megengine.core.tensor.megbrain_graph.dump_graph

dump_graph(output_vars, *, keep_var_name=1, keep_opr_name=False, keep_param_name=False, keep_opr_priority=False, strip_info_file=None, append_json=False, metadata=None)[源代码]

serialize the computing graph of output_vars and get byte result.

参数

  • output_vars (Union[Dict[str, VarNode], List[VarNode]]) – output variables which are the graph’s end point.

  • keep_var_name (int) – 保留变量名的级别:

  • keep_opr_name (bool) – 是否要保留算子的名字

  • keep_param_name (bool) – 是否要保留参数的名字，为了加载模型后可以简单地对参数做操作

  • keep_opr_priority (bool) – 是否保留算子的优先级设置

  • strip_info_file – 路径地址或文件句柄。如果不为空，则导出的代码条信息会被写入``strip_info_file``中。

  • append_json – 当 strip_info_file 非空时会做检查。如果是真，代码条信息就会被添加到 strip_info_file 的尾部；如果是假，就会覆盖掉 strip_info_file.

注解

The underlying C++ API only accepts a var list. If a dict is given, the vars would be renamed to the given names.

返回类型

Tuple[bytes, CompGraphDumpResult]

返回

dump result as byte string, and an instance of namedtuple CompGraphDumpResult, whose fields are:

  • nr_opr number of operators dumped

  • tot_bytes total bytes for the whole graph

  • tensor_value_bytes bytes consumed for dumping tensor values

  • inputs names of input tensors

  • params list of names of dumped params

  • outputs names of output vars

上一页

megengine.core.tensor.megbrain_graph.modify_opr_algo_strategy_inplace

下一页

megengine.core.tensor.megbrain_graph.load_graph