# Copyright 2022 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.
# ============================================================================
"""Rewrite module api: SymbolTree."""
from typing import Optional, Union, List
import mindspore as ms
from mindspore.nn import Cell
from mindspore import _checkparam as Validator
from .node import Node
from ..symbol_tree import Position, SymbolTreeBuilder, SymbolTree as SymbolTreeImpl
ParamTypes = (int, str, float, bool, Node)
MsDtypes = (ms.float16, ms.float32, ms.float64)
[文档]class SymbolTree:
"""
SymbolTree stores information about a network, including statements of the network's forward
computation process and the topological relationship between statement input and output.
The statements in the network are saved in the SymbolTree in the form of nodes, and by processing
the nodes in the SymbolTree, you can delete the network code, insert and replace it, and get the
modified network code and network instances.
Args:
handler (SymbolTreeImpl): SymbolTree internal implementation instance. It is recommended to call the `create`
method in SymbolTree to create a SymbolTree, rather than calling SymbolTree's constructor directly.
Don't care what `SymbolTreeImpl` is, just treat it as a handle.
"""
def __init__(self, handler: SymbolTreeImpl):
Validator.check_value_type("handler", handler, [SymbolTreeImpl], "SymbolTree")
self._symbol_tree: SymbolTreeImpl = handler
[文档] @classmethod
def create(cls, network):
"""
Create a SymbolTree object by passing in the network instance `network`.
This interface parses the `network` instance, expands each source
code statement of the forward computation process, and parses it into nodes,
which is stored in the SymbolTree. The specific process is as follows:
1. Obtain the source code of the network instance.
2. Perform AST parsing on the network and obtain the AST nodes (abstract syntax trees) of each
statement in the network.
3. Expand complex statements in the network forward evaluation process into multiple simple statements.
4. Create a SymbolTree object. Each SymbolTree corresponds to one network instance.
5. Use the rewrite node to store each statement of the network forward computation process. The node records
the input, output, and other information of the statement.
6. Save the rewrite node to the SymbolTree, and update and maintain the topological connection between
the nodes.
7. Return the SymbolTree object corresponding to the network instance.
If a user-defined network of type :class:`mindspore.nn.Cell` is called in the forward computation process
of the network, rewrite will generate a node of type `NodeType.Tree` for the corresponding statement. This
type of node stores a new SymbolTree, which parses and maintains the node information of the user-defined
network.
If the following types of statements are called in the forward computation process of the network, rewrite
will parse the internal statements in the statement and generate corresponding nodes:
- :class:`mindspore.nn.SequentialCell`
- Functions(Excludes Python built-in functions and third-party library functions)
- Control flow statements, such as `if` statements
Note:
Because the specific execution branch of control flows are still unknown during the rewrite operation
of the network, no topology information will be established between the nodes inside the control flow
and the nodes outside.
Users cannot obtain nodes inside the control flow when they acquire nodes outside the control flow using
interfaces like :func:`mindspore.rewrite.Node.get_inputs` and :func:`mindspore.rewrite.Node.get_users` .
Users also cannot obtain nodes outside the control flow, if they use these interfaces inside the control
flow.
Therefore, when users modify the network, they need to manually handle the node information inside and
outside the control flow.
The current rewrite module has the following syntax limitations:
- Only networks of type :class:`mindspore.nn.Cell` are supported as input to the rewrite module.
- Parsing one-line control flow syntax(e.g. one-line if-else, one-line for loop) is not currently supported.
- Parsing decorator syntax is not currently supported.
- Parsing local classes and embedded classes is not currently supported, that is, the definition
of classes need to be placed on the outermost layer.
- Parsing closure syntax is not currently supported, that is, the definition of out-of-class
functions need to be placed at the outermost layer.
- Parsing lambda expression syntax is not currently supported.
- Parsing global variables is not currently supported, that is, global variables need to be converted to
class variables or local variables before they can be used.
- Parsing methods in the parent classes is not currently supported.
For statements that do not support parsing, rewrite will generate nodes of type `NodeType.Python`
for corresponding statements to ensure that the network after rewrite can run normally.
The `Python` node does not support modifying the input and output of statements, and there may be
a problem between variable names and those generated by the rewrite. In this case, users need to
adjust the variable names manually.
Args:
network (Cell): `network` used to create SymbolTree.
Returns:
Symboltree, a SymbolTree created based on `network`.
Raises:
TypeError: If `network` is not a `Cell` instance.
Examples:
>>> from mindspore.rewrite import SymbolTree
>>> # Define the network structure of LeNet5. Refer to
>>> # https://gitee.com/mindspore/docs/blob/r2.3.q1/docs/mindspore/code/lenet.py
>>> net = LeNet5()
>>> stree = SymbolTree.create(net)
>>> print(type(stree))
<class 'mindspore.rewrite.api.symbol_tree.SymbolTree'>
"""
Validator.check_value_type("network", network, [Cell], "SymbolTree")
return cls(SymbolTreeBuilder(network).build())
@staticmethod
def _check_args_type(args):
for arg in args:
if arg not in MsDtypes and not isinstance(arg, ParamTypes):
raise TypeError(f"For call-function Node, got unsupported arg: {arg}, type: {type(arg)}")
@staticmethod
def _check_kwargs_type(kwargs):
for k, v in kwargs.items():
if not isinstance(k, str):
raise TypeError(f"For call-function Node, key in kwarg must be a str, but got: {type(v)}",)
if v not in MsDtypes and not isinstance(v, ParamTypes):
raise TypeError(f"For call-function Node, got unsupported kwarg value: {v}, type: {type(v)}")
def get_handler(self) -> SymbolTreeImpl:
return self._symbol_tree
[文档] def nodes(self, all_nodes: bool = False):
"""
Get the generator of the node in the current SymbolTree, which is used to iterate
through the nodes in SymbolTree.
Args:
all_nodes (bool): Get all nodes including nodes in CallFunction node, CellContainer node
and sub symbol tree. Default: ``False`` .
Returns:
A generator for nodes in SymbolTree.
Raises:
TypeError: If `all_nodes` is not bool.
Examples:
>>> from mindspore.rewrite import SymbolTree
>>> # Define the network structure of LeNet5. Refer to
>>> # https://gitee.com/mindspore/docs/blob/r2.3.q1/docs/mindspore/code/lenet.py
>>> net = LeNet5()
>>> stree = SymbolTree.create(net)
>>> print([node.get_name() for node in stree.nodes()])
['input_x', 'conv1', 'relu', 'max_pool2d', 'conv2', 'relu_1', 'max_pool2d_1',
'unaryop_not', 'if_node', 'flatten', 'fc1', 'relu_2', 'fc2', 'relu_3', 'fc3', 'return_1']
"""
Validator.check_value_type("all_nodes", all_nodes, [bool], "nodes")
nodes = self._symbol_tree.all_nodes() if all_nodes else self._symbol_tree.nodes()
for node in nodes:
yield Node(node)
[文档] def get_node(self, node_name: str) -> Optional[Node]:
"""
Get the node with the name `node_name` in the SymbolTree.
Args:
node_name (str): The name of node.
Returns:
Node with name of `node_name` . Return ``None`` if there is no node named `node_name` in SymbolTree.
Examples:
>>> from mindspore.rewrite import SymbolTree
>>> # Define the network structure of LeNet5. Refer to
>>> # https://gitee.com/mindspore/docs/blob/r2.3.q1/docs/mindspore/code/lenet.py
>>> net = LeNet5()
>>> stree = SymbolTree.create(net)
>>> node = stree.get_node('conv1')
>>> print(node.get_name())
conv1
"""
Validator.check_value_type("node_name", node_name, [str], "SymbolTree")
node_impl = self._symbol_tree.get_node_from_name(node_name)
if node_impl is None:
return None
return Node(node_impl)
def get_inputs(self) -> List[Node]:
return [Node(node_impl) for node_impl in self._symbol_tree.get_input_nodes()]
[文档] def before(self, node: Union[Node, str]):
"""
Returns a location information before `node`. The return value of this interface is
used as a parameter for the insert operation.
Args:
node (Union[Node, str]): Indicate the position before which node. Can be a node or name of node.
Returns:
A `Position` to indicate where to insert node.
Raises:
TypeError: if `node` is not a Node or str.
Examples:
>>> from mindspore.rewrite import SymbolTree
>>> # Define the network structure of LeNet5. Refer to
>>> # https://gitee.com/mindspore/docs/blob/r2.3.q1/docs/mindspore/code/lenet.py
>>> net = LeNet5()
>>> stree = SymbolTree.create(net)
>>> for node in stree.nodes():
... if node.get_name() == "conv1":
... position = stree.before(node)
"""
Validator.check_value_type("node", node, [Node, str], "SymbolTree")
if isinstance(node, Node):
node = node.get_handler()
return self._symbol_tree.before(node)
[文档] def after(self, node: Union[Node, str]):
"""
Returns a location information after `node`. The return value of this interface is
used as a parameter for the insert operation.
Args:
node (Union[Node, str]): Indicate the position after which node. Can be a node or name of node.
Returns:
A `Position` to indicate where to insert node.
Raises:
TypeError: If `node` is not a Node or str.
Examples:
>>> from mindspore.rewrite import SymbolTree
>>> # Define the network structure of LeNet5. Refer to
>>> # https://gitee.com/mindspore/docs/blob/r2.3.q1/docs/mindspore/code/lenet.py
>>> net = LeNet5()
>>> stree = SymbolTree.create(net)
>>> for node in stree.nodes():
... if node.get_name() == "conv1":
... position = stree.after(node)
"""
Validator.check_value_type("node", node, [Node, str], "SymbolTree")
if isinstance(node, Node):
node = node.get_handler()
return self._symbol_tree.after(node)
[文档] def insert(self, position, node: Node) -> Node:
"""
Insert a `node` into `SymbolTree` at `position`.
`position` is obtained from `before` api or `after` api of `SymbolTree`.
Args:
position (Position): Indicate where to insert `node`.
node (Node): An instance of Node to be inserted.
Returns:
An instance of Node being inserted.
Raises:
ValueError: If `position` is not belong to current `SymbolTree`.
TypeError: If `position` is not a `Position`.
TypeError: If `node` is not a `Node`.
Examples:
>>> from mindspore.rewrite import SymbolTree, ScopedValue
>>> import mindspore.nn as nn
>>> # Define the network structure of LeNet5. Refer to
>>> # https://gitee.com/mindspore/docs/blob/r2.3.q1/docs/mindspore/code/lenet.py
>>> net = LeNet5()
>>> stree = SymbolTree.create(net)
>>> node = stree.get_node("conv1")
>>> position = stree.after(node)
>>> new_node = node.create_call_cell(cell=nn.ReLU(), targets=['x'],
... args=[ScopedValue.create_naming_value('x')], name='new_relu')
>>> stree.insert(position, new_node)
"""
Validator.check_value_type("position", position, [Position], "SymbolTree")
Validator.check_value_type("node", node, [Node], "SymbolTree")
return Node(self._symbol_tree.insert_node(node.get_handler(), position.node, position.before_node))
[文档] def erase(self, node: Union[Node, str]) -> Optional[Node]:
"""
Erase a `node` from rewrite.
Args:
node (Union[Node, str]): A `Node` to be erased. Can be a node or name of node.
Returns:
An instance of `Node` being erased if node is in `SymbolTree` else None.
Raises:
TypeError: The type of `node` is not Node or str.
Examples:
>>> from mindspore.rewrite import SymbolTree
>>> # Define the network structure of LeNet5. Refer to
>>> # https://gitee.com/mindspore/docs/blob/r2.3.q1/docs/mindspore/code/lenet.py
>>> net = LeNet5()
>>> stree = SymbolTree.create(net)
>>> node = stree.get_node("conv1")
>>> stree.erase(node)
"""
Validator.check_value_type("node", node, [Node, str], "SymbolTree")
if isinstance(node, Node):
node = node.get_handler()
return Node(self._symbol_tree.erase_node(node))
[文档] def replace(self, old_node: Node, new_nodes: List[Node]) -> Node:
"""
Replace the `old_node` with nodes in the `new_nodes` list.
Nodes in `new_nodes` will be inserted into SymbolTree sequentially, and then `old_node` will be deleted.
Note:
- Replace support one-to-one replacement or one-to-multi replacement. If you need multi-to-multi
replacement, please refer to `PatternEngine`.
- Caller should maintain the topological relationship between each node in the `new_nodes` , as well as
the topological relationship between nodes in the `new_nodes` and nodes in the original tree.
Args:
old_node (Node): Node to be replaced.
new_nodes (List[Node]): Nodes of the node_tree to replace in.
Returns:
An instance of Node represents root of node_tree been replaced in.
Raises:
TypeError: If `old_node` is not a `Node`.
TypeError: If `new_nodes` is not a `list` or node in `new_nodes` is not a `Node`.
Examples:
>>> from mindspore.rewrite import SymbolTree, ScopedValue
>>> import mindspore.nn as nn
>>> # Define the network structure of LeNet5. Refer to
>>> # https://gitee.com/mindspore/docs/blob/r2.3.q1/docs/mindspore/code/lenet.py
>>> net = LeNet5()
>>> stree = SymbolTree.create(net)
>>> node = stree.get_node("conv1")
>>> new_node = node.create_call_cell(cell=nn.ReLU(), targets=['x'],
... args=[ScopedValue.create_naming_value('x')], name='new_relu')
>>> stree.replace(node, [new_node])
"""
Validator.check_value_type("old_node", old_node, [Node], "SymbolTree")
Validator.check_element_type_of_iterable("new_nodes", new_nodes, [Node], "SymbolTree")
nodes_impl = [node.get_handler() for node in new_nodes]
return Node(self._symbol_tree.replace(old_node.get_handler(), nodes_impl))
def set_output(self, index: int, return_value: str) -> Node:
Validator.check_value_type("index", index, [int], "SymbolTree")
Validator.check_value_type("return_value", return_value, [str], "SymbolTree")
return Node(self._symbol_tree.set_output(return_value, index))
def dump(self):
self._symbol_tree.dump()
[文档] def print_node_tabulate(self, all_nodes: bool = False):
r"""
Print the topology information of nodes in SymbolTree, including node type, node name, node code,
and node input-output relationship.
The information is output to the screen using the print interface, including the following information:
- **node type** (str): The type of node, refer to class:`mindspore.rewrite.NodeType` .
- **name** (str): The name of node.
- **codes** (str): The code statement in the SymbolTree corresponding to the node.
- **arg providers** (Dict[int, Tuple[str, int]]): The format is `{[idx, (n, k)]}` , which means the
`idx` th parameter of the node is provided by the `k` th output of node `n` .
- **target users** (Dict[int, List[Tuple[str, int]]]): The format is '{[idx, [(n, k)]]}' , which means
the `idx` th output of the node is used as the `k` th parameter of node `n` .
Args:
all_nodes (bool): Print information of all nodes, including nodes in CallFunction
node, CellContainer node and sub symbol tree. Default: ``False`` .
Raises:
TypeError: If `all_nodes` is not bool.
Examples:
>>> from mindspore.rewrite import SymbolTree
>>> # Define the network structure of LeNet5. Refer to
>>> # https://gitee.com/mindspore/docs/blob/r2.3.q1/docs/mindspore/code/lenet.py
>>> net = LeNet5()
>>> stree = SymbolTree.create(net)
>>> stree.print_node_tabulate()
"""
Validator.check_value_type("all_nodes", all_nodes, [bool], "print_node_tabulate")
dump_str = self._symbol_tree.get_node_tabulate(all_nodes)
print(dump_str)
[文档] def get_code(self) -> str:
"""
Get source code corresponding to the network information in SymbolTree.
If the network has already been modified, the source code of modified network is returned.
Returns:
A str represents source code of modified network.
Examples:
>>> from mindspore.rewrite import SymbolTree
>>> # Define the network structure of LeNet5. Refer to
>>> # https://gitee.com/mindspore/docs/blob/r2.3.q1/docs/mindspore/code/lenet.py
>>> net = LeNet5()
>>> stree = SymbolTree.create(net)
>>> codes = stree.get_code()
>>> print(codes)
"""
return self._symbol_tree.get_code()
[文档] def get_network(self) -> Cell:
"""
Get the network object generated based on SymbolTree.
The source code is saved to a file in the 'rewritten_network' folder of the current directory.
Note:
- The modification of network by rewrite module is based on the modification of AST tree of
original network instance, and the new network instance will obtain attribute information
from original network instance, so the new network instance and the original network instance
have data association, and the original network should no longer be used.
- Due to the data association between the new network and the original network instance, manually creating
a network instance using the source code file generated by rewrite is not currently supported.
Returns:
A network object generated from SymbolTree.
Examples:
>>> from mindspore.rewrite import SymbolTree
>>> # Define the network structure of LeNet5. Refer to
>>> # https://gitee.com/mindspore/docs/blob/r2.3.q1/docs/mindspore/code/lenet.py
>>> net = LeNet5()
>>> stree = SymbolTree.create(net)
>>> new_net = stree.get_network()
"""
return self._symbol_tree.get_network()
def set_saved_file_name(self, file_name: str):
Validator.check_value_type("file_name", file_name, [str], "Saving network")
self._symbol_tree.set_saved_file_name(file_name)
def get_saved_file_name(self):
return self._symbol_tree.get_saved_file_name()
def save_network_to_file(self):
self._symbol_tree.save_network_to_file()
[文档] def unique_name(self, name: str = "output"):
"""
Based on the given `name` , returns a new name that is unique within the symbol tree.
This interface can be used when a variable name that does not conflict is required.
Args:
name (str, optional): The prefix of the name. Defaults to ``"output"`` .
Returns:
str, A new, unique name within a symbol tree in the format `name_n`, where `n` is a numeric subscript.
If there is no name conflict when entered `name`, there is no numeric subscript.
Raises:
TypeError: The type of `name` is not str.
"""
Validator.check_value_type("name", name, [str], "SymbolTree")
return self._symbol_tree.unique_name(name)
def flatten_nodes(self, node, erase_nodes_after_return: bool = False):
Validator.check_value_type("node", node, [Node], "flatten_nodes")
Validator.check_value_type("erase_nodes_after_return", erase_nodes_after_return, [bool], "flatten_nodes")
return self._symbol_tree.flatten_nodes(node.get_handler(), erase_nodes_after_return)
def flatten_static_if_control_flow(self):
return self._symbol_tree.flatten_static_if_control_flow()
# pylint: disable=missing-docstring
def get_origin_network(self):
return self._symbol_tree.get_origin_network()
# pylint: disable=missing-docstring
def all_nodes(self, subtree_nodes: bool = True):
Validator.check_value_type("subtree_nodes", subtree_nodes, [bool], "all_nodes")
return [Node(n) for n in self._symbol_tree.all_nodes(subtree_nodes)]