Method opset_matching

ppq/IR/search.py:564–619

opset_matching is used for operator set searching, it returns relevant op set instead of specific paths, should be used when results of path_matching overflow. opset_matching 用于图检索，匹配算子集合（无序） 一个最简单的例子： sp_expr = lamdba x: x.type == 'Conv'

(
        self,
        sp_expr: Callable,
        rp_expr: Callable,
        ep_expr: Callable,
        direction: str = 'up'
        )

        return _ret_collection

    def opset_matching(
        self,
        sp_expr: Callable,
        rp_expr: Callable,
        ep_expr: Callable,
        direction: str = 'up'
        ) -> OperationSet:
        """opset_matching is used for operator set searching, it returns
        relevant op set instead of specific paths, should be used when results
        of path_matching overflow.

        opset_matching 用于图检索，匹配算子集合（无序）
            一个最简单的例子：
                sp_expr = lamdba x: x.type == 'Conv'
                rp_expr = lamdba x, y: y.type == 'Relu'
                ep_expr = lamdba x: x.type == 'Conv'
                limitation = None
                direction = 'down'
            该指令检索出从任意 Conv 出发，经过任意多个Relu，到任意 Conv 的所有相关算子
            注意返回结果是无序的，相比于 path matching，opset matching 的性能更高。

            你需要注意到 rp_expr 是一个二元表达式，其中 x 代表起点，y 代表终点

            在极端情况下，path matching 的结果是无法返回的（由于其结果数可能过多）
            此时应当使用 opset matching
        Args:
            sp_expr (Union[Pattern, Callable]):
                start point expression, 用于匹配检索起点的表达式
            rp_expr (Union[Pattern, Callable]):
                relay point expression, 用于匹配检索中继点的表达式
            ep_expr (Union[Pattern, Callable]):
                end point expression, 用于匹配检索终点的表达式
            limitation (Union[Limitation, Callable]):
                用于过滤路径的表达式
            direction (str, optional):
                图检索方向，up, down. Defaults to 'down'.
            greedy (bool, optional):
                是否执行贪心匹配，设置为True则尝试匹配直到最后一个end point.
        """
        ret_collection, candidates = OperationSet(), OperationSet()
        for operation in self.graph.operations.values():
            # filter out operation which can not be a start point.
            if sp_expr(operation): candidates.add(operation)

        for operation in candidates:
            # match patterns, add patterns towards result collection.
            partial_matchings = self._opset_matching(
                start_point=operation, rp_expr=rp_expr,
                ep_expr=ep_expr, direction=direction)

            if len(partial_matchings) > 0:
                ret_collection.update(partial_matchings)

        # clear cache
        self._cache.clear()
        return ret_collection

    def activation_matching(