Merge pull request #13201 from reyoung/revert_callstack

Revert "Revert "Add Python Callstacks when Op::Run error (#12759)""

paddle/fluid/framework/op_proto_maker.cc

@@ -132,7 +132,9 @@ void OpProtoAndCheckerMaker::operator()(proto::OpProto* proto,
 
   AddAttr<std::string>(OpNamescopeAttrName(), "Operator name with namesope.")
       .SetDefault("");
-
+  AddAttr<std::vector<std::string>>(OpCreationCallstackAttrName(),
+                                    "Callstack for Op Creatation.")
+      .SetDefault({});
   Validate();
 }
 

paddle/fluid/framework/op_proto_maker.h

@@ -46,6 +46,7 @@ class OpProtoAndCheckerMaker {
   static const char *OpRoleAttrName() { return "op_role"; }
   static const char *OpRoleVarAttrName() { return "op_role_var"; }
   static const char *OpNamescopeAttrName() { return "op_namescope"; }
+  static const char *OpCreationCallstackAttrName() { return "op_callstack"; }
 
   void operator()(proto::OpProto *proto, OpAttrChecker *attr_checker);
 

paddle/fluid/framework/operator.cc

@@ -14,15 +14,17 @@ limitations under the License. */
 #define GLOG_NO_ABBREVIATED_SEVERITIES
 #define GOOGLE_GLOG_DLL_DECL
 
+#include "paddle/fluid/framework/operator.h"
 #include <gflags/gflags.h>
 #include <glog/logging.h>
-
 #include <algorithm>
-
+#include <sstream>
+#include <string>
+#include <vector>
 #include "paddle/fluid/framework/data_transform.h"
 #include "paddle/fluid/framework/executor.h"
 #include "paddle/fluid/framework/lod_tensor.h"
-#include "paddle/fluid/framework/operator.h"
+#include "paddle/fluid/framework/op_proto_maker.h"
 #include "paddle/fluid/framework/shape_inference.h"
 #include "paddle/fluid/framework/var_type.h"
 #include "paddle/fluid/platform/profiler.h"
@@ -140,7 +142,10 @@ static LoD GetLoD(const Scope& scope, const std::string& name) {
 }
 
 void OperatorBase::Run(const Scope& scope, const platform::Place& place) {
+  try {
+    if (VLOG_IS_ON(4)) {
       VLOG(4) << place << " " << DebugStringEx(&scope);
+    }
     if (platform::is_gpu_place(place)) {
 #ifndef PADDLE_WITH_CUDA
       PADDLE_THROW("Cannot run operator on place %s", place);
@@ -152,7 +157,33 @@ void OperatorBase::Run(const Scope& scope, const platform::Place& place) {
     platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
     platform::RecordEvent record_event(Type(), pool.Get(place));
     RunImpl(scope, place);
+    if (VLOG_IS_ON(3)) {
       VLOG(3) << place << " " << DebugStringEx(&scope);
+    }
+  } catch (platform::EnforceNotMet exception) {
+    if (Attrs().count("sub_block") != 0) {
+      throw exception;
+    }
+
+    auto& callstack = Attr<std::vector<std::string>>(
+        OpProtoAndCheckerMaker::OpCreationCallstackAttrName());
+
+    if (callstack.empty()) {
+      throw exception;
+    }
+    std::ostringstream sout;
+    sout << "Invoke operator " << Type() << " error.\n";
+    sout << "Python Callstacks: \n";
+    for (auto& line : callstack) {
+      sout << line;
+    }
+    sout << "C++ Callstacks: \n";
+    sout << exception.err_str_;
+    exception.err_str_ = sout.str();
+    throw exception;
+  } catch (...) {
+    std::rethrow_exception(std::current_exception());
+  }
 }
 
 bool OperatorBase::HasInputs(const std::string& name) const {
@@ -180,7 +211,7 @@ const std::vector<std::string>& OperatorBase::Inputs(
 }
 
 bool OperatorBase::HasOutputs(const std::string& name) const {
-  if (outputs_.find(name) != outputs_.end()) {
+  if (outputs_.end() != outputs_.find(name)) {
     return true;
   } else {
     return false;

paddle/fluid/operators/tensorrt_engine_op.h

@@ -36,7 +36,7 @@ namespace operators {
 using FluidDT = framework::proto::VarType_Type;
 using TRT_DT = nvinfer1::DataType;
 
-namespace {
+namespace {  // NOLINT
 
 TRT_DT FluidDataType2TRT(FluidDT type) {
   switch (type) {

paddle/fluid/operators/top_k_op.cc

@@ -30,6 +30,8 @@ class TopkOp : public framework::OperatorWithKernel {
                    "Output(Indices) of TopkOp should not be null.");
 
     auto input_dims = ctx->GetInputDim("X");
+    PADDLE_ENFORCE_EQ(input_dims.size(), 2,
+                      "Rank of TopK op's input must be 2.");
     const int k = static_cast<int>(ctx->Attrs().Get<int>("k"));
 
     PADDLE_ENFORCE_GE(k, 1, "k must >= 1");

paddle/fluid/pybind/const_value.cc

@@ -48,6 +48,9 @@ void BindConstValue(pybind11::module* m) {
   op_proto_and_checker_maker.def(
       "kOpNameScopeAttrName",
       framework::OpProtoAndCheckerMaker::OpNamescopeAttrName);
+  op_proto_and_checker_maker.def(
+      "kOpCreationCallstackAttrName",
+      framework::OpProtoAndCheckerMaker::OpCreationCallstackAttrName);
 }
 
 }  // namespace pybind

python/paddle/fluid/framework.py

@@ -18,6 +18,7 @@ import collections
 import contextlib
 import re
 import six
+import traceback
 
 import numpy as np
 
@@ -34,6 +35,8 @@ except ImportError as e:
 except Exception as e:
     raise e
 from . import unique_name
+import os
+PADDLE_ON_MODEL_CE = os.environ.get('PADDLE_ON_MODEL_CE', None) is not None
 
 __all__ = [
     'Program',
@@ -489,7 +492,8 @@ class OpProtoHolder(object):
         return {
             core.op_proto_and_checker_maker.kOpRoleAttrName(),
             core.op_proto_and_checker_maker.kOpRoleVarAttrName(),
-            core.op_proto_and_checker_maker.kOpNameScopeAttrName()
+            core.op_proto_and_checker_maker.kOpNameScopeAttrName(),
+            core.op_proto_and_checker_maker.kOpCreationCallstackAttrName()
         }
 
 
@@ -572,6 +576,11 @@ class Operator(object):
         if role_var_name in op_attrs and len(op_attrs[role_var_name]) == 0:
             del op_attrs[role_var_name]
 
+        if not PADDLE_ON_MODEL_CE:
+            callstack_var_name = op_maker.kOpCreationCallstackAttrName()
+            op_attrs[callstack_var_name] = list(
+                reversed(traceback.format_stack()))[1:]
+
         if len(self.desc.type()) != 0:
             return
         if type is None:

python/paddle/fluid/tests/unittests/test_operator_desc.py

@@ -69,7 +69,7 @@ class TestOperator(unittest.TestCase):
             set(mul_op.attr_names),
             set([
                 "x_num_col_dims", "y_num_col_dims", "op_role", "op_role_var",
-                "op_namescope"
+                "op_namescope", "op_callstack"
             ]))
         self.assertEqual(mul_op.has_attr("x_num_col_dims"), True)
         self.assertEqual(mul_op.attr_type("x_num_col_dims"), core.AttrType.INT)