Update python pattern expression · Pull Request !3715 · MindSpore/mindspore

<!--  Thanks for sending a pull request!  Here are some tips for you:

If this is your first time, please read our contributor guidelines: https://gitee.com/mindspore/mindspore/blob/master/CONTRIBUTING.md
-->

**What type of PR is this?**
> Uncomment only one ` /kind <>` line, hit enter to put that in a new line, and remove leading whitespaces from that line:
> /kind feature

**What does this PR do / why do we need it**:
Updated python pass pattern expression. In the old version, pattern and target are expressed as python functions. These two functions will be parsed and resolved as ANF node which will be used in the pattern matching and node replacement process. However, this method has three major drawbacks:

1) it's hard to express correspondence between pattern and target. Although we can use placeholders in function parameter list to link nodes captured during the matching process and nodes to be used in the target replacement process, but it's rather **implicit** and confusing.
2) it's hard to express new nodes in the target(which is a common usage).  If we want to use new nodes in the target(unencountered nodes in the pattern), we may need to add new patches for parser and resolver, which is quite tedious.
3）similar passes cannot be reduced. Since one pass can only be parsed and resolved to one sub-graph, every and each map between pattern and target sub-graph needs to be written as a separate pass no matter how similar they are. Say we wanna repalce "ReLU" and "ReLU6" to "Softmax", we need two passes in the old pattern expression which is quite ugly and error-prone.

In this PR, we propose a new way to explicitly express pattern and target. Pattern and target are built in a down-to-top manner, the final pattern and target are nested patterns.It involves 6 pattern classes(which are all derived from the same base Pattern class) to express nodes and relationship between nodes:
1)`AnyPattern` :  match with any node
2)`NewTensor` : create a new tensor node for the target
3)`IsPrimTypeOf`: match with primitive of certain type(s)
4)`CallWith`: match with primitive CNode with specified primitive and inputs
5)`IsIn`: express the IsIn relationship for patterns
6)`IsNot`: express the IsNot relationship for patterns (cannot be the root pattern though)
By using this new pattern expression, the drawbacks mentioned above are addressed.

**Which issue(s) this PR fixes**:

Fixes #

**Special notes for your reviewers**:

GVPMindSpore/mindspore

搜索帮助

GVP MindSpore/mindspore