OpRegistry::CreateOp调用，输入/输出名字写错，报错信息难以理解

[Xreki]

FCOp，通过组合MulOp, Row,RowwiseAddOp, SoftmaxOp等实现，在C++实现中通过调用OpRegistry::CreateOp来创建相应的Op。

FCOp的输入为X和W，MulOp的输入为X和Y，在创建MulOp将输入误写为：

    AppendOp(framework::OpRegistry::CreateOp(
             "mul", {{"X", {Input("X")}}, {"W", {Input("W")}}},
             {{"Out", {Output("mul_out")}}}, {}));

单测时报错信息如下：

133: ======================================================================
133: ERROR: test_all (__main__.TestFCOp)
133: ----------------------------------------------------------------------
133: Traceback (most recent call last):
133:   File "/home/liuyiqun01/github/Paddle/python/paddle/v2/framework/tests/op_test_util.py", line 55, in test_all
133:     op = Operator(self.type, **kwargs)
133:   File "/home/liuyiqun01/github/Paddle/build_paddle/build/python/build/lib-python/paddle/v2/framework/op.py", line 161, in __call__
133:     return self.get_op_info(t).method(**kwargs)
133:   File "/home/liuyiqun01/github/Paddle/build_paddle/build/python/build/lib-python/paddle/v2/framework/op.py", line 132, in __impl__
133:     return core.Operator.create(opdesc.SerializeToString())
133: RuntimeError: basic_string::_S_construct null not valid
133: 
133: ----------------------------------------------------------------------

这个报错信息太难以理解了，花了很多时间才发现是MulOp输入的名字写错了，正确的写法应该是：

    AppendOp(framework::OpRegistry::CreateOp(
             "mul", {{"X", {Input("X")}}, {"Y", {Input("W")}}},
             {{"Out", {Output("mul_out")}}}, {}));

重构后的Paddle，需要更友好、直观的错误提示。

[QiJune]

遇到过同样的问题，，打了一堆log来定位出错的地方

[reyoung]

Yes, I found a similar problem this weekend.

This should have been fixed in PR #3831. @QiJune @Xreki It is very kind if you can help to check whether the newest develop branch has the same problem or not.

[Xreki]

@reyoung 用的是最新的develop分支，仍然报同样的错误，挂在CheckAllInputOutputSet函数里面。

[lcy-seso]

我也遇到这个问题，使用的也是最新分枝，和 @Xreki 的情形相同。

[Xreki]

我又一次在output名字上栽了个跟头，想想这也是我个人的粗心造成的。具体是这样的：

我在写FCOp的时候，想把输出的key从Out换回Y，于是我在C++代码里面将代码改好了

AddOutput("Y", "The activated output matrix of FC operator");

引用处也都改成了Output("Y")。C++代码看起来是没什么问题了，可是跑单测的时候出了如下错误

136: Test command: /bin/env "PYTHONPATH=/home/liuyiqun01/github/Paddle/build_paddle/build/python/build/lib-python" "python2" "test_fc_op.py"
136: Test timeout computed to be: 9.99988e+06
136: Run
1/1 Test #136: test_fc_op .......................***Exception: SegFault 11.22 sec

经过几番调试，发现是python测试程序里面输出的名字没改过来，功能是Y = X1 * W1 + X2 * W2 + b，我打印出Op的DebugString，如下

136: Op(fc), inputs:{W[W2, W1], X[X2, X1], b[b]}, outputs:{Y[], add_out[add_out], mul_out[mul_out]}.
136: Op(mul), inputs:{X[X2], Y[W2]}, outputs:{Out[mul_out]}.
136: Op(mul), inputs:{X[X1], Y[W1]}, outputs:{Out[add_out]}.
136: Op(add), inputs:{X[mul_out], Y[add_out]}, outputs:{Out[mul_out]}.
136: Op(rowwise_add), inputs:{X[mul_out], b[b]}, outputs:{Out[add_out]}.
136: Op(identity), inputs:{X[add_out]}, outputs:{Y[@EMPTY@]}.
136: Op(scale), inputs:{X[add_out]}, outputs:{Out[@EMPTY@]}.

发现有两个EMPTY的Var，产生这个错误是我自己的疏忽。在这里我想说的两点是：
1. Paddle内部是不是应该对Op的输入输出变量名是否为EMPTY做一下检查？
2. Paddle应该尽量准确地打印错误信息，而不是这样由系统扔出来一个Exception: SegFault，真的很难查。。。

[lcy-seso]

我们是否可以一起努力，保证此 PR 的尽快merge：#3452

[lcy-seso]

另一个和命名相关的小问题 ：#3976

[Xreki]

@reyoung 我查到产生上述log的根源了，是由PADDLE_ENFORCE语句产生的。
为了验证PADDLE_ENFORCE，我做了如下实验：

• git pull当前最新develop分支，将mul_op.cc里面的一个PADDLE_ENFORCE改成肯定不满足的情况，如下

 45     PADDLE_ENFORCE_EQ(
 46         // x_mat_dims[1], y_mat_dims[0],
 47         1, 2,
 48         "First matrix's width must be equal with second matrix's height.");

• 按照设定，执行单测test_mul_op的时候，单测会挂掉，预期的输出是：

125: Check failed in mul_op.cc, line 45
125: First matrix's width must be equal with second matrix's height.
...

• 但实际的输出是

test 125
    Start 125: test_mul_op

125: Test command: /bin/env "PYTHONPATH=/home/liuyiqun01/github/Paddle/build_paddle/build/python/build/lib-python" "python2" "test_mul_op.py"
125: Test timeout computed to be: 9.99988e+06
125: EEEEEEEEEE
125: ======================================================================
125: ERROR: test_cpu_gpu_compare (__main__.TestMulGradOp)
125: ----------------------------------------------------------------------
125: Traceback (most recent call last):
125:   File "test_mul_op.py", line 45, in test_cpu_gpu_compare
125:     self.compare_grad(self.op, self.inputs)
125:   File "/home/liuyiqun01/github/Paddle/python/paddle/v2/framework/tests/gradient_checker.py", line 216, in compare_grad
125:     out_names, core.CPUPlace())
125:   File "/home/liuyiqun01/github/Paddle/python/paddle/v2/framework/tests/gradient_checker.py", line 164, in __get_gradient
125:     forward_op.infer_shape(scope)
125: RuntimeError: basic_string::_S_construct null not valid
125: 

我们设定的错误提示完全没有输出。。。

再来说下，CreateOp中，输入/输出key写错了这个错误的原因，可以看下operator.cc中OperatorBase的构造函数：

OperatorBase::OperatorBase(const std::string& type,
                           const VariableNameMap& inputs,
                           const VariableNameMap& outputs,
                           const AttributeMap& attrs)
    : type_(type), inputs_(inputs), outputs_(outputs), attrs_(attrs) {
  GenerateTemporaryNames();
  CheckAllInputOutputSet();
}

其实在CheckAllInputOutputSet()有对输入输出进行检查：

void OperatorBase::CheckAllInputOutputSet() const {
  auto& info_map = OpInfoMap::Instance();
  auto* op_info = info_map.GetNullable(Type());
  if (op_info == nullptr || op_info->proto_ == nullptr) return;

  for (auto& in : op_info->Proto().inputs()) {
    PADDLE_ENFORCE(inputs_.find(in.name()) != inputs_.end(),
                   "Type %s's input %s is not set", Type(), in.name());
  }

  for (auto& out : op_info->Proto().outputs()) {
    PADDLE_ENFORCE(outputs_.find(out.name()) != outputs_.end(),
                   "Type %s's output %s is not set", Type(), out.name());
  }
}

没错，当我们输入/输出的key配置错误时，就是挂在这两个PADDLE_ENFORCE。然而没有准确的错误提示信息，让我们调程序变的困难无比。

[Xreki]

@Superjom 说PADDLE_ENFORCE能在C++测试中打印Stack信息，只是在Python测试中存在问题。我做了如下一个测试。
新建一个enforce_failure_test.cc，其中的内容为：

#include "gtest/gtest.h"
#include "paddle/platform/enforce.h"

TEST(ENFORCE, Failure) {
  PADDLE_ENFORCE(false, "Example to show the behavior when enforce is failed.");
}

执行单测make test ARGS="-R enforce_failure_test -V"，输出如下：

$ make test ARGS="-R enforce_failure_test -V"  
Running tests...
UpdateCTestConfiguration  from :/home/liuyiqun01/github/Paddle/build_paddle/build/DartConfiguration.tcl
UpdateCTestConfiguration  from :/home/liuyiqun01/github/Paddle/build_paddle/build/DartConfiguration.tcl
Test project /home/liuyiqun01/github/Paddle/build_paddle/build
Constructing a list of tests
Done constructing a list of tests
Updating test list for fixtures
Added 0 tests to meet fixture requirements
Checking test dependency graph...
Checking test dependency graph end
test 84
    Start 84: enforce_failure_test

84: Test command: /home/liuyiqun01/github/Paddle/build_paddle/build/paddle/platform/enforce_failure_test
84: Test timeout computed to be: 9.99988e+06
84: Running main() from gtest_main.cc
84: [==========] Running 1 test from 1 test case.
84: [----------] Global test environment set-up.
84: [----------] 1 test from ENFORCE
84: [ RUN      ] ENFORCE.Failure
84: unknown file: Failure
84: C++ exception with description "basic_string::_S_construct null not valid" thrown in the test body.
84: [  FAILED  ] ENFORCE.Failure (1 ms)
84: [----------] 1 test from ENFORCE (1 ms total)
84: 
84: [----------] Global test environment tear-down
84: [==========] 1 test from 1 test case ran. (1 ms total)
84: [  PASSED  ] 0 tests.
84: [  FAILED  ] 1 test, listed below:
84: [  FAILED  ] ENFORCE.Failure
84: 
84:  1 FAILED TEST
1/1 Test #84: enforce_failure_test .............***Failed    0.00 sec

0% tests passed, 1 tests failed out of 1

Total Test time (real) =   0.01 sec

The following tests FAILED:
         84 - enforce_failure_test (Failed)
Errors while running CTest
make: *** [test] Error 8

[Xreki]

PADDLE_ENFORCE的问题已经在#4002 中解决，谢谢@reyoung 。当CreateOp中误将op的输出key Y写成W，能正确地输出错误提示信息、堆栈信息。

136: ERROR: test_check_output (__main__.TestFCOp)
136: ----------------------------------------------------------------------
136: Traceback (most recent call last):
136:   File "test_fc_op.py", line 40, in test_check_output
136:     self.check_output()
136:   File "/home/liuyiqun01/github/Paddle/python/paddle/v2/framework/tests/op_test.py", line 203, in check_output
136:     self.check_output_with_place(place)
136:   File "/home/liuyiqun01/github/Paddle/python/paddle/v2/framework/tests/op_test.py", line 171, in check_output_with_place
136:     self.op = create_op(self.scope, self.op_type, self.inputs, self.outputs)
136:   File "/home/liuyiqun01/github/Paddle/python/paddle/v2/framework/tests/op_test.py", line 41, in create_op
136:     return Operator(op_type, **kwargs)
136:   File "/home/liuyiqun01/github/Paddle/build_paddle/build/python/build/lib-python/paddle/v2/framework/op.py", line 172, in __call__
136:     return self.get_op_info(t).method(**kwargs)
136:   File "/home/liuyiqun01/github/Paddle/build_paddle/build/python/build/lib-python/paddle/v2/framework/op.py", line 140, in __impl__
136:     return core.Operator.create(opdesc.SerializeToString())
136: RuntimeError: Type mul's input Y is not set at [/home/liuyiqun01/github/Paddle/paddle/framework/operator.cc:167]
136: PaddlePaddle Call Stacks: 
136: 0       0x7fe466e44527p _ZN6paddle8platform13EnforceNotMetC2ENSt15__exception_ptr13exception_ptrEPKci + 663
136: 1       0x7fe4683b55fap _ZNK6paddle9framework12OperatorBase22CheckAllInputOutputSetEv + 426
...