloned_prim
[文档] def check_shape(self, *args):
"""
Check shapes of input args.
Note:
The shape of scalar is an empty tuple.
Args:
args (tuple(int)): shapes of input tensors.
Return:
None.
"""
return None
[文档] def check_dtype(self, *args):
"""
Check data types of input args.
Args:
args (:class:`mindspore.dtype`): data type of inputs.
Return:
None.
"""
return None
[文档]class PrimitiveWithInfer(Primitive):
"""
PrimitiveWithInfer is the base class of primitives in python and defines functions for tracking inference
in python.
There are four method can be overridden to define the infer logic of the primitive: __infer__(), infer_shape(),
infer_dtype(), and infer_value(). If __infer__() is defined in primitive, the __infer__() has the highest priority
to be called. If __infer__() is not defined, infer_shape() and infer_dtype() can be defined to describe the infer
logic of the shape and type. The infer_value() is used for constant propagation.
Args:
name (str): Name of the current Primitive.
Supported Platforms:
``Ascend`` ``GPU`` ``CPU``
Examples:
>>> from mindspore.ops import prim_attr_register, PrimitiveWithInfer
>>> # init a Primitive class with infer
>>> class Add(PrimitiveWithInfer):
... @prim_attr_register
... def __init__(self):
... pass
...
... def infer_shape(self, x, y):
... return x # output shape same as first input 'x'
...
... def infer_dtype(self, x, y):
... return x # output type same as first input 'x'
...
>>> # init a Primitive obj
>>> add = Add()
"""
def __init__(self, name):
Primitive.__init__(self, name)
self.set_prim_type(prim_type.py_infer_shape)
def _clone(self):
"""
Deeply clones the primitive object.
Calls the __init__() method with the same arguments. This method is called in parser if the
flag self.__setattr_flag__ is True.
"""
cloned_prim = Primitive._clone(self)
return cloned_prim
def infer_shape(self, *args):
"""
Infer output shape based on input shape.
Note:
The shape of scalar is an empty tuple.
Args:
args (tuple(int)): shapes of input tensors.
Return:
`tuple(int)`, shapes of output tensors.
"""
return None
def infer_dtype(self, *args):
"""
Infer output dtype based on input dtype.
Args:
args (:class:`mindspore.dtype`): data type of inputs.
Return:
:class:`mindspore.dtype`, data type of outputs.
"""
return None
def infer_value(self, *args):
"""
Infer output value based on input value at compile time.
Args:
args (Any): value of inputs.
Return:
Value of outputs. Return `None`, the value can not be inferred at compile time in this case.
"""
return None
def __infer__(self, *args):
"""Infer shape, type, and value at the same time by using dictionary as arguments."""
tracks = ['dtype', 'shape', 'value']
out = {}
for track in tracks:
fn = getattr(self, 'infer_' + track)
# fn may return None
out[track] = fn(*(x[track] for x in args))
# output does not contain dynamic shape, no need to calculate min/max shape
def has_dynamic_shape(shp):
if isinstance(shp, int):
return shp < 0
if isinstance(shp, (list, tuple)):
return any(has_dynamic_shape(e) for e in shp)
return False
# calculate min/max value for output
def get_specified_value(elems, attr):
has_specified_value = False
ret_vals = []
for elem in elems:
if attr in elem:
has_specified_value = True
ret_vals.append(elem[attr])
else:
ret_vals.append(elem['value'])
return has_specified_value, tuple(ret_vals)
has_min_value, min_values = get_specified_value(args, 'min_value')
has_max_value, max_values = get_specified_value(args, 'max_value')
if has_min_value and has_max_value:
if hasattr(self, '_infer_min_value'):
fn_infer_min_value = getattr(self, '_infer_min_value')
out['min_value'] = fn_infer_min_value(*min_values)
if hasattr(self, '_infer_max_value'):
fn_infer_max_value = getattr(self, '_infer_max_value')
out['max_value'] = fn_infer_max_value(*max_values)
has_shape_value, shape_values = get_specified_value(args, 'shape_value')
if has_shape_value and hasattr(self, '_infer_shape_value') and not None in shape_values:
fn_infer_shape_value = getattr(self, '_infer_shape_value')
out['shape_value'] = fn_infer_shape_value(*shape_values)
if not has_dynamic_shape(out['shape']):
return out
return out
[文档]def prim_attr_register(fn):
"""
Primitive attributes register.
Register the decorator of the built-in operator primitive '__init__'.
The function will add all the parameters of '__init__' as operator attributes ,
and init primitive name.
Args:
fn (function): __init__ function of primitive.
Returns:
function, original function.
Examples:
>>> from mindspore.ops import prim_attr_register, PrimitiveWithCheck
>>> class MatMul(PrimitiveWithCheck):
... @prim_attr_register
... def __init__(self, transpose_a=False, transpose_b=False):
... self.init_prim_io_names(inputs=['x1', 'x2'], outputs=['output'])
...
>>> # init a Primitive obj
>>> matmul = MatMul()
"""
@functools.wraps(fn)
def deco(self, *args, **kwargs):
class_name = self.__class__.__name__
if hasattr(self.__class__, "substitute_name"):
class_name = self.__class__.substitute_name
if isinstance(self, PrimitiveWithInfer):
PrimitiveWithInfer.__init__(self, class_name)
elif isinstance(self, PrimitiveWithCheck):
PrimitiveWithCheck.__init__(self, class_name)
else:
Primitive.__init__(self, self.__class__.__name__)
bound_args = inspect.signature(fn).bind(self, *args, **kwargs)
bound_args.apply_defaults()
arguments = bound_args.arguments
del arguments['self']
del self.init_attrs['name']
for name in arguments:
value = arguments[name]
self.add_prim_attr(name, value)
self.init_attrs[name] = value
fn(self, *args, **kwargs)
deco.decorated_func = fn
return deco
[文档]def constexpr(fn=None, get_instance=True, name=None, reuse_result=True, check=True):
"""
Creates a PrimitiveWithInfer operator that can infer the value at compile time. We can use it to define a function
to compute constant value using the constants in the constructor.
Args:
fn (function): A `fn` use as the infer_value of the output operator. Default: None.
get_instance (bool): If true, return the instance of operator,
otherwise return the operator class. Default: True.
name (str): Defines the operator name. If `name` is None, use the function name as op name. Default: None.
reuse_result (bool): If true, the operator will be executed once and reuse the result next time,
otherwise the operator will always be executed. Default: True.
check (bool): If ture, the parameters will be checked
and the warning message will raised if the parameter is not const value. Default: True.
Examples:
>>> from mindspore.ops import constexpr
>>> a = (1, 2)
>>> # make an operator to calculate tuple len
>>> @constexpr
... def tuple_len(x):
... return len(x)
...
>>> print(tuple_len(a))
2
>>> # make an operator class to calculate tuple len
>>> @constexpr(get_instance=False, name="TupleLen")
... def tuple_len_class(x):
... return len(x)
...
>>> print(tuple_len_class()(a))
2
"""
def deco(fn):
"""Decorator for CompileOp."""
class CompileOp(PrimitiveWithInfer):
"""
CompileOp is a temporary operator used to execute the constexpr function.
"""
def __init__(self):
op_name = name if name else fn.__name__
PrimitiveWithInfer.__init__(self, op_name)
self.set_const_prim(True)
self.fn = fn
if not reuse_result:
self.add_prim_attr('forbid_reuse_result', True)
def __infer__(self, *args):
value_args = []
for item in args:
if (item["dtype"] is not None and item["value"] is None and check):
logger.warning("The \"" + self.name + "\" is a constexpr function." \
" The input arguments must be all constant value.")
value_args.append(item["value"])
return {'dtype': None, 'shape': None, 'value': fn(*value_args)}
def __call__(self, *args, **kwargs):
return fn(*args, **kwargs)
if get_instance:
return CompileOp()
return CompileOp
if fn is not None:
return deco(fn)
return deco
@_wrap_func
def _run_op(obj, op_name, args):
"""Single op execution function supported by ge in PyNative mode."""
output = _pynative_executor.real_run_op(obj, op_name, args)
return output