fix bug in AddInputOutputOpsPass: check existence of key in HashMap(inferface_lbi2scope_sym_id)

[zzk0]

Bug 复现

1. 训练一个模型。
2. 使用 python/oneflow/serving/saved_model_builder.py 将模型保存下来，并且增加一个自定义 Signature。
3. 使用 python/oneflow/serving/inference_session.py 读取模型，在调用 CurJobBuildAndInferCtx_Complete 的时候报错。

问题定位

单步调试，直到发现是在 AddInputOutputOpsPass 中出了问题。仔细阅读了代码之后，发现：如果一个 Input 并不参与计算但它又存在，也就是说从 output 反向遍历不到的话，这个 Input 是不会加入到一个 HashMap 中的。可是，后面的代码会根据 signature 去 HashMap 找 Input 对应的 key。

代码错误发生在

int64_t scope_sym_id = inferface_lbi2scope_sym_id.at(input_def.lbi());

附：

下面是报错信息和保存的模型。

Traceback (most recent call last):
  File "mlp_load.py", line 20, in <module>
    sess.load_saved_model(saved_model_dir="./models", model_version=2)
  File "/home/percent1/oneflow/build_debug/python_scripts/oneflow/python/serving/inference_session.py", line 399, in load_saved_model
    self.compile(graph_def.op_list)
  File "/home/percent1/oneflow/build_debug/python_scripts/oneflow/python/serving/inference_session.py", line 316, in compile
    oneflow._oneflow_internal.CurJobBuildAndInferCtx_Complete()
IndexError: _Map_base::at
WARNING: Logging before InitGoogleLogging() is written to STDERR
F0729 02:03:32.338528 13585 global_process_ctx.cpp:47] Check failed: 'Global<ProcessCtx>::Get()' Must be non NULL 
*** Check failure stack trace: ***
    @     0x7f35ade39ee6  google::LogMessage::Fail()
    @     0x7f35ade39e2e  google::LogMessage::SendToLog()
    @     0x7f35ade3976f  google::LogMessage::Flush()
    @     0x7f35ade3cbcc  google::LogMessageFatal::~LogMessageFatal()
    @     0x7f35a8970a73  google::CheckNotNull<>()
    @     0x7f35a8e75a37  oneflow::GlobalProcessCtx::IsThisProcessMaster()
    @     0x7f35a81218d8  oneflow::DestroyLazyGlobalSession()
    @     0x7f35a8122799  DestroyLazyGlobalSession()
    @     0x7f35a7f27db9  _ZNO8pybind116detail15argument_loaderIJEE9call_implIvRPFvvEJENS0_9void_typeEEET_OT0_NS0_14index_sequenceIJXspT1_EEEEOT2_
    @     0x7f35a7f27c74  _ZNO8pybind116detail15argument_loaderIJEE4callIvNS0_9void_typeERPFvvEEENSt9enable_ifIXsrSt7is_voidIT_E5valueES4_E4typeEOT1_
    @     0x7f35a7f27618  _ZZN8pybind1112cpp_function10initializeIRPFvvEvJEJNS_4nameENS_5scopeENS_7siblingEEEEvOT_PFT0_DpT1_EDpRKT2_ENKUlRNS_6detail13function_callEE1_clESL_
    @     0x7f35a7f27689  _ZZN8pybind1112cpp_function10initializeIRPFvvEvJEJNS_4nameENS_5scopeENS_7siblingEEEEvOT_PFT0_DpT1_EDpRKT2_ENUlRNS_6detail13function_callEE1_4_FUNESL_
    @     0x7f35a72fc6e7  pybind11::cpp_function::dispatcher()
    @     0x563999cd9706  PyCFunction_Call
    @     0x563999c9725f  _PyObject_MakeTpCall
    @     0x563999d1f719  _PyEval_EvalFrameDefault
    @     0x563999ce516b  _PyFunction_Vectorcall.localalias.353
    @     0x563999c5956d  _PyEval_EvalFrameDefault.cold.2793
    @     0x563999ce516b  _PyFunction_Vectorcall.localalias.353
    @     0x563999bd7ef9  _PyObject_Vectorcall.lto_priv.9
    @     0x563999c973ef  call_unbound_noarg
    @     0x563999d62b99  slot_tp_finalize
    @     0x563999c962b5  collect.constprop.446
    @     0x563999d8f2aa  _PyGC_CollectNoFail
    @     0x563999da5b55  PyImport_Cleanup
    @     0x563999da5d13  Py_FinalizeEx
    @     0x563999da8200  Py_RunMain
    @     0x563999da8389  Py_BytesMain
    @     0x7f35d34aabf7  __libc_start_main
    @     0x563999d38553  (unknown)
Aborted (core dumped)

模型，删除了中间的东西，下面的 Input_15 是没用到的

name: "mlp"
version: 2
checkpoint_dir: "variables"
graphs {
  key: "mlp_inference"
  value {
    op_list {
      name: "Input_14"
      device_tag: "gpu"
      scope_symbol_id: 4611686018427486206
      input_conf {
        out: "out"
        blob_conf {
          shape {
            dim: 1
            dim: 1
            dim: 28
            dim: 28
          }
          data_type: kFloat
          is_dynamic: false
          parallel_distribution {
            sbp_parallel {
              split_parallel {
                axis: 0
              }
            }
          }
        }
      }
    }
    op_list {
      name: "Input_15"
      device_tag: "gpu"
      scope_symbol_id: 4611686018427486206
      input_conf {
        out: "out"
        blob_conf {
          shape {
            dim: 1
          }
          data_type: kInt32
          is_dynamic: false
          parallel_distribution {
            sbp_parallel {
              split_parallel {
                axis: 0
              }
            }
          }
        }
      }
    }
...(省略)
    op_list {
      name: "Return_17"
      device_tag: "cpu"
      scope_symbol_id: 4611686018427502590
      return_conf {
        in: "dense2-bias_add/out_0"
        out: "out"
      }
    }
    signatures {
      key: "mlp"
      value {
        inputs {
          key: "image"
          value {
            lbi {
              op_name: "Input_14"
              blob_name: "out"
            }
            blob_conf {
              shape {
                dim: 1
                dim: 1
                dim: 28
                dim: 28
              }
              data_type: kFloat
              is_dynamic: false
              parallel_distribution {
                sbp_parallel {
                  split_parallel {
                    axis: 0
                  }
                }
              }
            }
          }
        }
        inputs {
          key: "label"
          value {
            lbi {
              op_name: "Input_15"
              blob_name: "out"
            }
            blob_conf {
              shape {
                dim: 1
              }
              data_type: kInt32
              is_dynamic: false
              parallel_distribution {
                sbp_parallel {
                  split_parallel {
                    axis: 0
                  }
                }
              }
            }
          }
        }
        outputs {
          key: "output"
          value {
            lbi {
              op_name: "dense2-bias_add"
              blob_name: "out_0"
            }
          }
        }
      }
    }
    default_signature_name: "mlp"
  }
}
default_graph_name: "mlp_inference"