# 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.
# ============================================================================
"""lazy_inline"""
from __future__ import absolute_import
import inspect
from functools import wraps
from mindspore import log as logger
[文档]def lazy_inline(fn=None, attrs=None, policy=None):
"""
Make the cell to be reusable. The corresponding sub graph will not be inline at first
and will be inline with the policy.
Registering the decorator of the built-in function `__init__` of a cell, the decorator
will add the parameters of `__init__` according to the `attrs` as the attributes of this cell.
For a detailed description of the function, see `Using the lazy_inline decorator
<https://www.mindspore.cn/docs/en/master/model_train/program_form/static_graph_syntax/
static_graph_expert_programming.html#using-lazy-inline-decorator>`_ .
.. warning::
This feature is only supported on Ascend and is not supported on other hardwares.
The construct parameters must be positional or key word arguments and have not default values.
The cell has not switch sub graph.
Args:
fn (function): `__init__` function of a cell.
attrs (Union[list[string], string]): The attributes list to add for the cell.
policy (Union[None, "front"]): The policy of inline. Default is None.
- ``None``: The cell will be compiled to sub graph and will not be inline.
- ``"front"``: The cell will be compiled to sub graph first and will be inline at front end.
Returns:
function, original function.
Supported Platforms:
``Ascend``
Examples:
>>> import numpy as np
>>> from mindspore import Tensor
>>> import mindspore.nn as nn
>>> from mindspore import lazy_inline
>>> from mindspore import context
>>> from mindspore import ops
>>> def conv3x3(in_channels, out_channels, stride=1, padding=1, pad_mode='pad'):
... return nn.Conv2d(in_channels, out_channels,
... kernel_size=3, stride=stride, padding=padding, pad_mode=pad_mode)
...
>>> def conv1x1(in_channels, out_channels, stride=1, padding=0, pad_mode='pad'):
... return nn.Conv2d(in_channels, out_channels,
... kernel_size=1, stride=stride, padding=padding, pad_mode=pad_mode)
...
>>> class Block(nn.Cell):
... expansion = 4
...
... @lazy_inline
... def __init__(self,
... in_channels,
... out_channels,
... stride=1,
... down_sample=False):
... super(Block, self).__init__()
...
... out_chls = out_channels
... self.conv1 = conv1x1(in_channels, out_chls, stride=1, padding=0)
... self.bn1 = nn.BatchNorm2d(out_chls)
...
... self.conv2 = conv3x3(out_chls, out_chls, stride=stride, padding=1)
... self.bn2 = nn.BatchNorm2d(out_chls)
...
... self.conv3 = conv1x1(out_chls, out_channels, stride=1, padding=0)
... self.bn3 = nn.BatchNorm2d(out_channels)
...
... self.relu = nn.ReLU()
... self.downsample = down_sample
...
... self.conv_down_sample = conv1x1(in_channels, out_channels,
... stride=stride, padding=0)
... self.bn_down_sample = nn.BatchNorm2d(out_channels)
... self.add = ops.Add()
...
... def construct(self, x):
... identity = x
...
... out = self.conv1(x)
... out = self.bn1(out)
... out = self.relu(out)
...
... out = self.conv2(out)
... out = self.bn2(out)
... out = self.relu(out)
...
... out = self.conv3(out)
... out = self.bn3(out)
...
... if self.downsample:
... identity = self.conv_down_sample(identity)
... identity = self.bn_down_sample(identity)
...
... out = self.add(out, identity)
... out = self.relu(out)
...
... return out
...
>>> class Net(nn.Cell):
... def __init__(self, block, num_classes=100):
... super(Net, self).__init__()
...
... self.conv1 = nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3, pad_mode='pad')
... self.bn1 = nn.BatchNorm2d(64)
... self.relu = nn.ReLU()
... self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, pad_mode='valid')
...
... self.layer = self.MakeLayer(
... block, 50, in_channels=64, out_channels=2048, stride=2)
... self.avgpool = nn.AvgPool2d(7, 1)
... self.flatten = ops.Flatten()
...
... def MakeLayer(self, block, layer_num, in_channels, out_channels, stride):
... layers = []
... resblk = block(in_channels, out_channels,
... stride=stride, down_sample=True)
... layers.append(resblk)
...
... for _ in range(1, layer_num):
... resblk = block(out_channels, out_channels, stride=1)
... layers.append(resblk)
...
... return nn.SequentialCell(layers)
...
... def construct(self, x):
... x = self.conv1(x)
... x = self.bn1(x)
... x = self.relu(x)
... x = self.maxpool(x)
... x = self.layer(x)
... x = self.avgpool(x)
... x = self.flatten(x)
... return x
...
>>> def test_compile():
... net = Net(Block)
... inp = Tensor(np.ones([1, 3, 224, 224]).astype(np.float32))
... net(inp)
...
>>> context.set_context(mode=context.GRAPH_MODE)
>>> os.environ["MS_DEV_SAVE_GRAPHS"] = "2"
>>> os.environ["MS_DEV_SAVE_GRAPHS_PATH"] = os.path.realpath("./lazy")
...
>>> test_compile()
"""
def lazy_inline_wrap(fn):
if inspect.isclass(fn):
tips = "The lazy_inline should decorate the __init__ function, not the class {}.".format(fn.__name__) \
+ " File: " + inspect.getfile(fn)
raise ValueError(tips)
if fn.__name__ != "__init__":
tips = "The lazy_inline should decorate the __init__ function, not the function: {}.".format(fn.__name__) \
+ " line: " + str(fn.__code__.co_firstlineno) + " in " \
+ fn.__code__.co_filename
raise ValueError(tips)
def check_parameters(self):
if hasattr(fn, "has_tips_"):
return
if hasattr(self, "construct"):
params = inspect.signature(self.construct).parameters
err = False
tips = "The function construct's parameters: "
for name, parm in params.items():
if parm.default != inspect.Parameter.empty:
if err:
tips += " , " + name
else:
err = True
tips += " " + name
if err:
tips += " must be key word or positional arguments and can't have default values." \
+ " line: " + str(self.construct.__code__.co_firstlineno) \
+ " in " + self.construct.__code__.co_filename
logger.info(tips)
fn.has_tips_ = True
else:
tips = "The " + self.__class__.__name__ + " must be a cell and must has a construct function." \
+ " line: " + str(fn.__code__.co_firstlineno) + " in " + fn.__code__.co_filename
logger.warning(tips)
fn.has_tips_ = True
@wraps(fn)
def lazy_inline_deco(self, *args, **kwargs):
check_parameters(self)
new_args = []
if attrs is None:
bound_args = inspect.signature(fn).bind(self, *args, **kwargs)
new_args = bound_args.arguments
del new_args['self']
new_args = new_args.values()
fn(self, *args, **kwargs)
if isinstance(policy, str) and policy == "front":
self.no_inline = False
elif policy is not None:
raise ValueError(f"policy must be None or 'front'")
if attrs is None:
self.cell_init_args = "lazy_inline_" + type(self).__name__ + str(new_args)
return
if isinstance(attrs, list):
for attr in attrs:
if not isinstance(attr, str):
raise ValueError(f"attr must be a string")
if hasattr(self, attr):
new_args.append(getattr(self, attr))
elif isinstance(attrs, str):
if hasattr(self, attrs):
new_args = getattr(self, attrs)
else:
raise ValueError(f"attrs must be list or string")
self.cell_init_args = "lazy_inline_" + type(self).__name__ + str(new_args)
return lazy_inline_deco
if fn is not None:
return lazy_inline_wrap(fn)
return lazy_inline_wrap