Merge pull request #7093 from reyoung/feature/check_nan_executor

Feature/check nan executor

paddle/framework/executor.cc

@@ -14,18 +14,17 @@ limitations under the License. */
 
 #include "paddle/framework/executor.h"
 
-#include <algorithm>
-#include <iostream>
-#include <memory>
 #include <set>
-#include <vector>
 
+#include "gflags/gflags.h"
 #include "paddle/framework/feed_fetch_type.h"
 #include "paddle/framework/lod_rank_table.h"
-#include "paddle/framework/lod_tensor.h"
 #include "paddle/framework/lod_tensor_array.h"
 #include "paddle/framework/op_registry.h"
-#include "paddle/framework/scope.h"
+
+DEFINE_bool(check_nan_inf, false,
+            "Checking whether operator produce NAN/INF or not. It will be "
+            "extremely slow so please use this flag wisely.");
 
 namespace paddle {
 namespace framework {
@@ -58,6 +57,19 @@ static void CreateTensor(Variable* var, proto::VarDesc::VarType var_type) {
   }
 }
 
+static void CheckTensorNANOrInf(const std::string& name,
+                                const framework::Tensor& tensor) {
+  if (tensor.memory_size() == 0) {
+    return;
+  }
+  if (tensor.type().hash_code() != typeid(float).hash_code() &&
+      tensor.type().hash_code() != typeid(double).hash_code()) {
+    return;
+  }
+  PADDLE_ENFORCE(!framework::HasInf(tensor), "Tensor %s has Inf", name);
+  PADDLE_ENFORCE(!framework::HasNAN(tensor), "Tensor %s has NAN", name);
+}
+
 void Executor::Run(const ProgramDesc& pdesc, Scope* scope, int block_id,
                    bool create_local_scope, bool create_vars) {
   // TODO(tonyyang-svail):
@@ -101,6 +113,15 @@ void Executor::Run(const ProgramDesc& pdesc, Scope* scope, int block_id,
     auto op = paddle::framework::OpRegistry::CreateOp(*op_desc);
     VLOG(3) << op->DebugString();
     op->Run(*local_scope, place_);
+    if (FLAGS_check_nan_inf) {
+      for (auto& vname : op->OutputVars(true)) {
+        auto* var = local_scope->FindVar(vname);
+        if (var == nullptr) continue;
+        if (var->IsType<framework::LoDTensor>()) {
+          CheckTensorNANOrInf(vname, var->Get<framework::LoDTensor>());
+        }
+      }
+    }
   }
   if (create_vars && create_local_scope) {
     scope->DeleteScope(local_scope);

paddle/framework/tensor_util.h

@@ -210,10 +210,10 @@ inline void CopyToVector(const Tensor& src, std::vector<T>* dst) {
 }
 
 // Returns true if a tensor contains NAN, i.e., Not A Number.
-extern bool HasNAN(const framework::Tensor& tensor);
+bool HasNAN(const framework::Tensor& tensor);
 
 // Returns true if a tensor contains Inf, i.e., Infinity.
-extern bool HasInf(const framework::Tensor& tensor);
+bool HasInf(const framework::Tensor& tensor);
 
 inline void SerializeToStream(std::ostream& os, const Tensor& tensor,
                               const platform::DeviceContext& dev_ctx) {

paddle/framework/tensor_util_test.cc

@@ -231,7 +231,7 @@ TEST(CopyToVector, Tensor) {
 #endif
 }
 
-TEST(IsNAN, CPU) {
+TEST(HasNAN, CPU) {
   using namespace paddle::framework;
   using namespace paddle::platform;
   Tensor src;
@@ -243,7 +243,7 @@ TEST(IsNAN, CPU) {
   ASSERT_TRUE(HasNAN(src));
 }
 
-TEST(IsInf, CPU) {
+TEST(HasInf, CPU) {
   using namespace paddle::framework;
   using namespace paddle::platform;
   Tensor src;

python/paddle/v2/fluid/__init__.py

@@ -36,7 +36,7 @@ def __read_gflags_from_env__():
     """
     import sys
     import core
-    read_env_flags = ['use_pinned_memory']
+    read_env_flags = ['use_pinned_memory', 'check_nan_inf']
     if core.is_compile_gpu():
         read_env_flags.append('fraction_of_gpu_memory_to_use')
     core.init_gflags([sys.argv[0]] +