# This is the Python adaptation and derivative work of Myia (https://github.com/mila-iqia/myia/).
#
# 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.
# ============================================================================
"""Providing interface methods."""
from __future__ import absolute_import
import types
import sys
import os
import time
import ast
import inspect
import importlib
from collections import OrderedDict
from functools import wraps
import numpy as np
import mindspore as ms
from mindspore import context
from mindspore import log as logger
from mindspore._extends.remote import kernel_build_server
from .tensor import Tensor as PythonTensor
from .tensor import CSRTensor as PythonCSRTensor
from .tensor import COOTensor as PythonCOOTensor
from .tensor import RowTensor as PythonRowTensor
from .initializer import initializer
from .._c_expression import GraphExecutor_, Tensor, MetaTensor, CSRTensor, RowTensor, COOTensor, PynativeExecutor_
from .._c_expression import verify_inputs_signature, init_exec_dataset, _set_dataset_mode_config, init_pipeline
from ..parallel._tensor import _load_tensor_by_layout
from ..parallel._ps_context import _is_role_pserver, _is_role_sched, _enable_distributed_mindrt
from ..parallel._utils import _get_device_num, _get_global_rank, _need_to_full, _check_full_batch, _to_full_tensor, \
_get_parameter_broadcast, _get_pipeline_stages
from .._checkparam import Validator
from ._utils import is_shape_unknown
from ..common.mutable import mutable
# store ms_function class compiled pipeline cache
ms_compile_cache = set()
# store cell compiled pipeline cache,
cells_compile_cache = {}
BROADCAST_PHASE = "_broadcast_"
_PYNATIVE_PARRALLEL_FUNC_NAME = "after_shard"
def _convert_python_data(data):
"""
Convert C++ data to python.
Args:
data : The data need be convert.
Returns:
data, a data convert C++ to python
"""
if isinstance(data, Tensor) and not isinstance(data, PythonTensor):
return PythonTensor(data, internal=True)
if isinstance(data, CSRTensor) and not isinstance(data, PythonCSRTensor):
return PythonCSRTensor(csr_tensor=data)
if isinstance(data, COOTensor) and not isinstance(data, PythonCOOTensor):
return PythonCOOTensor(coo_tensor=data)
if isinstance(data, RowTensor) and not isinstance(data, PythonRowTensor):
return PythonRowTensor(row_tensor=data)
if isinstance(data, tuple):
return tuple(_convert_python_data(x) for x in data)
if isinstance(data, list):
return list(_convert_python_data(x) for x in data)
if isinstance(data, dict):
return dict((_convert_python_data(key), _convert_python_data(value)) for key, value in data.items())
return data
def _wrap_func(fn):
"""
Wrapper function, convert return data to tensor or tuple of tensor.
Args:
fn (Function): The function need be wrapped.
Returns:
Function, a new function with return suitable format data.
"""
@wraps(fn)
def wrapper(*arg, **kwargs):
results = fn(*arg, **kwargs)
return _convert_python_data(results)
return wrapper
def _check_all_tensor(sequence):
for element in sequence:
if not isinstance(element, Tensor) and not (isinstance(element, tuple) and _check_all_tensor(element)):
return False
return True
sys_path = list(sys.path)
# Get the entry script path.
if sys.argv and sys.argv[0] != '':
entry_script_path = os.path.realpath(sys.argv[0])
entry_script_path_dir = os.path.split(entry_script_path)[0]
if entry_script_path_dir in sys_path:
sys_path.remove(entry_script_path_dir)
def _in_sys_path(file_path):
for path in sys_path:
if file_path.startswith(path):
return True
return False
def __get_compile_cache_dep_files(file_path, compile_cache_dep_files, pkg):
"""Get the dependency files of the network"""
with open(file_path) as fh:
root = ast.parse(fh.read(), file_path)
for node in ast.iter_child_nodes(root):
module_name = ""
if isinstance(node, ast.ImportFrom):
module_name = node.module
if node.level == 1:
module_name = "." + module_name
elif not isinstance(node, ast.Import):
continue
# Do not care the files in mindspore package
if module_name.startswith("mindspore"):
continue
for n in node.names:
if n.name.startswith("mindspore"):
continue
if module_name == "":
whole_module = n.name
else:
whole_module = module_name
if not n.name is None:
whole_module += "." + n.name
try:
module_spec = importlib.util.find_spec(whole_module, pkg)
except (ModuleNotFoundError, ValueError):
whole_module = whole_module[0:whole_module.rfind('.')]
module_spec = importlib.util.find_spec(whole_module, pkg)
if module_spec is None:
continue
module = importlib.util.module_from_spec(module_spec)
if hasattr(module, '__file__'):
dep_file_path = module.__file__
else:
continue
# Exclude the installed modules.
if not _in_sys_path(dep_file_path) and not dep_file_path in compile_cache_dep_files:
logger.debug(f"dependent file path: {dep_file_path}")
compile_cache_dep_files.append(dep_file_path)
__get_compile_cache_dep_files(dep_file_path, compile_cache_dep_files, module.__package__)
def _get_compile_cache_dep_files():
"""Get the dependency files of the network"""
if entry_script_path is None:
logger.warning("Can not get the entry script file path.")
return []
compile_cache_dep_files = []
logger.debug(f"entry script file path: {entry_script_path}")
compile_cache_dep_files.append(entry_script_path)
__get_compile_cache_dep_files(entry_script_path, compile_cache_dep_files, None)
return compile_cache_dep_files
def _restore_mutable_attr(args_list, compile_args):
"""Restore the mutable attr for every arg."""
new_compile_args = ()
for idx in range(len(args_list)):
# Currently the CSRTensor and COOTensor still keep the mutable attr and we don't support make them mutable.
if isinstance(args_list[idx], (CSRTensor, COOTensor)):
new_compile_args += (compile_args[idx],)
continue
if hasattr(args_list[idx], "__ms_mutable__") and getattr(args_list[idx], "__ms_mutable__"):
new_compile_args += (mutable(compile_args[idx]),)
else:
new_compile_args += (compile_args[idx],)
return new_compile_args
class _MindsporeFunctionExecutor:
"""
Represents a function compiled by graph compiler.
_MindsporeFunctionExecutor will compile the original function for every combination
of argument types and shapes it is given (as well as their values, optionally).
Args:
fn (Function): The root function to compile.
input_signature (Function): User defines signature to verify input.
ms_create_time(TimeStamp): The time ms_function created
obj (Object): If function is a method, obj is the owner of function,
else, obj is none.
Returns:
The result of pipeline running in graph mode.
"""
def __init__(self, fn, ms_create_time, input_signature=None, obj=None, jit_config=None):
init_pipeline()
if not isinstance(fn, (types.FunctionType, types.MethodType)):
raise RuntimeError('fn {} is not function or method'.format(fn))
self.fn = fn
self.input_signature = input_signature
self.obj = None
if obj and hasattr(obj, fn.__name__):
self.obj = obj
self.shard_parent_obj = obj
self.enable_tuple_broaden = False
self._graph_executor = GraphExecutor_.get_instance()
self._create_time = ms_create_time
self.jit_config_dict = jit_config.jit_config_dict if jit_config else None
def _set_compile_cache_dep_files(self):
# If enable compile cache, get the dependency files list
enable_compile_cache = context.get_context("enable_compile_cache")
if enable_compile_cache is None:
enable_compile_cache = os.getenv('MS_COMPILER_CACHE_ENABLE')
if enable_compile_cache is True or enable_compile_cache == "1":
self._graph_executor.set_compile_cache_dep_files(_get_compile_cache_dep_files())
def _parallel_process_for_ms_function(self, phase):
"""Set parameter and optimizer states data according to sliced shape for shard"""
obj = self.shard_parent_obj if self.obj is None else self.obj
obj.parameter_layout_dict = self._graph_executor.get_parameter_layout(phase)
obj.parallel_parameter_name_list = self._graph_executor.get_parallel_parameter_name_list(phase)
replace = obj.init_parameters_data(auto_parallel_mode=True)
new_param = {x.name: replace[x] for x in replace if id(x) != id(replace[x])}
self._graph_executor.updata_param_node_default_input(phase, new_param)
obj.load_parameter_slice(None)
if _pynative_executor.get_optimizer():
params = obj.trainable_params()
opt_params = _pynative_executor.get_optimizer().trainable_params()
opt_states = []
for opt_param in opt_params:
for param in params:
if opt_param.name.find(param.name) > 0:
opt_states.append(opt_param)
obj.parameter_layout_dict[opt_param.name] = obj.parameter_layout_dict[param.name]
continue
states_tuple = (opt_states[:len(params)], opt_states[len(params):]) if len(opt_states) != len(params) \
else (opt_states[:len(params)],)
for states in states_tuple:
for param, state in zip(params, states):
if param.shape != state.shape:
if state.has_init:
state.set_data(initializer("zeros", param.shape), True)
else:
layout = obj.parameter_layout_dict[param.name]
new_tensor = _load_tensor_by_layout(state.data, layout)
state.set_data(new_tensor, True)
_pynative_executor.get_top_cell().parameter_layout_dict = obj.parameter_layout_dict
def compile(self, args_list, method_name):
"""Returns pipeline for the given args."""
# Check whether hook function registered on Cell object.
if self.obj and hasattr(self.obj, "_hook_fn_registered"):
if self.obj._hook_fn_registered():
logger.warning(f"For 'Cell', it's not support hook function when using ms_function. If you want to "
f"use hook function, please use context.set_context to set pynative mode and remove "
f"`ms_function`.")
# Chose dynamic shape tensors or actual input tensors as compile args.
compile_args = self._generate_compile_args(args_list)
# Restore the mutable attr for every arg.
compile_args = _restore_mutable_attr(args_list, compile_args)
generate_name = self.fn.__module__ + "." + self.fn.__name__ + "." + self.fn.__code__.co_filename + "." + \
str(self.fn.__code__.co_firstlineno) + '.' + str(id(self.fn))
if _pynative_executor.grad_flag():
generate_name = generate_name + ".grad"
if is_pynative_parallel():
generate_name = generate_name[:generate_name.rfind(str(id(self.fn)))] + str(id(self.shard_parent_obj))
self.fn.__parse_method__ = method_name
# Add key with obj
if self.obj is not None:
if self.obj.__module__ != self.fn.__module__:
logger.info(f'`obj` module not equal to `fn` module: {self.obj.__module__}, {self.fn.__module__}')
self.obj.__parse_method__ = method_name
generate_name = generate_name + '.' + str(self.obj.create_time) + '.' + str(id(self.obj))
else:
# Different instance of same class may use same memory(means same obj_id) at diff times.
# To avoid unexpected phase matched, add create_time to generate_name.
generate_name = generate_name + '.' + str(self._create_time)
self.enable_tuple_broaden = False
if hasattr(self.obj, "enable_tuple_broaden"):
self.enable_tuple_broaden = self.obj.enable_tuple_broaden
self._graph_executor.set_enable_tuple_broaden(self.enable_tuple_broaden)
key = self._graph_executor.generate_arguments_key(compile_args, self.enable_tuple_broaden)
phase = generate_name + '.' + str(key)
if phase in ms_compile_cache:
return phase
# If enable compile cache, get the dependency files list and set to graph executor.
self._set_compile_cache_dep_files()
if self.jit_config_dict:
self._graph_executor.set_jit_config(self.jit_config_dict)
if self.obj is None:
is_compile = self._graph_executor.compile(self.fn, compile_args, phase, True)
else:
self._graph_executor.set_weights_values(self.obj.parameters_dict())
is_compile = self._graph_executor.compile(self.obj, compile_args, phase, True)
# init sliced parameter and optimizer state
if is_pynative_parallel() and self.fn.__name__ == _PYNATIVE_PARRALLEL_FUNC_NAME:
self._parallel_process_for_ms_function(phase)
# init the rest optimizer states
if is_pynative_parallel() and _pynative_executor.get_optimizer():
opt_states = _pynative_executor.get_optimizer().trainable_params()
self._optimizer_state_init(opt_states)
if not is_compile:
raise RuntimeError("Executor compile failed.")
ms_compile_cache.add(phase)
return phase
@_wrap_func
def __call__(self, *args):
args_list = args
if self.obj is not None:
args_list = args_list[1:]
if is_pynative_parallel() and not hasattr(self.shard_parent_obj, "keep_input_unchanged") and \
self.fn.__name__ == _PYNATIVE_PARRALLEL_FUNC_NAME:
device_num = context.get_auto_parallel_context('device_num')
new_args_list = ()
for arg in args_list:
if isinstance(arg, PythonTensor):
new_shape = (arg.shape[0] * device_num,) + arg.shape[1:]
new_args_list += (PythonTensor(np.zeros(shape=new_shape), arg.dtype),)
else:
new_args_list += (arg,)
phase = self.compile(new_args_list, self.fn.__name__)
else:
phase = self.compile(args_list, self.fn.__name__)
if context.get_context("precompile_only"):
return None
new_inputs = self._generate_run_args(args_list)
output = self._graph_executor(tuple(new_inputs), phase)
if context.get_context("mode") == context.PYNATIVE_MODE:
_pynative_executor.set_graph_phase(phase)
output = _pynative_executor.grad_ms_function(output, *new_inputs)
return output
@staticmethod
def _optimizer_state_init(opt_states):
"""set data for all optimizer states in case it is executed in graph mode"""
prefix_list = ["moments", "accum", "moment1", "moment2", "lamb_m", "lamb_v", "mean_grad",
"mean_square", "prev"]
for opt_param in opt_states:
prefix = opt_param.name[:opt_param.name.find(".")]
if opt_param.has_init and (prefix in prefix_list or opt_param.name == "global_step"):
opt_param.init_data()
def _generate_compile_args(self, args_list):
"""Chose dynamic shape tensors or actual input tensors as compile args."""
# Case: If the shape of input args is dynamic, get dynamic shape tensor from context and use it to compile.
compile_args = _pynative_executor.get_dynamic_input(args_list)
# Case: The `set_inputs()` of Cell object has been set, using these dynamic shape args as compile args.
if isinstance(self.obj, ms.nn.Cell) and self.obj.get_inputs():
compile_args = self.obj.get_inputs()
for args in compile_args:
Validator.check_isinstance("args set in `set_inputs()` of Cell", args, PythonTensor)
Validator.check_dynamic_shape(compile_args, args_list)
# Case: If dynamic shape tensors have been assigned to `input_signature`, they are preferred as compile args.
if self.input_signature is not None:
if not isinstance(self.input_signature, (tuple, list)):
self.input_signature = (self.input_signature,)
self.input_signature = list(self.input_signature)
dyn_shape = False
for sig_args in self.input_signature:
Validator.check_isinstance("args in `input_signature` of `ms_function`", sig_args, MetaTensor)
if is_shape_unknown(sig_args.shape):
dyn_shape = True
if not dyn_shape:
if not verify_inputs_signature(self.input_signature, args_list):
raise ValueError("The input args is incompatible with the args in `input_signature`!")
else:
# Checkout whether the `sens` has been added to args_list.
if len(self.input_signature) == len(args_list) - 1:
logger.warning(f"The number of actual input args `{len(args_list)}` is one more than the number "
f"of input_signature args `{len(self.input_signature)}`. The last actual args may "
f"be `sens` and added it to compile args.")
self.input_signature.append(args_list[-1])
Validator.check_dynamic_shape(self.input_signature, args_list)
compile_args = tuple(self.input_signature)
_pynative_executor.set_dynamic_input(self.obj, *compile_args)
return compile_args
def _generate_run_args(self, args_list):
"""
Generate input args, which are required for running.
Args:
args_list (Tuple): Actual input args.
Returns:
new_inputs, new input args, which are required for running.
"""
new_inputs = []
for i in args_list:
if isinstance(i, (Tensor, CSRTensor, COOTensor)):
new_inputs.append(i)
elif hasattr(i, "__ms_mutable__") and getattr(i, "__ms_mutable__"):
new_inputs.append(i)
elif context.get_context("grad_for_scalar") and isinstance(i, (int, float)):
new_inputs.append(i)
elif self.enable_tuple_broaden and isinstance(i, tuple) and _check_all_tensor(i):
new_inputs.append(i)
return new_inputs
# The attributes used to identify a given object.
attr_op = {"__str__": lambda x: x.__str__(),
"__hash__": lambda x: str(x.__hash__()),
"__module__": lambda x: x.__module__,
"__name__": lambda x: x.__name__,
"__qualname__": lambda x: x.__qualname__,
"__len__": lambda x: str(x.__len__()),
"__code__": lambda x: x.__code__.co_filename + str(x.__code__.co_firstlineno)
}
def _get_obj_id(input_obj):
"""Get hash id of single object."""
obj_id = ".".join(
(map(lambda x: attr_op.get(x)(input_obj) if hasattr(input_obj, x) and getattr(input_obj, x) else "", attr_op)))
return obj_id + str(id(input_obj))
def _get_ms_function_hash(hash_input):
"""Get hash value of single object or list of objects."""
if isinstance(list, tuple):
return ".".join(map(_get_obj_id, hash_input))
return _get_obj_id(hash_input)
[文档]def ms_function(fn=None, input_signature=None, hash_args=None, jit_config=None):
"""
Create a callable MindSpore graph from a Python function.
This allows the MindSpore runtime to apply optimizations based on graph.
Note:
If `input_signature` is specified, each input of `fn` must be a Tensor. And the input arguments for `fn`
will not accept `**kwargs`.
Args:
fn (Function): The Python function that will be run as a graph. Default: None.
input_signature (Tensor): The Tensor which describes the input arguments. The shape and dtype of the Tensor
will be supplied to this function. If input_signature is specified, each input to `fn` must be a `Tensor`.
And the input parameters of `fn` cannot accept `**kwargs`. The shape and dtype of actual inputs should
keep the same as input_signature. Otherwise, TypeError will be raised. Default: None.
hash_args (Object, List or Tuple of Objects): The local free variables used inside `fn`, like functions or
objects of class defined outside `fn`. Calling `fn` again with change of `hash_args` will trigger
recompilation.
jit_config (JitConfig): Jit config for compile. Default: None.
Returns:
Function, if `fn` is not None, returns a callable function that will execute the compiled function; If `fn` is
None, returns a decorator and when this decorator invokes with a single `fn` argument, the callable function is
equal to the case when `fn` is not None.
Supported Platforms:
``Ascend`` ``GPU`` ``CPU``
Examples:
>>> import numpy as np
>>> from mindspore import Tensor
>>> from mindspore import ms_function
>>> from mindspore import ops
...
>>> x = Tensor(np.ones([1, 1, 3, 3]).astype(np.float32))
>>> y = Tensor(np.ones([1, 1, 3, 3]).astype(np.float32))
...
>>> # create a callable MindSpore graph by calling ms_function
>>> def tensor_add(x, y):
... z = x + y
... return z
...
>>> tensor_add_graph = ms_function(fn=tensor_add)
>>> out = tensor_add_graph(x, y)
...
>>> # create a callable MindSpore graph through decorator @ms_function
>>> @ms_function
... def tensor_add_with_dec(x, y):
... z = x + y
... return z
...
>>> out = tensor_add_with_dec(x, y)
...
>>> # create a callable MindSpore graph through decorator @ms_function with input_signature parameter
>>> @ms_function(input_signature=(Tensor(np.ones([1, 1, 3, 3]).astype(np.float32)),
... Tensor(np.ones([1, 1, 3, 3]).astype(np.float32))))
... def tensor_add_with_sig(x, y):
... z = x + y
... return z
...
>>> out = tensor_add_with_sig(x, y)
...
... # Set hash_args as fn, otherwise cache of compiled `closure_fn` will not be reused.
... # While fn differs during calling again, recompilation will be triggered.
>>> def func(x):
... return ops.exp(x)
...
>>> def closure_fn(x, fn):
... @ms_function(hash_args=fn)
... def inner_fn(a):
... return fn(a)
... return inner_fn(x)
...
>>> inputs = Tensor(np.ones([10, 10, 10]).astype(np.float32))
>>> for i in range(10):
... closure_fn(inputs, func)
"""
def wrap_mindspore(func):
if hash_args:
hash_obj = _get_ms_function_hash(hash_args)
else:
hash_obj = int(time.time() * 1e9)
@wraps(func)
def staging_specialize(*args):
process_obj = None
if args and not isinstance(args[0], PythonTensor) and hasattr(args[0], func.__name__):
process_obj = args[0]
# only the function or cell instance wrapped by shard will fall into this branch
if is_pynative_parallel() and func.__name__ == _PYNATIVE_PARRALLEL_FUNC_NAME:
process_obj = args[0]
args = args[1:]
out = _MindsporeFunctionExecutor(func, hash_obj, input_signature, process_obj, jit_config)(*args)
return out
return staging_specialize
if fn is not None:
return wrap_mindspore(fn)
return wrap_mindspore
[文档]def ms_class(cls):
"""
Class decorator for user-defined classes.
This allows MindSpore to identify user-defined classes and thus obtain their attributes and methods.
Args:
cls (Class): User-defined class.
Returns:
Class with __ms_class__ attribute.
Raises:
TypeError: If ms_class is used for non-class types or nn.Cell.
AttributeError: If the private attributes or magic methods of the class decorated by ms_class is called.
Supported Platforms:
``Ascend`` ``GPU`` ``CPU``
Examples:
>>> import mindspore.nn as nn
>>> from mindspore import ms_class
...
>>> @ms_class
... class UserDefinedNet:
... def __init__(self):
... self.value = 10
...
... def func(self, x):
... return 2 * x
...
>>> class Net(nn.Cell):
... def __init__(self):
... super(Net, self).__init__()
... self.net = UserDefinedNet()
...
... def construct(self, x):
... out = self.net.value + self.net.func(x)
... return out
...
>>> net = Net()
>>> out = net(5)
>>> print(out)
20
"""
# Check if cls is of type class.
if not inspect.isclass(cls):
raise TypeError(f'Decorator ms_class can only be used for class type, but got {cls}.')
logger.info(f'Found ms_class: {cls}.')
setattr(cls, '__ms_class__', True)
return cls
def _function_forbid_reuse(func):
if not inspect.isfunction(func):
raise TypeError(f'Decorator _function_forbid_reuse can only be used for function type, but got {func}.')
setattr(func, '__function_forbid_reuse__', True)
return func
def is_pynative_parallel():
run_mode = context.get_context('mode')
parallel_mode = context.get_auto_parallel_context('parallel_mode')
return run_mode == context.PYNATIVE_MODE and parallel_mode in (
context.ParallelMode.SEMI_AUTO_PARALLEL, context.ParallelMode.AUTO_PARALLEL)
def _get_auto_split_param_names(parameter_layout_dict):
auto_split_param_names = []
for key, value in parameter_layout_dict.items():
for dim in value[1]:
if dim != -1:
auto_split_param_names.append(key)
break
return auto_split_param_names
def _build_broadcast_graph(broadcast_params_dict, broadcast_phase):
"""Build broadcast graph."""
from mindspore.nn.wrap.cell_wrapper import _BroadCastCell
if not broadcast_params_dict:
broadcast_params_dict = {}
broadcast_params = []
for param in broadcast_params_dict.values():
broadcast_params.append(Tensor(param.asnumpy()))
_broadcast_net = _BroadCastCell(broadcast_params)
_broadcast_net.phase = broadcast_phase
broadcasted_params = _broadcast_net()
for param_name, param in zip(broadcast_params_dict.keys(), broadcasted_params):
broadcast_params_dict[param_name].set_data(param)
def _parameter_broadcast(obj, auto_parallel_mode):
"""Parameter broadcast."""
auto_split_param_names = []
if auto_parallel_mode:
auto_split_param_names = _get_auto_split_param_names(obj.parameter_layout_dict)
broadcast_params_dict = obj.parameters_broadcast_dict()
if auto_split_param_names and broadcast_params_dict:
broadcast_params_dict = OrderedDict()
for param_name, param in obj.parameters_broadcast_dict().items():
if param_name not in auto_split_param_names:
broadcast_params_dict[param_name] = param
broadcast_phase = "_broadcast_subgraph"
_build_broadcast_graph(broadcast_params_dict, broadcast_phase)
class _PynativeExecutor:
"""
A pynative executor used to compile/manage/run single op.
The main functions include: construct op graph, compile op graph, auto grad and run op graph.
Args:
obj (Object): The python network that will be run in pynative mode.
args (Tuple(Tensor...)): The inputs of network in tuple form.
Returns:
gradients (Tuple(Tensor...)): The gradients of network parameters and inputs.
Supported Platforms:
``Ascend`` ``GPU`` ``CPU``
"""
def __init__(self):
self._executor = PynativeExecutor_.get_instance()
self._executor.set_py_exe_path(sys.executable)
self._executor.set_kernel_build_server_dir(os.path.split(kernel_build_server.__file__)[0] + os.sep)
self._optimizer = None
self._top_cell = None
@staticmethod
def parameter_broadcast(obj, phase, auto_parallel_mode):
"""
Run broadcast for parameter.
Args:
obj (Cell): The cell instance.
phase (str): The phase of cell instance.
auto_parallel_mode (bool): The flag of running auto parallel.
Return:
None.
"""
if BROADCAST_PHASE not in phase and _get_parameter_broadcast():
_parameter_broadcast(obj, auto_parallel_mode)
def new_graph(self, obj, *args, **kwargs):
"""
Initialize resources for building forward and backward graph.
Args:
obj (Function/Cell): The function or cell instance.
args (tuple): Function or cell input arguments.
kwargs (dict): keyword arguments.
Return:
None.
"""
self._executor.new_graph(obj, *args, *(kwargs.values()))
def end_graph(self, obj, output, *args, **kwargs):
"""
Clean resources after building forward and backward graph.
Args:
obj (Function/Cell): The function or cell instance.
output (Tensor/tuple/list): Function or cell output object.
args (tuple): Function or cell input arguments.
kwargs (dict): keyword arguments.
Return:
None.
"""
self._executor.end_graph(obj, output, *args, *(kwargs.values()))
def check_graph(self, obj, *args, **kwargs):
"""
Determines the order of the function or cell.
Args:
obj (Function/Cell): The function or cell instance.
args (tuple): Function or cell input arguments.
kwargs (dict): keyword arguments.
Return:
bool, specifies whether the running function or cell is high-order.
"""
return self._executor.check_graph(obj, *args, *(kwargs.values()))
def check_run(self, grad, obj, *args, **kwargs):
"""
Whether the forward graph need to construct.
Args:
grad (GradOperation): The gradoperation object.
obj (Function/Cell): The function or cell instance.
args (tuple): Function or cell input arguments.
kwargs (dict): keyword arguments.
Return:
bool, specifies whether the forward graph need to construct.
"""
return self._executor.check_run(grad, obj, *args, *(kwargs.values()))
def set_grad_position(self, grad, grad_position):
"""
Set position of grad.
Args:
grad (GradOperation): The gradoperation object.
grad_position (Union(int, tuple[int])): If int, get the gradient with respect to single input.
If tuple, get the gradients with respect to selected inputs. 'grad_position' begins with 0. Default: 0.
Return:
None.
"""
return self._executor.set_grad_position(grad, grad_position)
def set_weights_id(self, grad, weights_id):
"""
Set weights id of grad.
Args:
grad (GradOperation): The gradoperation object.
weights_id (str): The id of weights that consists of names of all Parameters.
Return:
None.
"""
return self._executor.set_weights_id(grad, weights_id)
def grad(self, grad, obj, weights, grad_position, *args, **kwargs):
"""
Get grad graph.
Args:
grad (GradOperation): The gradoperation object.
obj (Function/Cell): The function or cell instance.
weights (ParameterTuple): The weights of cell instance.
grad_position (Union(int, tuple[int])): If int, get the gradient with respect to single input.
If tuple, get the gradients with respect to selected inputs. 'grad_position' begins with 0. Default: 0.
args (tuple): Function or cell input arguments.
kwargs (dict): keyword arguments.
Return:
None.
"""
self._executor.grad_net(grad, obj, weights, grad_position, *args, *(kwargs.values()))
def del_cell(self, obj):
"""
Clean resource for cell.
Args:
obj (Function/Cell): The function or cell instance.
Return:
None.
"""
self._executor.clear_cell(obj)
def clear_res(self):
"""
Clean resource for _PynativeExecutor.
Return:
None.
"""
return self._executor.clear_res()
def clear_grad(self, obj, *args, **kwargs):
"""
Clean resource after building grad graph.
Args:
obj (Function/Cell): The function or cell instance.
args (tuple): Function or cell input arguments.
kwargs (dict): keyword arguments.
Return:
None.
"""
self._executor.clear_grad(obj, *args, *(kwargs.values()))
def sync(self):
"""
SyncStream.
Return:
None.
"""
self._executor.sync()
def set_lazy_build(self, enable):
"""
The switch of lazy build.
Args:
enable (bool): Specifies whether the lazy build is enable.
Return:
None.
"""
self._executor.set_lazy_build(enable)
def grad_ms_function(self, output, *args):
"""
Building grad graph decorated by ms_function.
Args:
output (tuple): The function or cell decorated by ms_function output object.
args (tuple): Function or cell decorated by ms_function input arguments.
Return:
None.
"""
return self._executor.grad_ms_function(output, *args)
def set_graph_phase(self, phase):
"""
Set the phase of cell/function instance.
Args:
phase (str): The phase of cell/function instance.
Return:
None.
"""
self._executor.set_graph_phase(phase)
def grad_flag(self):
"""
The flag of building grad graph.
Return:
bool, whether building grad graph.
"""
return self._executor.grad_flag()
def set_grad_flag(self, flag):
"""
Set the flag of building grad graph.
Args:
flag (bool): Specifying whether building grad graph.
Return:
None.
"""
self._executor.set_grad_flag(flag)
def set_dynamic_input(self, obj, *args):
"""
Set dynamic shape tensor of input arguments.
Args:
obj (Function/Cell): The function or cell instance.
args (tuple): Function or cell dynamic input arguments.
Return:
None.
"""
self._executor.set_dynamic_input(obj, *args)
def get_dynamic_input(self, *actual_args):
"""
Get dynamic shape arguments according to actual input arguments.
Args:
actual_args(tuple): Actual input arguments of Function or Cell.
Return:
dynamic_shape_args(tuple): Dynamic shape arguments of Function or Cell.
"""
return self._executor.get_dynamic_input(*actual_args)
def is_first_cell(self):
"""
The flag of first cell instance.
Return:
bool, specifies whether is the first cell.
"""
return self._executor.is_first_cell()
def set_hook_changed(self, cell):
"""
The flag of registering or removing a hook function on Cell instance.
Args:
cell (Cell): The cell instance.
Return:
None.
"""
self._executor.set_hook_changed(cell)
def get_optimizer(self):
"""
Get the optimizer.
Return:
The optimizer.
"""
return self._optimizer
def get_top_cell(self):
"""
Get the top cell object.
Return:
The top cell object.
"""
return self._top_cell
def __call__(self, sens_param, obj, *args, **kwargs):
"""
PyNative executor run grad graph.
Args:
obj (Function/Cell): The function or cell instance.
args (tuple): Function or cell input arguments.
kwargs (dict): keyword arguments.
Return:
The return object after running grad graph.
"""
args = args + tuple(kwargs.values())
return self._executor(sens_param, obj, args)
class _CellGraphExecutor:
"""
An executor used to compile/manage/run graph for a Cell.
Including data_graph, train_graph, eval_graph and predict graph.
Args:
obj (Function/Cell): The function or cell instance need compile.
args (tuple): Function or cell input arguments.
Returns:
Graph, return the result of pipeline running.
"""
def __init__(self):
# create needed graph by lazy mode
self.is_init = False
self.enable_tuple_broaden = False
self._graph_executor = GraphExecutor_.get_instance()
self._graph_executor.set_py_exe_path(sys.executable)
self._graph_executor.set_kernel_build_server_dir(os.path.split(kernel_build_server.__file__)[0] + os.sep)
def init_dataset(self, queue_name, dataset_size, batch_size, dataset_types, dataset_shapes,
input_indexs, phase='dataset', need_run=True):
"""
Initialization interface for calling data subgraph.
Args:
queue_name (str): The name of tdt queue on the device.
dataset_size (int): The size of dataset.
batch_size (int): The size of batch.
dataset_types (list): The output types of element in dataset.
dataset_shapes (list): The output shapes of element in dataset.
input_indexs (list): The index of data with net.
phase (str): The name of phase, e.g., train_dataset/eval_dataset. Default: 'dataset'.
Returns:
bool, specifies whether the data subgraph was initialized successfully.
"""
if not init_exec_dataset(queue_name=queue_name,
size=dataset_size,
batch_size=batch_size,
types=dataset_types,
shapes=dataset_shapes,
input_indexs=input_indexs,
phase=phase,
need_run=need_run):
raise RuntimeError("Failure to init and dataset subgraph!")
self._graph_executor.set_queue_name(queue_name)
return True
def set_queue_name(self, queue_name):
"""
while a mode use shared dataset with others, need set queue_name which saved in data_set
:param queue_name:
:return:
"""
self._graph_executor.set_queue_name(queue_name)
def _set_dataset_mode(self, args_list):
"""set dataset mode."""
# decide whether to sink based on whether the inputs is virtual or args_list is ()
if (args_list and isinstance(args_list[0], Tensor) and args_list[0].virtual_flag) or \
(args_list is not None and args_list == ()):
_set_dataset_mode_config('sink')
else:
_set_dataset_mode_config('normal')
def _build_data_graph(self, obj, phase):
self._graph_executor.build_data_graph(obj.parameters_dict(), phase, obj.parameters_broadcast_dict())
@staticmethod
def _use_vm_mode():
enable_ge = context.get_context("enable_ge")
enable_debug_runtime = context.get_context("enable_debug_runtime")
exe_mode = context.get_context("mode") == context.PYNATIVE_MODE
return not enable_ge or (enable_debug_runtime and exe_mode)
def _set_compile_cache_dep_files(self, phase):
# If enable compile cache, get the dependency files list
enable_compile_cache = context.get_context("enable_compile_cache")
if enable_compile_cache is None:
enable_compile_cache = os.getenv('MS_COMPILER_CACHE_ENABLE')
if "train" in phase and (enable_compile_cache is True or enable_compile_cache == "1"):
self._graph_executor.set_compile_cache_dep_files(_get_compile_cache_dep_files())
def compile(self, obj, *args, phase='predict', do_convert=True, auto_parallel_mode=False, jit_config_dict=None):
"""
Compiles graph.
Args:
obj (Function/Cell): The function or cell instance need compile.
args (tuple): Function or cell input arguments.
phase (str): The name of compile phase. Default: 'predict'.
do_convert (bool): When set to True, convert ME graph to GE graph after compiling graph.
auto_parallel_mode: When set to True, use auto parallel mode to compile graph.
jit_config_dict (dict): Jit config for compile. Default: None.
Return:
Str, the full phase of the cell.
Bool, if the graph has been compiled before, return False, else return True.
"""
obj.__parse_method__ = 'construct'
if not hasattr(obj, obj.__parse_method__):
raise AttributeError(
'The class {} dose not have method {}'.format(obj.__class__.__name__, obj.__parse_method__))
args_list = args
self.enable_tuple_broaden = False
if hasattr(obj, "enable_tuple_broaden"):
self.enable_tuple_broaden = obj.enable_tuple_broaden
self._graph_executor.set_enable_tuple_broaden(self.enable_tuple_broaden)
key = self._graph_executor.generate_arguments_key(args_list, self.enable_tuple_broaden)
obj.arguments_key = str(key)
phase = phase + '.' + str(obj.create_time) + '.' + str(id(obj)) + '.' + obj.arguments_key
if phase in obj.compile_cache and self.has_compiled(phase):
logger.debug("%r graph has existed.", phase)
return phase, False
obj.check_names()
_check_full_batch()
self._set_dataset_mode(args_list)
self._set_compile_cache_dep_files(phase)
is_sink_mode = args and isinstance(args[0], Tensor) and args[0].virtual_flag
if auto_parallel_mode and _need_to_full() and not is_sink_mode and obj.auto_parallel_compile_and_run():
args_list = _to_full_tensor(args, _get_device_num(), _get_global_rank())
enable_ge = context.get_context("enable_ge")
self._graph_executor.set_weights_values(obj.parameters_dict())
if jit_config_dict:
self._graph_executor.set_jit_config(jit_config_dict)
result = self._graph_executor.compile(obj, args_list, phase, self._use_vm_mode())
obj.compile_cache.add(phase)
if not result:
raise RuntimeError("Executor compile failed.")
graph = self._graph_executor.get_func_graph(phase)
if graph is None:
raise RuntimeError("Compile graph failed for phase {}.".format(phase))
self._auto_parallel_process(obj, phase, is_sink_mode, auto_parallel_mode, *args)
if not do_convert:
return phase, True
# the following GE init process is not needed when use vm or ms backend
if enable_ge:
pass
elif "export" in phase:
self._build_data_graph(obj, phase)
elif BROADCAST_PHASE not in phase and _get_parameter_broadcast():
_parameter_broadcast(obj, auto_parallel_mode)
return phase, True
def _auto_parallel_process(self, obj, phase, is_sink_mode, auto_parallel_mode, *args):
"""compile graph in auto parallel mode."""
if not auto_parallel_mode:
replace = obj.init_parameters_data(auto_parallel_mode=auto_parallel_mode)
self._update_param_node_default_input(phase, replace)
return
obj.parameter_layout_dict = self._graph_executor.get_parameter_layout(phase)
obj.parallel_parameter_name_list = self._graph_executor.get_parallel_parameter_name_list(phase)
replace = obj.init_parameters_data(auto_parallel_mode=True)
if _get_pipeline_stages() > 1 and (not hasattr(obj, "is_first_iteration") or not obj.is_first_iteration):
obj.remove_redundant_parameters()
if not context.get_context("enable_debug_runtime") or context.get_context("enable_ge"):
obj.load_parameter_slice(None)
self._update_param_node_default_input(phase, replace)
# set parallel inputs in sink mode
if is_sink_mode:
obj.set_parallel_input_with_inputs(*args)
def _update_param_node_default_input(self, phase, replace):
new_param = {x.name: replace[x] for x in replace if id(x) != id(replace[x])}
return self._graph_executor.updata_param_node_default_input(phase, new_param)
def _get_shard_strategy(self, obj):
real_phase = obj.phase + '.' + str(obj.create_time) + '.' + str(id(obj)) + '.' + obj.arguments_key
return self._graph_executor.get_strategy(real_phase)
def _get_num_parallel_ops(self, obj):
real_phase = obj.phase + '.' + str(obj.create_time) + '.' + str(id(obj)) + '.' + obj.arguments_key
return self._graph_executor.get_num_parallel_ops(real_phase)
def _get_allreduce_fusion(self, obj):
real_phase = obj.phase + '.' + str(obj.create_time) + '.' + str(id(obj)) + '.' + obj.arguments_key
return self._graph_executor.get_allreduce_fusion(real_phase)
def __call__(self, obj, *args, phase='predict'):
if context.get_context("precompile_only") or\
(_is_role_pserver() and not _enable_distributed_mindrt()) or _is_role_sched():
return None
return self.run(obj, *args, phase=phase)
def has_compiled(self, phase='predict'):
"""
Specify whether have been compiled.
Args:
phase (str): The phase name. Default: 'predict'.
Returns:
bool, specifies whether the specific graph has been compiled.
"""
return self._graph_executor.has_compiled(phase)
@_wrap_func
def _exec_pip(self, obj, *args, phase=''):
"""Execute the generated pipeline."""
fn = obj.construct
obj.__parse_method__ = fn.__name__
return self._graph_executor(args, phase)
def run(self, obj, *args, phase='predict'):
"""
Run the specific graph.
Args:
phase (str): The phase name. Default: 'predict'.
Returns:
Tensor/Tuple, return execute result.
"""
if phase == 'save':
exe_phase = phase + '.' + str(obj.create_time) + '.' + str(id(obj)) + '.' + obj.arguments_key
return self._graph_executor((), exe_phase)
phase_real = phase + '.' + str(obj.create_time) + '.' + str(id(obj)) + '.' + obj.arguments_key
if self.has_compiled(phase_real):
return self._exec_pip(obj, *args, phase=phase_real)
raise KeyError('{} graph is not exist.'.format(phase_real))
def del_net_res(self, net_id):
self._graph_executor.del_net_res(net_id)
def _get_func_graph_proto(self, obj, exec_id, ir_type="onnx_ir", use_prefix=False):
"""Get graph proto from pipeline."""
if use_prefix:
exec_id = exec_id + '.' + obj.arguments_key
if self._graph_executor.has_compiled(exec_id) is False:
return None
return self._graph_executor.get_func_graph_proto(exec_id, ir_type)
def get_optimize_graph_proto(self, obj):
"""Return optimize graph binary proto."""
exec_id = obj.phase + "." + str(obj.create_time) + '.' + str(id(obj)) + '.' + obj.arguments_key
if self._graph_executor.has_compiled(exec_id) is False:
return None
graph_proto = self._graph_executor.get_optimize_graph_proto(exec_id)
if isinstance(graph_proto, str) and graph_proto == "":
logger.warning("Can not get optimize graph proto. Instead, try to find function graph.")
graph_proto = obj.get_func_graph_proto()
return graph_proto
def export(self, file_name, graph_id, enc_key=None, encrypt_func=None):
"""
Export graph.
Args:
file_name (str): File name of model to export
graph_id (str): id of graph to be exported
"""
from .._c_expression import export_graph
export_graph(file_name, 'AIR', graph_id, encrypt_func, enc_key)
def fetch_info_for_quant_export(self, exec_id):
"""Get graph proto from pipeline."""
if self._graph_executor.has_compiled(exec_id) is False:
return None
return self._graph_executor.fetch_info_for_quant_export(exec_id)
[文档]def ms_memory_recycle():
"""
Recycle memory used by MindSpore.
When train multi Neural network models in one process, memory used by MindSpore is very large,
this is because MindSpore cached runtime memory for every model.
To recycle these cached memory, users can call this function after training of one model.
"""
if ms_compile_cache:
_cell_graph_executor.del_net_res(ms_compile_cache)
ms_compile_cache.clear()
for cell_cache in cells_compile_cache.values():
if cell_cache:
_cell_graph_executor.del_net_res(cell_cache)
cell_cache.clear()
_ms_memory_recycle()
_cell_graph_executor = _CellGraphExecutor()
_pynative_executor = _PynativeExecutor()
__all__ = ['ms_function', 'ms_memory_recycle', 'ms_class']